Daily Papers

Reinforcement learning with verifiable reward (RLVR) has become a promising paradigm for post-training large language models (LLMs) to improve their reasoning capability. However, when the rollout accuracy is low on hard problems, the reward becomes sparse, limiting learning efficiency and causing exploration bottlenecks. Existing approaches either rely on teacher models for distillation or filter out difficult problems, which limits scalability or restricts reasoning improvement through exploration. We propose EvoCoT, a self-evolving curriculum learning framework based on two-stage chain-of-thought (CoT) reasoning optimization. EvoCoT constrains the exploration space by self-generating and verifying CoT trajectories, then gradually shortens CoT steps to expand the space in a controlled way. The framework enables LLMs to stably learn from initially unsolved hard problems under sparse rewards. We apply EvoCoT to multiple LLM families, including Qwen, DeepSeek, and Llama. Experiments show that EvoCoT enables LLMs to solve previously unsolved problems, improves reasoning capability without external CoT supervision, and is compatible with various RL fine-tuning methods. We release the source code to support future research.

Knowledge distillation transfers a teacher's competence to a small student but is brittle in the small-student regime: forcing the student to imitate logits from a much larger teacher concentrates it on the teacher's sharpest modes, hurting generalization on benchmark families beyond the training corpus. Reinforcement learning (RL) avoids logit imitation by training on the student's own rollouts. However, on questions where every rollout fails-yielding zero advantage and being silently discarded-injecting a stronger teacher's response into the policy gradient breaks the on-policy assumption and induces drift. We introduce Zone of Proximal Policy Optimization (ZPPO), inspired by Vygotsky's zone of proximal development, which keeps the teacher inside the prompt rather than the policy gradient. On hard questions, ZPPO constructs two reformulated prompts: a Binary Candidate-included Question (BCQ) pairs one correct teacher response with one incorrect student response as anonymized candidates the student must discriminate, and a Negative Candidate-included Question (NCQ) aggregates the student's wrong rollouts into a single prompt to surface their shared failure modes. A prompt replay buffer recirculates each hard question until it either graduates-the student's mean rollout accuracy on it reaches half- or is FIFO-evicted under finite capacity, amplifying BCQ and NCQ inside the student's current zone of proximal development. On the Qwen3.5 family at four student scales (0.8B-9B) with a 27B teacher, post-trained as vision-language models and evaluated on a 31-benchmark suite (16 VLM, 10 LLM, 5 Video), ZPPO outperforms off/on-policy distillation and GRPO, with the largest gains at the smallest scale.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a promising approach to improve the reasoning abilities of Large Language Models (LLMs). Among RLVR algorithms, Group Relative Policy Optimization (GRPO) and its variants have demonstrated strong performance and high training efficiency. However, GRPO-style objectives exhibit two issues on high accuracy prompts including mastered prompts (rollout accuracy =1) and majority-correct prompts (rollout accuracy in (0.5,1)). For mastered prompts, group-relative advantages vanish, yielding no training signal and unconstrained policy drift that can cause forgetting. For majority-correct prompts, the induced query weight shrinks as accuracy increases, weakening consolidation from partial correctness to mastery. To alleviate this, we propose Mastery-Consolidated Policy Optimization (MCPO), which introduces (i) a hinge-KL regularizer applied exclusively to mastered prompts to bound harmful policy drift between successive gradient steps, and (ii) a weighting mechanism that prioritizes majority-correct prompts to better allocate optimization effort. Extensive experiments across three mathematical benchmarks demonstrate that MCPO consistently improves pass@1 performance. Counter-intuitively, rather than restricting exploration, MCPO boosts pass@k metrics, indicating that mastery consolidation further catalyzes solution diversity.

Reinforcement fine-tuning (RFT) has emerged as a powerful post-training technique to incentivize the reasoning ability of large language models (LLMs). However, LLMs can respond very inconsistently to RFT: some show substantial performance gains, while others plateau or even degrade. To understand this divergence, we analyze the per-step influence of the RL objective and identify two key conditions for effective post-training: (1) RL-informative rollout accuracy, and (2) strong data co-influence, which quantifies how much the training data affects performance on other samples. Guided by these insights, we propose behavior injection, a task-agnostic data-augmentation scheme applied prior to RL. Behavior injection enriches the supervised finetuning (SFT) data by seeding exploratory and exploitative behaviors, effectively making the model more RL-ready. We evaluate our method across two reasoning benchmarks with multiple base models. The results demonstrate that our theoretically motivated augmentation can significantly increases the performance gain from RFT over the pre-RL model.

Reinforcement learning with verifiable rewards (RLVR) has achieved remarkable success in enhancing the reasoning capabilities of large language models (LLMs). However, existing RLVR methods often suffer from exploration inefficiency due to mismatches between the training data's difficulty and the model's capability. LLMs fail to discover viable reasoning paths when problems are overly difficult, while learning little new capability when problems are too simple. In this work, we formalize the impact of problem difficulty by quantifying the relationship between loss descent speed and rollout accuracy. Building on this analysis, we propose SEELE, a novel supervision-aided RLVR framework that dynamically adjusts problem difficulty to stay within the high-efficiency region. SEELE augments each training sample by appending a hint (part of a full solution) after the original problem. Unlike previous hint-based approaches, SEELE deliberately and adaptively adjusts the hint length for each problem to achieve an optimal difficulty. To determine the optimal hint length, SEELE employs a multi-round rollout sampling strategy. In each round, it fits an item response theory model to the accuracy-hint pairs collected in preceding rounds to predict the required hint length for the next round. This instance-level, real-time difficulty adjustment aligns problem difficulty with the evolving model capability, thereby improving exploration efficiency. Experimental results show that SEELE outperforms Group Relative Policy Optimization (GRPO) and Supervised Fine-tuning (SFT) by +11.8 and +10.5 points, respectively, and surpasses the best previous supervision-aided approach by +3.6 points on average across six math reasoning benchmarks.

Large multimodal language models (MLLMs) are increasingly deployed as web agents, yet many multimodal browsing benchmarks can be solved by shallow, fixed workflows that lean on high-recall image search and nearby text-masking the genuinely multimodal challenges of fine-grained visual reasoning, provenance verification, and long-horizon tool use. We introduce MMSearch-Plus, a benchmark of 311 tasks that highly demand multimodal understanding while preserving the difficulty profile of strong text-only browsing suites. Each item is constructed to contain multiple weak, localized visual signals that must be extracted, propagated through iterative text-image search, and cross-validated under retrieval noise before answering. Our curation procedure, Spatial-Temporal Extrapolation, seeds questions whose answers require extrapolating from spatial cues (micro-text, part-level appearance, layouts, signage) and temporal traces (broadcast overlays, seasonal context) to out-of-image facts such as events, dates, and venues. We provide a model-agnostic agent framework with browsing tools and evaluate a range of closed and open MLLMs. The strongest agent (o3) attains 15.1% without search and 36.0% accuracy with rollout under our framework, while a strong open-source model (Qwen-2.5-VL-72B-Instruct) achieves 0.0% without search and 6.9% after 20 rounds of search. Beyond answer accuracy, we assess bounding-box production and cropped-image search, and conduct an error analysis that surfaces failures in source verification, part-based reasoning, and long-horizon planning.

Pixel-space continuous-token autoregressive (AR) generation directly models images as sequences of raw pixel patches, avoiding discrete tokenization or a separately pretrained tokenizer. However, it faces coupled challenges: high-dimensional patch generation causes large single-step errors, and teacher-forced training creates a train--inference gap that makes these errors accumulate across AR steps. Existing fixes such as x-prediction and input noise injection only partially mitigate these issues. Exact rollout training better matches inference-time conditions, but is impractical due to prohibitively slow sequential sampling. We propose Parallel Rollout Approximation (PRA), a scalable framework that addresses both challenges jointly. PRA generates low-dimensional intermediate states instead of high-dimensional pixel patches, then maps them back to pixel-space tokens with a pixel decoder, preserving a pixel-in, pixel-out AR interface. It also constructs inference-like pixel inputs through the same intermediate-state-to-pixel path used at inference, independently across positions, approximating the pixel-feedback interface encountered during inference-time rollout while retaining parallel teacher-forced training. On class-conditional ImageNet-1K generation at 256times256 resolution, PRA-S with 135M parameters achieves an FID of 2.58, surpassing the previous billion-scale pixel-space AR result of 3.60. Scaling to PRA-L with 511M parameters further improves FID to 1.94, establishing a new state of the art among pixel-space AR models. Beyond generation, PRA achieves higher ImageNet classification probing accuracy than other AR and diffusion baselines, suggesting its potential for unified pixel-space image generation and understanding.

Reinforcement Learning with Verifiable Rewards (RLVR) has significantly advanced the reasoning capabilities of Large Language Models (LLMs). However, methods such as GRPO and DAPO suffer from substantial computational cost, since they rely on sampling many rollouts for each prompt. Moreover, in RLVR the relative advantage is often sparse: many samples become nearly all-correct or all-incorrect, yielding low within-group reward variance and thus weak learning signals. In this paper, we introduce arrol (Accelerating RLVR via online Rollout Pruning), an online rollout pruning method that prunes rollouts during generation while explicitly steering the surviving ones more correctness-balanced to enhance learning signals. Specifically, arrol trains a lightweight quality head on-the-fly to predict the success probability of partial rollouts and uses it to make early pruning decisions. The learned quality head can further weigh candidates to improve inference accuracy during test-time scaling. To improve efficiency, we present a system design that prunes rollouts inside the inference engine and re-batches the remaining ones for log-probability computation and policy updates. Across GRPO and DAPO on Qwen-3 and LLaMA-3.2 models (1B-8B), arrol improves average accuracy by +2.30 to +2.99 while achieving up to 1.7x training speedup, and yielding up to +8.33 additional gains in average accuracy in test-time scaling. The code is available at https://github.com/Hsu1023/ARRoL.

Post-training for long-horizon agentic tasks has a tension between compute efficiency and generalization. While supervised fine-tuning (SFT) is compute efficient, it often suffers from out-of-domain (OOD) degradation. Conversely, end-to-end reinforcement learning (E2E RL) preserves OOD capabilities, but incurs high compute costs due to many turns of on-policy rollout. We introduce PivotRL, a novel framework that operates on existing SFT trajectories to combine the compute efficiency of SFT with the OOD accuracy of E2E RL. PivotRL relies on two key mechanisms: first, it executes local, on-policy rollouts and filters for pivots: informative intermediate turns where sampled actions exhibit high variance in outcomes; second, it utilizes rewards for functional-equivalent actions rather than demanding strict string matching with the SFT data demonstration. We theoretically show that these mechanisms incentivize strong learning signals with high natural gradient norm, while maximally preserving policy probability ordering on actions unrelated to training tasks. In comparison to standard SFT on identical data, we demonstrate that PivotRL achieves +4.17% higher in-domain accuracy on average across four agentic domains, and +10.04% higher OOD accuracy in non-agentic tasks. Notably, on agentic coding tasks, PivotRL achieves competitive accuracy with E2E RL with 4x fewer rollout turns. PivotRL is adopted by NVIDIA's Nemotron-3-Super-120B-A12B, acting as the workhorse in production-scale agentic post-training.

Data-driven surrogate models improve the efficiency of simulating continuous dynamical systems, yet their autoregressive rollouts are often limited by instability and spectral blow-up. While global regularization techniques can enforce contractive dynamics, they uniformly damp high-frequency features, introducing a contraction-dissipation dilemma. Furthermore, long-horizon trajectory optimization methods that explicitly correct drift are bottlenecked by memory constraints. In this work, we propose Jacobian-Adaptive Weighting for Stability (JAWS), a probabilistic regularization strategy designed to mitigate these limitations. By framing operator learning as Maximum A Posteriori (MAP) estimation with spatially heteroscedastic uncertainty, JAWS dynamically modulates the regularization strength based on local physical complexity. This allows the model to enforce contraction in smooth regions to suppress noise, while relaxing constraints near singular features to preserve gradients, effectively realizing a behavior similar to numerical shock-capturing schemes. Experiments demonstrate that this spatially-adaptive prior serves as an effective spectral pre-conditioner, which reduces the base operator's burden of handling high-frequency instabilities. This reduction enables memory-efficient, short-horizon trajectory optimization to match or exceed the long-term accuracy of long-horizon baselines. Evaluated on the 1D viscous Burgers' equation, our hybrid approach improves long-term stability, shock fidelity, and out-of-distribution generalization while reducing training computational costs.

Vision-Language Models (VLMs) excel at many multimodal tasks, yet they frequently struggle with tasks requiring precise understanding and handling of fine-grained visual elements. This is mainly due to information loss during image encoding or insufficient attention to critical regions. Recent work has shown promise by incorporating pixel-level visual information into the reasoning process, enabling VLMs to access high-resolution visual details during their thought process. However, this pixel-level information is often overused, leading to inefficiency and distraction from irrelevant visual details. To address these challenges, we propose the first framework for adaptive pixel reasoning that dynamically determines necessary pixel-level operations based on the input query. Specifically, we first apply operation-aware supervised fine-tuning to establish baseline competence in textual reasoning and visual operations, then design a novel rollout-guided reinforcement learning framework relying on feedback of the model's own responses, which enables the VLM to determine when pixel operations should be invoked based on query difficulty. Experiments on extensive multimodal reasoning benchmarks show that our model achieves superior performance while significantly reducing unnecessary visual operations. Impressively, our model achieves 73.4\% accuracy on HR-Bench 4K while maintaining a tool usage ratio of only 20.1\%, improving accuracy and simultaneously reducing tool usage by 66.5\% compared to the previous methods.

Model-based reinforcement learning promises to learn an optimal policy from fewer interactions with the environment compared to model-free reinforcement learning by learning an intermediate model of the environment in order to predict future interactions. When predicting a sequence of interactions, the rollout length, which limits the prediction horizon, is a critical hyperparameter as accuracy of the predictions diminishes in the regions that are further away from real experience. As a result, with a longer rollout length, an overall worse policy is learned in the long run. Thus, the hyperparameter provides a trade-off between quality and efficiency. In this work, we frame the problem of tuning the rollout length as a meta-level sequential decision-making problem that optimizes the final policy learned by model-based reinforcement learning given a fixed budget of environment interactions by adapting the hyperparameter dynamically based on feedback from the learning process, such as accuracy of the model and the remaining budget of interactions. We use model-free deep reinforcement learning to solve the meta-level decision problem and demonstrate that our approach outperforms common heuristic baselines on two well-known reinforcement learning environments.

Test-Time Scaling (TTS) improves the performance of Large Language Models (LLMs) by using additional inference-time computation to explore multiple reasoning paths through search. Yet how to allocate a fixed rollout budget most effectively during search remains underexplored, often resulting in inefficient use of compute at test time. To bridge this gap, we formulate test-time search as a resource allocation problem and derive the optimal allocation strategy that maximizes the probability of obtaining a correct solution under a fixed rollout budget. Within this formulation, we reveal a core limitation of existing search methods: solution-level allocation tends to favor reasoning directions with more candidates, leading to theoretically suboptimal and inefficient use of compute. To address this, we propose Direction-Oriented Resource Allocation (DORA), a provably optimal method that mitigates this bias by decoupling direction quality from candidate count and allocating resources at the direction level. To demonstrate DORA's effectiveness, we conduct extensive experiments on challenging mathematical reasoning benchmarks including MATH500, AIME2024, and AIME2025. The empirical results show that DORA consistently outperforms strong baselines with comparable computational cost, achieving state-of-the-art accuracy. We hope our findings contribute to a broader understanding of optimal TTS for LLMs.

Reinforcement learning with verifiable rewards (RLVR) is a promising approach for enhancing reasoning and agentic behavior in large language models. However, rollout-intensive policy optimization is often limited by insufficient reward contrast, arising when overly simple or complex prompts generate low-variance feedback and when outcome-only rewards assign the same terminal assessment to every decision in a multi-turn rollout. Past efforts have focused on allocating available rollout resources to promising prompts, yet they only leverage sample informativeness at the prompt level and neglect variation in prefix-level informativeness across turns within the same rollout. This work targets multi-turn agentic RL by modeling each ReAct-style thought-action-observation turn as a semantically distinct node, allowing budget allocation to extend from prompt roots to turn-level prefixes with further continuations, which naturally forms tree-structured rollouts. We introduce Tree Rollout Allocation for Contrastive Exploration (TRACE), a unified rollout allocation framework that enhances reward contrast within a fixed sampling budget. Technically, TRACE allocates rollout budget to both prompt roots and intermediate prefixes that are most likely to yield mixed terminal rewards. A shared generalizable predictor estimates conditional success probability at these anchors from prefix histories to guide this allocation. The resulting adaptive tree structure enriches outcome-only feedback and amplifies the policy-update signal. Empirically, TRACE achieves competitive performance and efficiency gains on typical agentic benchmarks, e.g., improving Qwen3-14B Multi-Hop QA average accuracy by 2.8 points over competitive baselines at equal sampling cost.

Reinforcement Learning with Verifiable Rewards (RLVR) has become a key paradigm to improve the reasoning capabilities of Multimodal Large Language Models (MLLMs). However, prevalent group-based algorithms such as GRPO require multi-rollout sampling for each prompt. While more efficient single-rollout variants have recently been explored in text-only settings, we find that they suffer from severe instability in multimodal contexts, often leading to training collapse. To address this training efficiency-stability trade-off, we introduce MSSR (Multimodal Stabilized Single-Rollout), a group-free RLVR framework that achieves both stable optimization and effective multimodal reasoning performance. MSSR achieves this via an entropy-based advantage-shaping mechanism that adaptively regularizes advantage magnitudes, preventing collapse and maintaining training stability. While such mechanisms have been used in group-based RLVR, we show that in the multimodal single-rollout setting they are not merely beneficial but essential for stability. In in-distribution evaluations, MSSR demonstrates superior training compute efficiency, achieving similar validation accuracy to the group-based baseline with half the training steps. When trained for the same number of steps, MSSR's performance surpasses the group-based baseline and shows consistent generalization improvements across five diverse reasoning-intensive benchmarks. Together, these results demonstrate that MSSR enables stable, compute-efficient, and effective RLVR for complex multimodal reasoning tasks.

Group-relative policy optimization methods train language models by generating multiple rollouts per prompt and normalizing rewards with a shared mean reward baseline. In resource-constrained settings where the rollout budget is small, accuracy often degrades. We find that noise in the shared baseline induces advantage sign flips, where some rollouts receive an incorrect advantage sign, and the update direction is reversed. To address this, we propose Median-Centered Group Relative Policy Optimization (MC-GRPO), a simple and effective solution for small-rollout training. Our main idea is to replace the mean baseline with a median baseline: the median is far less sensitive to outlier rewards than the mean, mitigating the sign flips under small rollout size (G). We generate one additional rollout for median reference (G+1), and compute advantages by using the group median. With an odd-sized group, exactly one completion is the median and receives zero advantage, we exclude this pivot rollout from backpropagation so the number of gradient-contributing samples per prompt remains G, preserving the core update cost of standard G-rollout training. Across various GRPO-family methods and a wide range of models and scales, this median-centered training consistently improves stability and final accuracy in the low-rollout regime, reducing the gap between G=2 and G=8 to within 1%. Code is available at https://github.com/lotusroot-kim/MC-GRPO

Reinforcement learning (RL) is the dominant paradigm for post-training large language models. However, in the online, on-policy setting, rollout generation dominates the computational cost of training. Group-based policy optimization methods compute advantages from multiple rollouts per prompt, yet they indiscriminately allocate budget to prompts with collapsed reward distributions, wasting expensive rollouts on negligible learning signals. We demonstrate that group-based updates are most effective in regimes of high reward variance. Since the policy evolves throughout training, prompt informativeness must be estimated online rather than precomputed, but exhaustively evaluating every prompt is computationally prohibitive. We introduce Pilot-Commit, a budget-aware rollout allocation framework for group-based RL post-training. Pilot-Commit decouples prompt evaluation from exploitation: a pilot stage estimates per-prompt informativeness using a fraction of the budget, and the remaining rollouts are allocated to high-leverage prompts while low-signal prompts are skipped. Across multiple math reasoning benchmarks and model scales from 1.5B to 14B parameters, Pilot-Commit matches baseline accuracy with significantly lower sampling costs, reaching target accuracy up to 1.9times faster than GRPO and 4.0times faster than DAPO in cumulative rollouts.

Reinforcement learning with verifiable rewards (RLVR) has proven effective in eliciting complex reasoning in large language models (LLMs). However, standard RLVR training often leads to excessively verbose processes (in reasoning tasks) and inefficient exploration trajectories (in agentic settings), as outcome-only rewards provide no incentive for efficiency and the high variance in response length within relatively small rollout groups results in noisy optimization signals. To address this, we propose Rollout Response Recomposition (RoRecomp), a plug-and-play method that guides models toward concise reasoning by strategically recomposing the training data. RoRecomp separates responses into two distinct batch types: 1) priority batches, which combine short-correct and long-incorrect responses selected from online batches to provide a clear gradient signal for brevity, and 2) compensation batches, which utilize remaining responses from a replay buffer to maintain stability and prevent model collapse. To comprehensively evaluate effectiveness, we test RoRecomp across three settings where results demonstrate substantial efficiency gains: reducing reasoning length by 27.7% in zero RL training, reducing unnecessary tool calls by 46.8% while improving accuracy in agentic RL, and achieving up to 52.5% length reduction in thinking compression, all with minimal performance impact.

Reinforcement learning (RL) is essential for enhancing the complex reasoning capabilities of large language models (LLMs). However, existing RL training pipelines are computationally inefficient and resource-intensive, with the rollout phase accounting for over 70% of total training time. Quantized RL training, particularly using FP8 precision, offers a promising approach to mitigating this bottleneck. A commonly adopted strategy applies FP8 precision during rollout while retaining BF16 precision for training. In this work, we present the first comprehensive study of FP8 RL training and demonstrate that the widely used BF16-training + FP8-rollout strategy suffers from severe training instability and catastrophic accuracy collapse under long-horizon rollouts and challenging tasks. Our analysis shows that these failures stem from the off-policy nature of the approach, which introduces substantial numerical mismatch between training and inference. Motivated by these observations, we propose Jet-RL, an FP8 RL training framework that enables robust and stable RL optimization. The key idea is to adopt a unified FP8 precision flow for both training and rollout, thereby minimizing numerical discrepancies and eliminating the need for inefficient inter-step calibration. Extensive experiments validate the effectiveness of Jet-RL: our method achieves up to 33% speedup in the rollout phase, up to 41% speedup in the training phase, and a 16% end-to-end speedup over BF16 training, while maintaining stable convergence across all settings and incurring negligible accuracy degradation.

As LLM-based agents are deployed in production systems, understanding their behavioral consistency (whether they produce similar action sequences when given identical tasks) becomes critical for reliability. We study consistency in the context of SWE-bench, a challenging software engineering benchmark requiring complex, multi-step reasoning. Comparing Claude~4.5~Sonnet, GPT-5, and Llama-3.1-70B across 50 runs each (10 tasks times 5 runs), we find that across models, higher consistency aligns with higher accuracy: Claude achieves the lowest variance (CV: 15.2\%) and highest accuracy (58\%), GPT-5 is intermediate (CV: 32.2\%, accuracy: 32\%), and Llama shows the highest variance (CV: 47.0\%) with lowest accuracy (4\%). However, within a model, consistency can amplify both correct and incorrect interpretations. Our analysis reveals a critical nuance: consistency amplifies outcomes rather than guaranteeing correctness. 71\% of Claude's failures stem from "consistent wrong interpretation": making the same incorrect assumption across all runs. Interestingly, GPT-5 achieves similar early strategic agreement as Claude (diverging at step 3.4 vs.\ 3.2) but exhibits 2.1times higher variance, suggesting that divergence timing alone does not determine consistency. These findings suggest that for production deployment, interpretation accuracy matters more than execution consistency, with implications for agent evaluation and training.

Reinforcement Learning from Verifiable Rewards (RLVR) has emerged as a powerful paradigm for enhancing the complex reasoning capabilities of Large Reasoning Models. However, standard outcome-based supervision suffers from a critical limitation that penalizes trajectories that are largely correct but fail due to several missteps as heavily as completely erroneous ones. This coarse feedback signal causes the model to discard valuable largely correct rollouts, leading to a degradation in rollout diversity that prematurely narrows the exploration space. Process Reward Models have demonstrated efficacy in providing reliable step-wise verification for test-time scaling, naively integrating these signals into RLVR as dense rewards proves ineffective.Prior methods attempt to introduce off-policy guided whole-trajectory replacement that often outside the policy model's distribution, but still fail to utilize the largely correct rollouts generated by the model itself and thus do not effectively mitigate the narrowing of the exploration space. To address these issues, we propose SCOPE (Step-wise Correction for On-Policy Exploration), a novel framework that utilizes Process Reward Models to pinpoint the first erroneous step in suboptimal rollouts and applies fine-grained, step-wise off-policy rectification. By applying precise refinement on partially correct rollout, our method effectively salvages partially correct trajectories and increases diversity score by 13.5%, thereby sustaining a broad exploration space. Extensive experiments demonstrate that our approach establishes new state-of-the-art results, achieving an average accuracy of 46.6% on math reasoning and exhibiting robust generalization with 53.4% accuracy on out-of-distribution reasoning tasks.

We introduce ROLL, an efficient, scalable, and user-friendly library designed for Reinforcement Learning Optimization for Large-scale Learning. ROLL caters to three primary user groups: tech pioneers aiming for cost-effective, fault-tolerant large-scale training, developers requiring flexible control over training workflows, and researchers seeking agile experimentation. ROLL is built upon several key modules to serve these user groups effectively. First, a single-controller architecture combined with an abstraction of the parallel worker simplifies the development of the training pipeline. Second, the parallel strategy and data transfer modules enable efficient and scalable training. Third, the rollout scheduler offers fine-grained management of each sample's lifecycle during the rollout stage. Fourth, the environment worker and reward worker support rapid and flexible experimentation with agentic RL algorithms and reward designs. Finally, AutoDeviceMapping allows users to assign resources to different models flexibly across various stages.

We study whether demonstration-curation metrics that detect defective training episodes also improve the downstream behavior-cloning policy that trains on the curated data. On a contact-rich LIBERO pick-and-place benchmark with a controlled structural defect (early gripper release during the carry phase), we find that the two quantities are sharply decoupled. The metric with the highest defect-detection AUROC (0.804) produces the worst curated policy (13.3% task success), while a metric with a substantially lower AUROC (0.638) produces a policy that nearly matches the oracle trained on ground-truth clean data (90.0% vs. 93.3%). We further show that five of the seven metrics we evaluate exploit episode length as a trivial proxy for the defect label, a confound that inflates reported AUROCs to near-perfect values and disappears once episode length is controlled. Across all conditions, the contaminated baseline succeeds on only 3.3% of rollouts, and the two best curation methods close this to within 3 percentage points of the 93.3% oracle ceiling. Our results argue that curation methods should be evaluated by the policy they produce, not the defects they flag, and that any curation benchmark must control for episode length before reporting detection accuracy. We release the testbed, all metric implementations, and the evaluation pipeline.

Reinforcement learning (RL) has become a central post-training tool for improving the reasoning abilities of large language models (LLMs). In these systems, the rollout, the trajectory sampled from a prompt to termination, including intermediate reasoning steps and optional tool or environment interactions, determines the data the optimizer learns from, yet rollout design is often underreported. This survey provides an optimizer-agnostic view of rollout strategies for RL-based post-training of reasoning LLMs. We formalize rollout pipelines with unified notation and introduce Generate-Filter-Control-Replay (GFCR), a lifecycle taxonomy that decomposes rollout pipelines into four modular stages: Generate proposes candidate trajectories and topologies; Filter constructs intermediate signals via verifiers, judges, critics; Control allocates compute and makes continuation/branching/stopping decisions under budgets; and Replay retains and reuses artifacts across rollouts without weight updates, including self-evolving curricula that autonomously generate new training tasks. We complement GFCR with a criterion taxonomy of reliability, coverage, and cost sensitivity that characterizes rollout trade-offs. Using this framework, we synthesize methods spanning RL with verifiable rewards, process supervision, judge-based gating, guided and tree/segment rollouts, adaptive compute allocation, early-exit and partial rollouts, throughput optimization, and replay/recomposition for self-improvement. We ground the framework with case studies in math, code/SQL, multimodal reasoning, tool-using agents, and agentic skill benchmarks that evaluate skill induction, reuse, and cross-task transfer. Finally, we provide a diagnostic index that maps common rollout pathologies to GFCR modules and mitigation levers, alongside open challenges for building reproducible, compute-efficient, and trustworthy rollout pipelines.

Chain-of-thought (CoT) reasoning improves large language models (LLMs) on difficult tasks, but it also makes inference expensive because every intermediate step must be generated as a discrete token. Latent reasoning reduces visible token generation by propagating continuous states, yet replacing explicit derivations with latent computation can hurt tasks that require symbolic checking. We propose Latent-Then-Explicit Reasoning (LaTER), a two-stage paradigm that first performs bounded exploration in a continuous latent space and then switches to explicit CoT for verification and answer generation. In a training-free instantiation, LaTER projects final-layer hidden states back to the input embedding space, preserves the latent KV cache, and uses entropy and model-native stop-token probes to decide when to switch. We find that strong reasoning models already exhibit structured latent trajectories under this interface. On Qwen3-14B, training-free LaTER reduces total token usage by 16%-32% on several benchmarks while matching or improving accuracy on most of them; for example, it improves AIME 2025 from 70.0% to 73.3% while reducing tokens from 15,730 to 10,661. We further construct Latent-Switch-69K, a supervised corpus that pairs condensed solution intuitions with shortened explicit derivations. Fine-tuning with latent rollout and halting supervision yields additional gains: trained LaTER reaches 80.0% accuracy on AIME 2025, 10.0 points above the standard CoT baseline, while using 33% fewer tokens. Our code, data, and model are available at https://github.com/TioeAre/LaTER.

We identify a new dimension for enhancing rollout diversity in Group Relative Policy Optimization (GRPO) for LLMs. While GRPO relies on diverse rollouts, prevailing strategies primarily increase diversity by injecting more token-level randomness, which may introduce step-wise noise and lead to incoherent trajectories. We uncover that smaller models within the same model family inherently exhibit higher policy-level diversity, indicated by their superior pass@k relative to larger counterparts as sample counts increase. Unlike token-level noise, this diversity is temporally correlated, preserves logical consistency, and provides structured exploration signals for gradient estimation. We thus propose S2L-PO (Small-to-Large Policy Optimization), a framework that leverages fixed small models as natural explorers to train larger models. To balance exploration and exploitation, we design a progressive annealing strategy that transitions from offline small-model rollouts to the large learner's own sampling. This shift elegantly avoids mid-training performance drops caused by the small model's capacity limits, achieving faster convergence and unlocking a higher performance ceiling. S2L-PO improves accuracy on diverse mathematical reasoning benchmarks (e.g., +8.8% on AIME 24 using a 1.7B explorer to guide the 8B model) while reducing rollout compute.

Recent progress in Large Language Model (LLM) reasoning is increasingly driven by the refinement of post-training loss functions and alignment strategies. However, standard Reinforcement Learning (RL) paradigms like Group Relative Policy Optimization (GRPO) remain constrained by static uniformity: uniform prompt sampling and a fixed number of rollouts per prompt. For heterogeneous, heavy-tailed reasoning data, this creates structural inefficiencies that waste compute on already-solved patterns while under-training the long tail of hard problems. To address this, we propose Multi-Adversary Group Distributionally Robust Optimization (GDRO), an optimization-first framework that moves beyond uniform reasoning models by dynamically adapting the training distribution. We introduce an Online Difficulty Classifier that partitions prompts into dynamic pass@k difficulty groups. We then propose two independent GDRO games for post-training: (1) Prompt-GDRO, which employs an EMA-debiased multiplicative-weights bandit sampler to target the intensive difficulty margin and upweight persistently hard groups without frequency bias; and (2) Rollout-GDRO, which uses a shadow-price controller to reallocate rollouts across groups, maximizing gradient variance reduction on hard tasks under a fixed mean budget (compute-neutral). We provide no-regret guarantees for both controllers and additionally a variance-proxy analysis motivating a square-root optimal rollout allocation for Rollout-GDRO. We validate our framework on the DAPO 14.1k dataset using Qwen3-Base models. Prompt-GDRO and Rollout-GDRO achieve average relative gains of +10.6% and +10.1%, respectively, in pass@8 accuracy across 1.7B, 4B, and 8B scales compared to the GRPO baseline. Qualitative analysis shows an emergent curriculum: the adversaries shift resources to the evolving reasoning frontier, enhancing the reasoning model's performance.

Contemporary progress in large language models (LLMs) has revealed notable inferential capacities via reinforcement learning (RL) employing verifiable reward, facilitating the development of O1 and R1-like reasoning models. Directly training from base models with RL is called zero-RL. However, previous works rely upon activating LLMs' inherent capacities through fixed prompt templates. This strategy introduces substantial sampling inefficiencies for weak LLMs, as the majority of problems generate invalid outputs during accuracy-driven filtration in reasoning tasks, which causes a waste of samples. To solve this issue, we propose Cog-Rethinker, a novel hierarchical metacognitive RL framework for LLM reasoning. Our Cog-Rethinker mainly focuses on the rollout procedure in RL training. After the direct rollout, our Cog-Rethinker improves sample utilization in a hierarchical metacognitive two-stage framework. By leveraging human cognition during solving problems, firstly, it prompts policy to decompose zero-accuracy problems into subproblems to produce final reasoning results. Secondly, with zero-accuracy problems in previous rollout stage, it further prompts policy to refine these answers by referencing previous wrong solutions. Moreover, to enable cold-start of the two new reasoning patterns and maintain train-test consistency across prompt templates, our Cog-Rethinker applies supervised fine-tuning on the policy using correct samples of the two stages with direct rollout template. Experimental results demonstrate Cog-Rethinker's superior performance on various mathematical reasoning benchmarks, we also analyzed its improved sample efficiency that accelerates convergence compared to baseline methods.

On-policy distillation (OPD) transfers reasoning behavior by training a student on teacher feedback along student-generated trajectories, but standard full-rollout training ties every update to a costly completion and can over-allocate supervision to late positions with low marginal value for the current student. We revisit this assumption through the useful supervision horizon: student-induced rollouts can drift from teacher-preferred continuations, while aligned prefixes may already preserve the long-horizon OPD update direction. We propose ADWIN, an adaptive-window framework for OPD that treats rollout length as an online admissibility decision, training on short teacher-anchored prefixes while using delayed full-rollout probes to audit prefix--full alignment and adapt the next horizon with staleness control. Across math and code reasoning benchmarks in single-task, multi-task, and strong-to-weak settings, ADWIN improves the accuracy--compute trade-off over full-rollout OPD and prefix-based baselines, reducing end-to-end training cost by up to 4.1 times while achieving comparable or better accuracy.

Reinforcement learning (RL) has become a cornerstone in advancing large-scale pre-trained language models (LLMs). Successive generations, including GPT-o series, DeepSeek-R1, Kimi-K1.5, Grok 4, and GLM-4.5, have relied on large-scale RL training to enhance reasoning and coding capabilities. To meet the community's growing RL needs, numerous RL frameworks have been proposed. However, RL training remains computationally expensive, with rollout generation accounting for more than 90% of total runtime. In addition, its efficiency is often constrained by the long-tail distribution of rollout response lengths, where a few lengthy responses stall entire batches, leaving GPUs idle and underutilized. As model and rollout sizes continue to grow, this bottleneck increasingly limits scalability. To address this challenge, we propose Active Partial Rollouts in Reinforcement Learning (APRIL), which mitigates long-tail inefficiency. In the rollout phase, APRIL over-provisions rollout requests, terminates once the target number of responses is reached, and recycles incomplete responses for continuation in future steps. This strategy ensures that no rollouts are discarded while substantially reducing GPU idle time. Experiments show that APRIL improves rollout throughput by 22.5% on average (at most 44%) across commonly used RL algorithms (GRPO, DAPO, GSPO), accelerates convergence, and achieves 2.1% on average(at most 8%) higher final accuracy across tasks. Moreover, APRIL is both framework and hardware agnostic, already integrated into the slime RL framework, and deployable on NVIDIA and AMD GPUs alike. Taken together, this work unifies system-level and algorithmic considerations in proposing APRIL, with the aim of advancing RL training efficiency and inspiring further optimizations in RL systems. Our codebase is available at https://github.com/RLsys-Foundation/APRIL

Reasoning segmentation is an emerging vision-language task that requires reasoning over intricate text queries to precisely segment objects. However, existing methods typically suffer from overthinking, generating verbose reasoning chains that interfere with object localization in multimodal large language models (MLLMs). To address this issue, we propose DR^2Seg, a self-rewarding framework that improves both reasoning efficiency and segmentation accuracy without requiring extra thinking supervision. DR^2Seg employs a two-stage rollout strategy that decomposes reasoning segmentation into multimodal reasoning and referring segmentation. In the first stage, the model generates a self-contained description that explicitly specifies the target object. In the second stage, this description replaces the original complex query to verify its self-containment. Based on this design, two self-rewards are introduced to mitigate overthinking and the associated attention dispersion. Extensive experiments conducted on 3B and 7B variants of Qwen2.5-VL, as well as on both SAM2 and SAM3, demonstrate that DR^2Seg consistently improves reasoning efficiency and overall segmentation accuracy.

World models learn an internal representation of environment dynamics, enabling agents to simulate and reason about future states within a compact latent space for tasks such as planning, prediction, and inference. However, running world models rely on hevay computational cost and memory footprint, making model quantization essential for efficient deployment. To date, the effects of post-training quantization (PTQ) on world models remain largely unexamined. In this work, we present a systematic empirical study of world model quantization using DINO-WM as a representative case, evaluating diverse PTQ methods under both weight-only and joint weight-activation settings. We conduct extensive experiments on different visual planning tasks across a wide range of bit-widths, quantization granularities, and planning horizons up to 50 iterations. Our results show that quantization effects in world models extend beyond standard accuracy and bit-width trade-offs: group-wise weight quantization can stabilize low-bit rollouts, activation quantization granularity yields inconsistent benefits, and quantization sensitivity is highly asymmetric between encoder and predictor modules. Moreover, aggressive low-bit quantization significantly degrades the alignment between the planning objective and task success, leading to failures that cannot be remedied by additional optimization. These findings reveal distinct quantization-induced failure modes in world model-based planning and provide practical guidance for deploying quantized world models under strict computational constraints. The code will be available at https://github.com/huawei-noah/noah-research/tree/master/QuantWM.

Test-Time Scaling (TTS) enhances the reasoning capabilities of large language models by allocating additional inference compute to explore the solution space. However, existing parallel TTS methods typically keep branches isolated during search: intermediate discoveries remain branch-private and cannot guide other branches in time. This information isolation causes substantial redundant exploration, as branches repeatedly rediscover information already found elsewhere and require more search steps to collect complete decision information needed to reach correct answers. To bridge this gap, we propose Collaborative Parallel Thinking (CPT), a training-free inference framework that enables search-time information sharing across parallel branches. CPT extracts compact intermediate information from ongoing branches, maintains a deduplicated query-level information pool, and broadcasts pool entries through the input context, allowing each branch in subsequent search steps to reuse discoveries made by other branches rather than rediscover the same information. Empirically, experiments on HMMT and AIME benchmarks show that CPT establishes a stronger accuracy--latency Pareto frontier than strong baselines across rollout budgets and model scales, highlighting search-time collaboration as an effective direction for efficient parallel TTS.

We present FastNet version 1.0, a data-driven medium range numerical weather prediction (NWP) model based on a Graph Neural Network architecture, developed jointly between the Alan Turing Institute and the Met Office. FastNet uses an encode-process-decode structure to produce deterministic global weather predictions out to 10 days. The architecture is independent of spatial resolution and we have trained models at 1^{circ} and 0.25^{circ} resolution, with a six hour time step. FastNet uses a multi-level mesh in the processor, which is able to capture both short-range and long-range patterns in the spatial structure of the atmosphere. The model is pre-trained on ECMWF's ERA5 reanalysis data and then fine-tuned on additional autoregressive rollout steps, which improves accuracy over longer time horizons. We evaluate the model performance at 1.5^{circ} resolution using 2022 as a hold-out year and compare with the Met Office Global Model, finding that FastNet surpasses the skill of the current Met Office Global Model NWP system using a variety of evaluation metrics on a number of atmospheric variables. Our results show that both our 1^{circ} and 0.25^{circ} FastNet models outperform the current Global Model and produce results with predictive skill approaching those of other data-driven models trained on 0.25^{circ} ERA5.

We introduce LaviGen, a framework that repurposes 3D generative models for 3D layout generation. Unlike previous methods that infer object layouts from textual descriptions, LaviGen operates directly in the native 3D space, formulating layout generation as an autoregressive process that explicitly models geometric relations and physical constraints among objects, producing coherent and physically plausible 3D scenes. To further enhance this process, we propose an adapted 3D diffusion model that integrates scene, object, and instruction information and employs a dual-guidance self-rollout distillation mechanism to improve efficiency and spatial accuracy. Extensive experiments on the LayoutVLM benchmark show LaviGen achieves superior 3D layout generation performance, with 19% higher physical plausibility than the state of the art and 65% faster computation. Our code is publicly available at https://github.com/fenghora/LaviGen.

Multimodal Large Language Models (MLLMs) achieve stronger visual understanding by scaling input fidelity, yet the resulting visual token growth makes jointly sustaining high spatial resolution and long temporal context prohibitive. We argue that the bottleneck lies not in how post-encoding representations are compressed but in the volume of pixels the encoder receives, and address it with ResAdapt, an Input-side adaptation framework that learns how much visual budget each frame should receive before encoding. ResAdapt couples a lightweight Allocator with an unchanged MLLM backbone, so the backbone retains its native visual-token interface while receiving an operator-transformed input. We formulate allocation as a contextual bandit and train the Allocator with Cost-Aware Policy Optimization (CAPO), which converts sparse rollout feedback into a stable accuracy-cost learning signal. Across budget-controlled video QA, temporal grounding, and image reasoning tasks, ResAdapt improves low-budget operating points and often lies on or near the efficiency-accuracy frontier, with the clearest gains on reasoning-intensive benchmarks under aggressive compression. Notably, ResAdapt supports up to 16x more frames at the same visual budget while delivering over 15% performance gain. Code is available at https://github.com/Xnhyacinth/ResAdapt.

While reinforcement learning (RL) has empowered multi-turn reasoning agents with retrieval and tools, existing successes largely depend on extensive on-policy rollouts in high-cost, high-accuracy regimes. Under realistic resource constraints that cannot support large models or dense explorations, however, small language model agents fall into a low-cost, low-accuracy regime, where limited rollout budgets lead to sparse exploration, sparse credit assignment, and unstable training. In this work, we challenge this trade-off and show that small language models can achieve strong multi-hop reasoning under resource constraints. We introduce DAVID-GRPO, a budget-efficient RL framework that (i) stabilizes early learning with minimal supervision, (ii) assigns retrieval credit based on evidence recall, and (iii) improves exploration by resampling truncated near-miss trajectories. Evaluated on agents up to 1.5B parameters trained on only four RTX 3090 GPUs, DAVID-GRPO consistently outperforms prior RL methods designed for large-scale settings on six multi-hop QA benchmarks. These results show that with the right inductive biases, small agents can achieve low training cost with high accuracy.

Reinforcement learning, such as PPO and GRPO, has powered recent breakthroughs in LLM reasoning. Scaling rollout to sample more prompts enables models to selectively use higher-quality data for training, which can stabilize RL training and improve model performance. However, this comes at the cost of significant computational overhead. In this paper, we show that a substantial portion of this overhead can be avoided by skipping uninformative prompts before rollout. Our analysis of reward dynamics reveals a strong temporal consistency in prompt value: prompts that are uninformative in one epoch of training are likely to remain uninformative in future epochs. Based on these insights, we propose GRESO (GRPO with Efficient Selective Rollout), an online, lightweight pre-rollout filtering algorithm that predicts and skips uninformative prompts using reward training dynamics. By evaluating GRESO on a broad range of math reasoning benchmarks and models, such as Qwen2.5-Math-1.5B, DeepSeek-R1-Distill-Qwen-1.5B, and Qwen2.5-Math-7B, we show that GRESO achieves up to 2.4x wall-clock time speedup in rollout and up to 2.0x speedup in total training time without accuracy degradation.

Asynchronous reinforcement learning improves rollout throughput for large language model agents by decoupling sample generation from policy optimization, but it also introduces a critical failure mode for PPO-style off-policy correction. In heterogeneous training systems, the total importance ratio should ideally be decomposed into two semantically distinct factors: a training--inference discrepancy term that aligns inference-side and training-side distributions at the same behavior-policy version, and a policy-staleness term that constrains the update from the historical policy to the current policy. We show that practical asynchronous pipelines with delayed updates and partial rollouts often lose the required historical training-side logits, or old logits. This missing-old-logit problem entangles discrepancy repair with staleness correction, breaks the intended semantics of decoupled correction, and makes clipping and masking thresholds interact undesirably. To address this issue, we study both exact and approximate correction routes. We propose three exact old-logit acquisition strategies: snapshot-based version tracking, a dedicated old-logit model, and synchronization via partial rollout interruption, and compare their system trade-offs. From the perspective of approximate correction, we focus on preserving the benefits of decoupled correction through a more appropriate approximate policy when exact old logits cannot be recovered at low cost, without incurring extra system overhead. Following this analysis, we adopt a revised PPO-EWMA method, which achieves significant gains in both training speed and optimization performance. Code at https://github.com/millioniron/ROLL.

Reinforcement-Learning-based post-training has recently emerged as a promising paradigm for aligning text-to-image diffusion models with human preferences. In recent studies, increasing the rollout group size yields pronounced performance improvements, indicating substantial room for further alignment gains. However, scaling rollouts on large-scale foundational diffusion models (e.g., FLUX.1-12B) imposes a heavy computational burden. To alleviate this bottleneck, we explore the integration of FP4 quantization into Diffusion RL rollouts. Yet, we identify that naive quantized pipelines inherently introduce risks of performance degradation. To overcome this dilemma between efficiency and training integrity, we propose Sol-RL (Speed-of-light RL), a novel FP4-empowered Two-stage Reinforcement Learning framework. First, we utilize high-throughput NVFP4 rollouts to generate a massive candidate pool and extract a highly contrastive subset. Second, we regenerate these selected samples in BF16 precision and optimize the policy exclusively on them. By decoupling candidate exploration from policy optimization, Sol-RL integrates the algorithmic mechanisms of rollout scaling with the system-level throughput gains of NVFP4. This synergistic algorithm-hardware design effectively accelerates the rollout phase while reserving high-fidelity samples for optimization. We empirically demonstrate that our framework maintains the training integrity of BF16 precision pipeline while fully exploiting the throughput gains enabled by FP4 arithmetic. Extensive experiments across SANA, FLUX.1, and SD3.5-L substantiate that our approach delivers superior alignment performance across multiple metrics while accelerating training convergence by up to 4.64times, unlocking the power of massive rollout scaling at a fraction of the cost.

Benchmarks increasingly guide deployment, procurement and scientific screening, yet a score supports only the response it records, not necessarily the deployment action. We introduce deployment-complete benchmarking, which tests whether benchmark evidence determines a deployment action. A benchmark is complete for a claim exactly when the action is constant on each evidence fiber; mixed fibers expose missing deployment information, and completion curves quantify the evidence required to resolve ambiguity. In controlled response spaces, benchmark-channel conformal coverage of 94.98% transferred poorly to an unmeasured deployment channel (10.07%), whereas response-rank intervals achieved 94.91% coverage; even zero benchmark error certified only 45.4% of candidates at the largest residual size. Public audits revealed incompleteness, including 97.9% mixed Tox21 fibers and zero median certifiable fraction in main Matbench and JARVIS audits. In held-out replays, certify-then-acquire reduced false decisions from 1.19% to 0.027% in Tox21 and from 20.3% to 0.128% in JARVIS, while changing model choice and identifying deployment-relevant probes. Deployment-ready benchmarks should report evidence, supported actions, ambiguity and completion cost rather than scores alone.

We present WebGym, the largest-to-date open-source environment for training realistic visual web agents. Real websites are non-stationary and diverse, making artificial or small-scale task sets insufficient for robust policy learning. WebGym contains nearly 300,000 tasks with rubric-based evaluations across diverse, real-world websites and difficulty levels. We train agents with a simple reinforcement learning (RL) recipe, which trains on the agent's own interaction traces (rollouts), using task rewards as feedback to guide learning. To enable scaling RL, we speed up sampling of trajectories in WebGym by developing a high-throughput asynchronous rollout system, designed specifically for web agents. Our system achieves a 4-5x rollout speedup compared to naive implementations. Second, we scale the task set breadth, depth, and size, which results in continued performance improvement. Fine-tuning a strong base vision-language model, Qwen-3-VL-8B-Instruct, on WebGym results in an improvement in success rate on an out-of-distribution test set from 26.2% to 42.9%, significantly outperforming agents based on proprietary models such as GPT-4o and GPT-5-Thinking that achieve 27.1% and 29.8%, respectively. This improvement is substantial because our test set consists only of tasks on websites never seen during training, unlike many other prior works on training visual web agents.

Text-to-SQL technology has become crucial for translating natural language into SQL queries in various industries, enabling non-technical users to perform complex data operations. The need for accurate evaluation methods has increased as these systems have grown more sophisticated. However, we found that the Execution Accuracy (EX), the most promising evaluation metric, still shows a substantial portion of false positives and negatives compared to human evaluation. Thus, this paper introduces FLEX (False-Less EXecution), a novel approach to evaluating text-to-SQL systems using large language models (LLMs) to emulate human expert-level evaluation of SQL queries. Our method shows significantly higher agreement with human expert judgments, improving Cohen's kappa from 61 to 78.17. Re-evaluating top-performing models on the Spider and BIRD benchmarks using FLEX reveals substantial shifts in performance rankings, with an average performance decrease of 3.15 due to false positive corrections and an increase of 6.07 from addressing false negatives. This work contributes to a more accurate and nuanced evaluation of text-to-SQL systems, potentially reshaping our understanding of state-of-the-art performance in this field.

Reinforcement learning (RL) for large language models (LLMs) is dominated by the cost of rollout generation, which has motivated the use of low-precision rollouts (e.g., FP8) paired with a BF16 trainer to improve throughput and reduce memory pressure. This introduces a rollout-training mismatch that biases the policy gradient and can cause training to collapse outright on reasoning benchmarks. We show that the mismatch is non-stationary and acts as a double-edged sword: early in training it provides a stochastic exploration bonus, exposing the gradient to trajectories the trainer would otherwise under-sample, but the same perturbation transitions into a destabilizing source of bias as the policy concentrates. To solve this, we propose Adaptive Importance Sampling (AIS), a correction framework that adjusts the strength of its intervention on a per-batch basis. AIS combines three real-time diagnostics, namely weight reliability, divergence severity, and variance amplification, into a single mixing coefficient that interpolates between the uncorrected and fully importance-weighted gradients, suppressing the destabilizing component of the mismatch while preserving its exploratory benefit. We integrate AIS into GRPO and evaluate it on the diffusion-based LLaDA-8B-Instruct and the autoregressive Qwen3-8B and Qwen3.5-9B across mathematical reasoning and planning benchmarks. AIS matches the BF16 baseline on most tasks while retaining the 1.5 to 2.76x rollout speedup of FP8.

Recent studies on reasoning models explore the meta-awareness of language models, the ability to know how to think by itself. We argue that large reasoning models lack this meta-awareness property by proving severe misalignment between true rollouts and predicted meta information. We posit that aligning meta-prediction with true rollouts will lead to significant performance gains. To verify this hypothesis, we design a training pipeline that boosts Meta-Awareness via Self-Alignment (MASA), and prove that enhanced meta-awareness directly translates to improved accuracy. Unlike existing meta-cognitive reasoning models, our method does not require external training sources but leverages self-generated signals to train meta-awareness. Moreover, our method enables efficient training by i) filtering out zero-variance prompts that are either trivial or unsolvable and ii) cutting off lengthy rollouts when they are unlikely to lead to correct answers. The results are inspiring: our strategy yields significant improvements in both accuracy and training efficiency on in-domain tasks and shows strong generalization to out-of-domain benchmarks. More specifically, our method can speed up GRPO training by over 1.28x to reach the same performance, and achieve a 19.3% gain in accuracy on AIME25, and a 6.2 % average gain over six mathematics benchmarks. Training with meta-cognitive guidance enhances out-of-domain generalization, giving a 3.87 % boost on GPQA-Diamond and a 2.08 % overall accuracy gain across 13 benchmarks spanning logical, scientific, and coding domains.

Code Agent development is an extremely active research area, where a reliable performance metric is critical for tracking progress and guiding new developments. This demand is underscored by the meteoric rise in popularity of SWE-Bench. This benchmark challenges code agents to generate patches addressing GitHub issues given the full repository as context. The correctness of generated patches is then evaluated by executing a human-written test suite extracted from the repository after the issue's resolution. However, constructing benchmarks like SWE-Bench requires substantial manual effort to set up historically accurate execution environments for testing. Crucially, this severely limits the number of considered repositories, e.g., just 12 for SWE-Bench. Considering so few repositories, selected for their popularity runs the risk of leading to a distributional mismatch, i.e., the measured performance may not be representative of real-world scenarios potentially misguiding development efforts. In this work, we address this challenge and introduce SetUpAgent, a fully automated system capable of historically accurate dependency setup, test execution, and result parsing. Using SetUpAgent, we generate two new datasets: (i) SWEE-Bench an extended version of SWE-Bench encompassing hundreds of repositories, and (ii) SWA-Bench a benchmark focusing on applications rather than libraries. Comparing these datasets to SWE-Bench with respect to their characteristics and code agent performance, we find significant distributional differences, including lower issue description quality and detail level, higher fix complexity, and most importantly up to 40% lower agent success rates.

Estimating the test performance of a model, possibly under distribution shift, without having access to the ground-truth labels is a challenging, yet very important problem for the safe deployment of machine learning algorithms in the wild. Existing works mostly rely on information from either the outputs or the extracted features of neural networks to estimate a score that correlates with the ground-truth test accuracy. In this paper, we investigate -- both empirically and theoretically -- how the information provided by the gradients can be predictive of the ground-truth test accuracy even under distribution shifts. More specifically, we use the norm of classification-layer gradients, backpropagated from the cross-entropy loss after only one gradient step over test data. Our intuition is that these gradients should be of higher magnitude when the model generalizes poorly. We provide the theoretical insights behind our approach and the key ingredients that ensure its empirical success. Extensive experiments conducted with various architectures on diverse distribution shifts demonstrate that our method significantly outperforms current state-of-the-art approaches. The code is available at https://github.com/Renchunzi-Xie/GdScore

Reinforcement Learning with Verifiable Rewards (RLVR) has become the leading paradigm for enhancing reasoning in Large Language Models (LLMs). However, standard RLVR algorithms suffer from a well-documented pathology: while they improve Pass@1 accuracy through sharpened sampling, they simultaneously narrow the model's reasoning boundary and reduce generation diversity. We identify a root cause that existing methods overlook: the uniform penalization of errors. Current approaches -- whether data-filtering methods that select prompts by difficulty, or advantage normalization schemes -- treat all incorrect rollouts within a group identically. We show that this uniformity allows overconfident errors (incorrect reasoning paths that the RL process has spuriously reinforced) to persist and monopolize probability mass, ultimately suppressing valid exploratory trajectories. To address this, we propose the Asymmetric Confidence-aware Error Penalty (ACE). ACE introduces a per-rollout confidence shift metric, c_i = log(pi_theta(y_i|x) / pi_ref(y_i|x)), to dynamically modulate negative advantages. Theoretically, we demonstrate that ACE's gradient can be decomposed into the gradient of a selective regularizer restricted to overconfident errors, plus a well-characterized residual that partially moderates the regularizer's strength. We conduct extensive experiments fine-tuning Qwen2.5-Math-7B, Qwen3-8B-Base, and Llama-3.1-8B-Instruct on the DAPO-Math-17K dataset using GRPO and DAPO within the VERL framework. Evaluated on MATH-500 and AIME 2025, ACE composes seamlessly with existing methods and consistently improves the full Pass@k spectrum across all three model families and benchmarks.

Reinforcement learning (RL) has become a representative post-training paradigm for LLMs, enabling strong reasoning and agentic capabilities. However, rollout generation remains a dominant latency bottleneck because autoregressive sampling decodes responses sequentially and a small number of long-tailed generations often determine completion time. Speculative decoding (SD) offers a natural way to address this bottleneck, as it is a well-established technique for serving fixed LLMs that reduces latency by rapidly drafting tokens and accepting them through parallel verification while preserving the target-model distribution. However, its practical speedups do not directly carry over to RL rollouts: (i) the evolving target policy makes any fixed drafter increasingly mismatched with the policy's output distribution; and (ii) active batch sizes shrink throughout rollout decoding, shifting decoding from compute-bound to memory-bound regimes where parallel verification can exploit underutilized compute. Therefore, accelerating RL rollouts requires both a drafter that remains effective under long, high-temperature generations from an evolving policy and system-aware use of SD that avoids compute-bound regimes. We present EfficientRollout, a system-aware self-SD framework designed to address this gap for RL rollouts. EfficientRollout induces a quantized drafter from the target model (i.e. self-speculative decoding), keeping it coupled to the evolving policy without separate drafter pretraining or online adaptation. It further coordinates a system-aware SD toggle policy with acceptance-aware draft-length adaptation, enabling speculation only in beneficial regimes while matching the drafting budget to evolving drafter quality. EfficientRollout reduces rollout and end-to-end latency by up to 19.6% and 12.7%, respectively, over an accelerated AR rollout baseline, while preserving final model quality.

The goal of this paper is to improve the performance and reliability of vision-language-action (VLA) models through iterative online interaction. Since collecting policy rollouts in the real world is expensive, we investigate whether a learned simulator-specifically, an action-conditioned video generation model-can be used to generate additional rollout data. Unfortunately, existing world models lack the physical fidelity necessary for policy improvement: they are predominantly trained on demonstration datasets that lack coverage of many different physical interactions (particularly failure cases) and struggle to accurately model small yet critical physical details in contact-rich object manipulation. We propose a simple iterative improvement algorithm that uses real-world roll-out data to improve the fidelity of the world model, which can then, in turn, be used to generate supplemental synthetic data for improving the VLA model. In our experiments on a real robot, we use this approach to improve the performance of a state-of-the-art VLA model on multiple downstream tasks. We achieve a 39.2% absolute success rate improvement over the base policy and 11.6% improvement from training with the generated synthetic rollouts. Videos can be found at this anonymous website: https://sites.google.com/view/vla-w

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a key ingredient for unlocking complex reasoning capabilities in large language models. Recent work ProRL has shown promise in scaling RL by increasing the number of training steps. However, performance plateaus after thousands of steps, with clear diminishing returns from allocating more computation to additional training. In this work, we investigate a complementary paradigm for scaling RL, BroR-Lincreasing the number of rollouts per example to hundreds to exhaustively Broaden exploration, which yields continuous performance gains beyond the saturation point observed in ProRL when scaling the number of training steps. Our approach is motivated by a mass balance equation analysis allowing us to characterize the rate of change in probability mass for correct and incorrect tokens during the reinforcement process. We show that under a one-step RL assumption, sampled rollout tokens always contribute to correct-mass expansion, while unsampled tokens outside rollouts may lead to gains or losses depending on their distribution and the net reward balance. Importantly, as the number of rollouts per example N increases, the effect of unsampled terms diminishes, ensuring overall correct-mass expansion. To validate our theoretical analysis, we conduct simulations under more relaxed conditions and find that a sufficiently large rollout size N-corresponding to ample exploration-guarantees an increase in the probability mass of all correct tokens. Empirically, BroRL revives models saturated after 3K ProRL training steps and demonstrates robust, continuous improvement, achieving state-of-the-art results for the 1.5B model across diverse benchmarks.

Reinforcement learning (RL) has become an effective approach for fine-tuning large language models (LLMs), particularly to enhance their reasoning capabilities. However, RL fine-tuning remains highly resource-intensive, and existing work has largely overlooked the problem of data efficiency. In this paper, we propose two techniques to improve data efficiency in LLM RL fine-tuning: difficulty-targeted online data selection and rollout replay. We introduce the notion of adaptive difficulty to guide online data selection, prioritizing questions of moderate difficulty that are more likely to yield informative learning signals. To estimate adaptive difficulty efficiently, we develop an attention-based framework that requires rollouts for only a small reference set of questions. The adaptive difficulty of the remaining questions is then estimated based on their similarity to this set. To further reduce rollout cost, we introduce a rollout replay mechanism inspired by experience replay in traditional RL. This technique reuses recent rollouts, lowering per-step computation while maintaining stable updates. Experiments across 6 LLM-dataset combinations show that our method reduces RL fine-tuning time by 23% to 62% while reaching the same level of performance as the original GRPO algorithm. Our code is available at https://github.com/ASTRAL-Group/data-efficient-llm-rl.

Self-correction is essential for solving complex reasoning problems in vision-language models (VLMs). However, existing reinforcement learning (RL) methods struggle to learn it, as effective self-correction behaviors emerge only rarely, making learning signals extremely sparse. To address this challenge, we propose correction-specific rollouts (Octopus), an RL rollout augmentation framework that synthesizes dense self-correction examples by recombining existing rollouts. This augmentation simultaneously improves sample efficiency due to rollout reuse and stabilizes RL optimization through balanced supervision. Furthermore, we introduce a response-masking strategy that decouples self-correction from direct reasoning, avoiding signal conflicts and enabling both behaviors to be learned effectively. Building on this, we introduce Octopus-8B, a reasoning VLM with controllable self-correction capability. Across 7 benchmarks, it achieves SoTA performance among open-source VLMs, outperforming the best RLVR baseline by 1.0 score while requiring only 0.72times training time per step.

Reinforcement learning (RL) for large language models (LLMs) is increasingly bottlenecked by rollout (generation), where long output sequence lengths make attention and KV-cache memory dominate end-to-end step time. FP8 offers an attractive lever for accelerating RL by reducing compute cost and memory traffic during rollout, but applying FP8 in RL introduces unique engineering and algorithmic challenges: policy weights change every step (requiring repeated quantization and weight synchronization into the inference engine) and low-precision rollouts can deviate from the higher-precision policy assumed by the trainer, causing train-inference mismatch and potential instability. This report presents a practical FP8 rollout stack for LLM RL, implemented in the veRL ecosystem with support for common training backends (e.g., FSDP/Megatron-LM) and inference engines (e.g., vLLM/SGLang). We (i) enable FP8 W8A8 linear-layer rollout using blockwise FP8 quantization, (ii) extend FP8 to KV-cache to remove long-context memory bottlenecks via per-step QKV scale recalibration, and (iii) mitigate mismatch using importance-sampling-based rollout correction (token-level TIS/MIS variants). Across dense and MoE models, these techniques deliver up to 44% rollout throughput gains while preserving learning behavior comparable to BF16 baselines.

Reinforcement learning (RL) has become essential for post-training large language models (LLMs) in reasoning tasks. While scaling rollouts can stabilize training and enhance performance, the computational overhead is a critical issue. In algorithms like GRPO, multiple rollouts per prompt incur prohibitive costs, as a large portion of prompts provide negligible gradients and are thus of low utility. To address this problem, we investigate how to select high-utility prompts before the rollout phase. Our experimental analysis reveals that sample utility is non-uniform and evolving: the strongest learning signals concentrate at the ``learning edge", the intersection of intermediate difficulty and high uncertainty, which shifts as training proceeds. Motivated by this, we propose HIVE (History-Informed and online-VErified prompt selection), a dual-stage framework for data-efficient RL. HIVE utilizes historical reward trajectories for coarse selection and employs prompt entropy as a real-time proxy to prune instances with stale utility. By evaluating HIVE across multiple math reasoning benchmarks and models, we show that HIVE yields significant rollout efficiency without compromising performance.

Researchers have proposed numerous text-to-SQL techniques to streamline data analytics and accelerate the development of data-driven applications. To compare these techniques and select the best one for deployment, the community depends on public benchmarks and their leaderboards. Since these benchmarks heavily rely on human annotations during question construction and answer evaluation, the validity of the annotations is crucial. In this paper, we conduct an empirical study that (i) benchmarks annotation error rates for two widely used text-to-SQL benchmarks, BIRD and Spider 2.0-Snow, and (ii) corrects a subset of the BIRD development (Dev) set to measure the impact of annotation errors on text-to-SQL agent performance and leaderboard rankings. Through expert analysis, we show that BIRD Mini-Dev and Spider 2.0-Snow have error rates of 52.8% and 62.8%, respectively. We re-evaluate all 16 open-source agents from the BIRD leaderboard on both the original and the corrected BIRD Dev subsets. We show that performance changes range from -7% to 31% (in relative terms) and rank changes range from -9 to +9 positions. We further assess whether these impacts generalize to the full BIRD Dev set. We find that the rankings of agents on the uncorrected subset correlate strongly with those on the full Dev set (Spearman's r_s=0.85, p=3.26e-5), whereas they correlate weakly with those on the corrected subset (Spearman's r_s=0.32, p=0.23). These findings show that annotation errors can significantly distort reported performance and rankings, potentially misguiding research directions or deployment choices. Our code and data are available at https://github.com/uiuc-kang-lab/text_to_sql_benchmarks.

Execution Accuracy (EX), the widely used metric for evaluating the effectiveness of Natural Language to SQL (NL2SQL) solutions, is becoming increasingly unreliable. It is sensitive to syntactic variation, ignores that questions may admit multiple interpretations, and is easily misled by erroneous ground-truth SQL. To address this, we introduce ROSE, an intent-centered metric that focuses on whether the predicted SQL answers the question, rather than consistency with the ground-truth SQL under the reference-dependent paradigm. ROSE employs an adversarial Prover-Refuter cascade: SQL Prover assesses the semantic correctness of a predicted SQL against the user's intent independently, while Adversarial Refuter uses the ground-truth SQL as evidence to challenge and refine this judgment. On our expert-aligned validation set ROSE-VEC, ROSE achieves the best agreement with human experts, outperforming the next-best metric by nearly 24% in Cohen's Kappa. We also conduct a largescale re-evaluation of 19 NL2SQL methods, revealing four valuable insights. We release ROSE and ROSE-VEC to facilitate more reliable NL2SQL research.

Large Language Models (LLMs) increasingly rely on reinforcement learning with verifiable rewards (RLVR) to elicit reliable chain-of-thought reasoning. However, the training process remains bottlenecked by the computationally expensive rollout stage. Existing acceleration methods-such as parallelization, objective- and data-driven modifications, and replay buffers-either incur diminishing returns, introduce bias, or overlook redundancy across iterations. We identify that rollouts from consecutive training epochs frequently share a large portion of overlapping segments, wasting computation. To address this, we propose SPEC-RL, a novel framework that integrates SPECulative decoding with the RL rollout process. SPEC-RL reuses prior trajectory segments as speculative prefixes and extends them via a draft-and-verify mechanism, avoiding redundant generation while ensuring policy consistency. Experiments on diverse math reasoning and generalization benchmarks, including GSM8K, MATH-500, OlympiadBench, MMLU-STEM, and others, demonstrate that SPEC-RL reduces rollout time by 2-3x without compromising policy quality. As a purely rollout-stage enhancement, SPEC-RL integrates seamlessly with mainstream algorithms (e.g., PPO, GRPO, DAPO), offering a general and practical path to scale RLVR for large reasoning models. Our code is available at https://github.com/ShopeeLLM/Spec-RL

Test-time reinforcement learning generates multiple candidate answers via repeated rollouts and performs online updates using pseudo-labels constructed by majority voting. To reduce overhead and improve exploration, prior work introduces tree structured rollouts, which share reasoning prefixes and branch at key nodes to improve sampling efficiency. However, this paradigm still faces two challenges: (1) high entropy branching can trigger rollout collapse, where the branching budget concentrates on a few trajectories with consecutive high-entropy segments, rapidly reducing the number of effective branches; (2) early pseudo-labels are noisy and biased, which can induce self-reinforcing overfitting, causing the policy to sharpen prematurely and suppress exploration. To address these issues, we propose Entropy Confidence Hybrid Group Relative Policy Optimization (ECHO). During rollout, ECHO jointly leverages local entropy and group level confidence to adaptively control branch width, and further introduces online confidence-based pruning to terminate persistently low confidence branches, avoiding high entropy traps and mitigating collapse. During policy updates, ECHO employs confidence adaptive clipping and an entropy confidence hybrid advantage shaping approach to enhance training robustness and mitigate early stage bias. Experiments demonstrate that ECHO achieves consistent gains on multiple mathematical and visual reasoning benchmarks, and generalizes more effectively under a limited rollout budget.

Reinforcement learning with verifiable rewards (RLVR) for Large Reasoning Models hinges on baseline estimation for variance reduction, but existing approaches pay a heavy price: PPO requires a policy-model scale critic, while GRPO needs multiple rollouts per prompt to keep its empirical group mean stable. We introduce Policy Optimization with Internal State Value Estimation), which obtains a baseline at negligible cost by using the policy model's internal signals already computed during the policy forward pass. A lightweight probe predicts the expected verifiable reward from the hidden states of the prompt and generated trajectory, as well as token-entropy statistics, and is trained online alongside the policy. To preserve gradient unbiasedness despite using trajectory-conditioned features, we introduce a cross-rollout construction that predicts each rollout's value from an independent rollout's internal states. Because POISE estimates prompt value using only a single rollout, it enables higher prompt diversity for a fixed compute budget during training. This reduces gradient variance for more stable learning and also eliminates the compute overhead of sampling costs for detecting zero-advantage prompts. On Qwen3-4B and DeepSeek-R1-Distill-Qwen-1.5B across math reasoning benchmarks, POISE matches DAPO while requiring less compute. Moreover, its value estimator shows similar performance to a separate LLM-scale value model and generalizes to various verifiable tasks. By leveraging the model's own internal representations, POISE enables more stable and efficient policy optimization.

Financial institutions deploy Large Language Models (LLMs) for reconciliations, regulatory reporting, and client communications, but nondeterministic outputs (output drift) undermine auditability and trust. We quantify drift across five model architectures (7B-120B parameters) on regulated financial tasks, revealing a stark inverse relationship: smaller models (Granite-3-8B, Qwen2.5-7B) achieve 100% output consistency at T=0.0, while GPT-OSS-120B exhibits only 12.5% consistency (95% CI: 3.5-36.0%) regardless of configuration (p<0.0001, Fisher's exact test). This finding challenges conventional assumptions that larger models are universally superior for production deployment. Our contributions include: (i) a finance-calibrated deterministic test harness combining greedy decoding (T=0.0), fixed seeds, and SEC 10-K structure-aware retrieval ordering; (ii) task-specific invariant checking for RAG, JSON, and SQL outputs using finance-calibrated materiality thresholds (plus or minus 5%) and SEC citation validation; (iii) a three-tier model classification system enabling risk-appropriate deployment decisions; and (iv) an audit-ready attestation system with dual-provider validation. We evaluated five models (Qwen2.5-7B via Ollama, Granite-3-8B via IBM watsonx.ai, Llama-3.3-70B, Mistral-Medium-2505, and GPT-OSS-120B) across three regulated financial tasks. Across 480 runs (n=16 per condition), structured tasks (SQL) remain stable even at T=0.2, while RAG tasks show drift (25-75%), revealing task-dependent sensitivity. Cross-provider validation confirms deterministic behavior transfers between local and cloud deployments. We map our framework to Financial Stability Board (FSB), Bank for International Settlements (BIS), and Commodity Futures Trading Commission (CFTC) requirements, demonstrating practical pathways for compliance-ready AI deployments.

For deploying foundation models, practitioners increasingly need prescriptive scaling laws: given a pre training compute budget, what downstream accuracy is attainable with contemporary post training practice, and how stable is that mapping as the field evolves? Using large scale observational evaluations with 5k observational and 2k newly sampled data on model performance, we estimate capability boundaries, high conditional quantiles of benchmark scores as a function of log pre training FLOPs, via smoothed quantile regression with a monotone, saturating sigmoid parameterization. We validate the temporal reliability by fitting on earlier model generations and evaluating on later releases. Across various tasks, the estimated boundaries are mostly stable, with the exception of math reasoning that exhibits a consistently advancing boundary over time. We then extend our approach to analyze task dependent saturation and to probe contamination related shifts on math reasoning tasks. Finally, we introduce an efficient algorithm that recovers near full data frontiers using roughly 20% of evaluation budget. Together, our work releases the Proteus 2k, the latest model performance evaluation dataset, and introduces a practical methodology for translating compute budgets into reliable performance expectations and for monitoring when capability boundaries shift across time.

Context: Code reviews are crucial for software quality. Recent AI advances have allowed large language models (LLMs) to review and fix code; now, there are tools that perform these reviews. However, their reliability and accuracy have not yet been systematically evaluated. Objective: This study compares different LLMs' performance in detecting code correctness and suggesting improvements. Method: We tested GPT4o and Gemini 2.0 Flash on 492 AI generated code blocks of varying correctness, along with 164 canonical code blocks from the HumanEval benchmark. To simulate the code review task objectively, we expected LLMs to assess code correctness and improve the code if needed. We ran experiments with different configurations and reported on the results. Results: With problem descriptions, GPT4o and Gemini 2.0 Flash correctly classified code correctness 68.50% and 63.89% of the time, respectively, and corrected the code 67.83% and 54.26% of the time for the 492 code blocks of varying correctness. Without problem descriptions, performance declined. The results for the 164 canonical code blocks differed, suggesting that performance depends on the type of code. Conclusion: LLM code reviews can help suggest improvements and assess correctness, but there is a risk of faulty outputs. We propose a process that involves humans, called the "Human in the loop LLM Code Review" to promote knowledge sharing while mitigating the risk of faulty outputs.

Reinforcement learning with verifiable rewards (RLVR) has emerged as a promising paradigm for enhancing the reasoning capabilities of large language models (LLMs). In this context, models explore reasoning trajectories and exploit rollouts with correct answers as positive signals for policy optimization. However, these rollouts might involve flawed patterns such as answer-guessing and jump-in-reasoning. Such flawed-positive rollouts are rewarded identically to fully correct ones, causing policy models to internalize these unreliable reasoning patterns. In this work, we first conduct a systematic study of flawed-positive rollouts in RL and find that they enable rapid capability gains during the early optimization stage, while constraining reasoning capability later by reinforcing unreliable patterns. Building on these insights, we propose Flawed-Aware Policy Optimization (FAPO), which presents a parameter-free reward penalty for flawed-positive rollouts, enabling the policy to leverage them as useful shortcuts in the warm-up stage, securing stable early gains, while gradually shifting optimization toward reliable reasoning in the later refinement stage. To accurately and comprehensively detect flawed-positive rollouts, we introduce a generative reward model (GenRM) with a process-level reward that precisely localizes reasoning errors. Experiments show that FAPO is effective in broad domains, improving outcome correctness, process reliability, and training stability without increasing the token budget.

Action-conditioned robot world models generate future video frames of the manipulated scene given a robot action sequence, offering a promising alternative for simulating tasks that are difficult to model with traditional physics engines. However, these models are optimized for short-term prediction and break down when deployed autoregressively: each predicted clip feeds back as context for the next, causing errors to compound and visual quality to rapidly degrade. We address this through the following contributions. First, we introduce a reinforcement learning (RL) post-training scheme that trains the world model on its own autoregressive rollouts rather than on ground-truth histories. We achieve this by adapting a recent contrastive RL objective for diffusion models to our setting and show that its convergence guarantees carry over exactly. Second, we design a training protocol that generates and compares multiple candidate variable-length futures from the same rollout state, reinforcing higher-fidelity predictions over lower-fidelity ones. Third, we develop efficient, multi-view visual fidelity rewards that combine complementary perceptual metrics across camera views and are aggregated at the clip level for dense, low-variance training signal. Fourth, we show that our approach establishes a new state-of-the-art for rollout fidelity on the DROID dataset, outperforming the strongest baseline on all metrics (e.g., LPIPS reduced by 14% on external cameras, SSIM improved by 9.1% on the wrist camera), winning 98% of paired comparisons, and achieving an 80% preference rate in a blind human study.

Reinforcement Learning with Verifiable Reward (RLVR) has emerged as a powerful paradigm for unlocking reasoning capabilities in large language models, yet its full potential is hindered by two under-explored dimensions: Depth-the hardest problem a model can sample; Breadth-the number of instances consumed in a single iteration. We dissect the popular GRPO algorithm and reveal a systematic bias: the cumulative-advantage disproportionately weights samples with medium accuracy, while down-weighting the low-accuracy instances that are crucial for pushing reasoning boundaries. To rectify the depth neglect, we introduce Difficulty Adaptive Rollout Sampling (DARS), which re-weights hard problems through targeted multi-stage rollouts, thereby increasing the number of positive rollouts for hard problems. Empirically, naively enlarging rollout size only accelerates convergence and even hurts Pass@K. Our DARS, in contrast, delivers consistent Pass@K gains without extra inference cost at convergence. Just as we adaptively expanded the depth of exploration, we now ask whether aggressively scaling the breadth of training data can further amplify reasoning gains. To this end, we intensely scale batch size and replace PPO's mini-batch iterations with full-batch updates over multiple epochs. Increasing breadth significantly enhances Pass@1 performance. Large-breadth training sustains high token-level entropy, indicating continued exploration and reduced gradient noise. We further present DARS-B, which augments DARS with large breadth, and demonstrate simultaneous gains in Pass@K and Pass@1. The results confirm that breadth and adaptive exploration across depth operate as orthogonal dimensions in RLVR, which are key to unleashing the reasoning power of RLVR.

When a large language model under reinforcement learning commits a wrong reasoning step early in a trajectory, standard algorithms force it to keep generating until the maximum horizon, spending compute on tokens that never receive positive reward and polluting advantage estimates with post-failure noise. We propose ESPO (Early-Stopping Proximal Policy Optimization), which detects trajectory failure on-the-fly and terminates rollouts early. At each generation step, ESPO computes a surrogate regret using only the logits already computed during sampling, and terminates when the smoothed cumulative regret significantly exceeds its estimated values. Truncated trajectories are treated as absorbing failure states with a terminal reward, concentrating negative temporal-difference (TD) errors near the detected failure step without any additional reward model or human annotation. On DeepSeek-R1-Distill-Qwen-7B trained for mathematical reasoning, ESPO surpasses PPO on AIME~2024 (46.28% vs. 45.25%), AMC~2023 (85.83% vs. 82.94%), and MATH-500 (87.42% vs. 85.43%), while saving more than 20% rollout tokens cumulatively.

Graphical User Interface (GUI) agents have demonstrated remarkable progress in automating complex user interface interactions through reinforcement learning. However, current approaches face a fundamental dilemma: offline RL enables stable training on pre-collected trajectories, but struggles with multi-step task execution for lack of trajectory-level reward signals; online RL captures these signals through environment interaction, but suffers from sparse rewards and prohibitive deployment costs. To address it, we present Semi-online Reinforcement Learning, a novel paradigm that simulates online RL on offline trajectories. During each rollout process, we preserve the original model output within the multi-turn dialogue, where a Patch Module adaptively recovers the divergence between rollout and expert trajectories. To capture long-term training signals, Semi-online RL introduces discounted future returns into the reward computation and optimizes the policy with weighted step-level and episode-level advantages. We further introduce Semi-Online Performance (SOP), a metric that aligns better with true online performance, serving as a practical and effective proxy for real-world evaluation. Experiments show that ours Semi-online RL achieves SOTA performance among 7B models across four dynamic benchmarks, with significant gains over the base model (e.g., +12.0% on AndroidWorld, +23.8% on AITW), demonstrating significant progress in bridging the gap between offline training efficiency and online multi-turn reasoning. The code is available at https://github.com/X-PLUG/MobileAgent/tree/main/UI-S1.

While there are numerous benchmarks comparing the performance of modern language models (LMs), end-task evaluations often conflate notions of *factual accuracy* ("truth") and *reasoning ability* ("rationality", or "honesty" in the sense of correctly reporting implications of beliefs). Our goal is a dataset that clearly distinguishes these two notions. Our approach is to leverage and extend a collection of human-annotated *entailment trees*, engineered to express both good and bad chains of reasoning, and using a mixture of true and false facts, in particular including counterfactual examples, to avoid belief bias (also known as the "content effect"). The resulting dataset, called BaRDa, contains 3000 entailments (1787 valid, 1213 invalid), using 6681 true and 2319 false statements. Testing on four GPT-series models, GPT3(curie)/GPT3(davinici)/3.5/4, we find factual accuracy (truth) scores of 74.1/80.6/82.6/87.1 and reasoning accuracy scores of 63.1/78.0/71.8/79.2. This shows the clear progression of models towards improved factual accuracy and entailment reasoning, and the dataset provides a new benchmark that more cleanly separates and quantifies these two notions.

Reference-free faithfulness metrics verify each atomic claim a model makes against ground truth, and are increasingly used to evaluate grounded generation. We show they share a blind spot: they measure only precision -- are the stated claims supported? -- and therefore reward abstention, since a model can score near-perfect faithfulness by saying almost nothing. We make this measurable using Formula 1 telemetry, a domain where strategic ground truth is derived deterministically and, crucially, completely: for each decision we know the full set of facts that mattered. This completeness -- absent in open-domain faithfulness benchmarks -- lets us measure recall (coverage of the relevant facts) exactly, alongside precision. On a multilingual (EN/ES/PT) benchmark of 7,253 decision instances spanning 150 races, the most precise frontier model covers under half of the relevant facts and ranks last by F1, so requiring coverage reorders the systems; the same effect reappears in a second complete-oracle domain (NOAA weather forecasts). A prompt ablation shows the low coverage is not an under-prompting artifact: explicitly asking models to be thorough does not close the gap. We pair faithfulness with coverage into a single score, validate the metric (controlled perturbation; agreement across a model-free regex extractor and a cross-family LLM extractor, system-level Spearman 1.0), and give a verifier-guided generation method that improves precision and recall without references. We release the benchmark, structured annotations, metric, baselines, and an interactive demo.

Group-based policy optimization methods like GRPO and GSPO have become standard for training multimodal models, leveraging group-wise rollouts and relative advantage estimation. However, they suffer from a critical gradient vanishing problem when all responses within a group receive identical rewards, causing advantage estimates to collapse and training signals to diminish. Existing attempts to mitigate this issue fall into two paradigms: filtering-based and sampling-based methods. Filtering-based methods first generate rollouts broadly and then retroactively filter out uninformative groups, leading to substantial computational overhead. Sampling-based methods proactively select effective samples before rollout but rely on static criteria or prior dataset knowledge, lacking real-time adaptability. To address these issues, we propose VADE, a Variance-Aware Dynamic sampling framework via online sample-level difficulty Estimation. Our framework integrates three key components: online sample-level difficulty estimation using Beta distributions, a Thompson sampler that maximizes information gain through the estimated correctness probability, and a two-scale prior decay mechanism that maintains robust estimation under policy evolution. This three components design enables VADE to dynamically select the most informative samples, thereby amplifying training signals while eliminating extra rollout costs. Extensive experiments on multimodal reasoning benchmarks show that VADE consistently outperforms strong baselines in both performance and sample efficiency, while achieving a dramatic reduction in computational overhead. More importantly, our framework can serves as a plug-and-play component to be seamlessly integrated into existing group-based RL algorithms. Code and models are available at https://VADE-RL.github.io.

Publishing a large language model (LLM) benchmark on the Internet risks contaminating future LLMs: the benchmark may be unintentionally (or intentionally) used to train or select a model. A common mitigation is to keep the benchmark private and let participants submit their models or predictions to the organizers. However, this strategy will require trust in a single organization and still permits test-set overfitting through repeated queries. To overcome this issue, we propose a way to publish benchmarks without completely disclosing the ground-truth answers to the questions, while still maintaining the ability to openly evaluate LLMs. The main underlying idea is to reduces the best possible accuracy, i.e., Bayes accuracy, by injecting randomness to the answers by preparing several logically correct answers, and only include one of them as the solution in the benchmark. Not only is this helpful to keep us from disclosing the ground truth, but this also offers a test for detecting data contamination. In principle, even fully capable models should not surpass the Bayes accuracy. If a model surpasses this ceiling despite this expectation, this is a strong signal of data contamination. We present experimental evidence that our method can detect data contamination accurately on a wide range of benchmarks, models, and training methodologies.

Test-time reinforcement learning (TTRL) enables large language models (LLMs) to self-improve on unlabeled inputs, but its effectiveness critically depends on how reward signals are estimated without ground-truth supervision. Most existing TTRL methods rely on majority voting (MV) over rollouts to produce deterministic rewards, implicitly assuming that the majority rollout provides a reliable learning signal. We show that this assumption is fragile: MV reduces the rollout distribution into a single outcome, discarding information about non-majority but correct actions candidates, and yields systematically biased reward estimates. To address this, we propose Distribution-AwareReward Estimation (DARE), which shifts reward estimation from a single majority outcome to the full empirical rollout distribution. DARE further augments this distribution-based reward with an exploration bonus and a distribution pruning mechanism for non-majority rollout exploration and reward denoise, yielding a more informative and robust reward estimation. Extensive experiments on challenging reasoning benchmarks show that DARE improves optimization stability and final performance over recent baselines, achieving relative improvements of 25.3% on challenging AIME 2024 and 5.3% on AMC.

To combine the advantages of Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), recent methods have integrated ''hints'' into post-training, which are prefix segments of complete reasoning trajectories, aiming for powerful knowledge expansion and reasoning generalization. However, existing hint-based RL methods typically ignore difficulty when scheduling hint ratios and estimating relative advantages, leading to unstable learning and excessive imitation of off-policy hints. In this work, we propose ADHint, which treats difficulty as a key factor in both hint-ratio schedule and relative-advantage estimation to achieve a better trade-off between exploration and imitation. Specifically, we propose Adaptive Hint with Sample Difficulty Prior, which evaluates each sample's difficulty under the policy model and accordingly schedules an appropriate hint ratio to guide its rollouts. We also introduce Consistency-based Gradient Modulation and Selective Masking for Hint Preservation to modulate token-level gradients within hints, preventing biased and destructive updates. Additionally, we propose Advantage Estimation with Rollout Difficulty Posterior, which leverages the relative difficulty of rollouts with and without hints to estimate their respective advantages, thereby achieving more balanced updates. Extensive experiments across diverse modalities, model scales, and domains demonstrate that ADHint delivers superior reasoning ability and out-of-distribution generalization, consistently surpassing existing methods in both pass@1 and avg@8. Our code and dataset will be made publicly available upon paper acceptance.

RL post-training of frontier language models is increasingly bottlenecked by autoregressive rollout generation, making rollout acceleration a central systems challenge. Many existing efficiency methods improve throughput by changing the rollout or optimization regime, for example, through off-policy execution, replay, or lower-precision generation. We study speculative decoding as a lossless acceleration primitive for RL rollouts that preserves the target model's output distribution. We implement speculative decoding in NeMo-RL with a vLLM backend, supporting both synchronous and asynchronous pipelines and enabling speculation during RL rollouts. This benefit is realizable across speculation mechanisms, such as pretrained MTP heads, small external draft models or even techniques such as Eagle3, which are traditionally applied after RL phase. This yields a deployment path for state-of-the-art speculative decoding inside RL training. In a reasoning post-training workload at 8B scale under synchronous RL, speculative decoding improves rollout throughput by 1.8x. Using a high-fidelity performance simulator, we project that combining speculative decoding with asynchronous RL yields up to 2.5x end-to-end training speedup at 235B scale.

Large Language Models (LLMs) are being increasingly used in software engineering tasks, with an increased focus on bug report resolution over the past year. However, most proposed systems fail to properly handle uncertain or incorrect inputs and outputs. Existing LLM-based tools and coding agents respond to every issue and generate a patch for every case, even when the input is vague or their own output is incorrect. There are no mechanisms in place to abstain when confidence is low. This leads to unreliable behaviour, such as hallucinated code changes or responses based on vague issue reports. We introduce BouncerBench, a benchmark that evaluates whether LLM-based software agents can refuse to act when inputs are ill-defined or refuse to respond when their own outputs are likely to be incorrect. Unlike prior benchmarks that implicitly incentivize models to generate responses even when uncertain, BouncerBench aims to improve precision by targeting two overlooked failure points: (1) vague or underspecified issue descriptions in tickets and (2) logically or functionally incorrect code patches created by the system. It measures whether proposed systems can distinguish actionable issues from vague tickets and valid patches from untrustworthy ones. We also implement a basic input and output bouncer, evaluating how well current LLMs can abstain when needed. Our results show that most models fail to abstain from underspecified inputs or incorrect outputs. Hence, we conclude that there is significant room for improvement before LLMs can be trusted to make correct decisions and recommendations in real-world software engineering workflows. BouncerBench provides a first step toward evaluating and building more cautious, trustworthy code agents. The replication package, dataset, and leaderboard can be found at bouncerbench.com

In this paper, we introduce OSUniverse: a benchmark of complex, multimodal desktop-oriented tasks for advanced GUI-navigation AI agents that focuses on ease of use, extensibility, comprehensive coverage of test cases, and automated validation. We divide the tasks in increasing levels of complexity, from basic precision clicking to multistep, multiapplication tests requiring dexterity, precision, and clear thinking from the agent. In version one of the benchmark, presented here, we have calibrated the complexity of the benchmark test cases to ensure that the SOTA (State of the Art) agents (at the time of publication) do not achieve results higher than 50%, while the average white collar worker can perform all these tasks with perfect accuracy. The benchmark can be scored manually, but we also introduce an automated validation mechanism that has an average error rate less than 2%. Therefore, this benchmark presents solid ground for fully automated measuring of progress, capabilities and the effectiveness of GUI-navigation AI agents over the short and medium-term horizon. The source code of the benchmark is available at https://github.com/agentsea/osuniverse.

Best-of-N (BoN) Average Displacement Error (ADE)/ Final Displacement Error (FDE) is the most used metric for evaluating trajectory prediction models. Yet, the BoN does not quantify the whole generated samples, resulting in an incomplete view of the model's prediction quality and performance. We propose a new metric, Average Mahalanobis Distance (AMD) to tackle this issue. AMD is a metric that quantifies how close the whole generated samples are to the ground truth. We also introduce the Average Maximum Eigenvalue (AMV) metric that quantifies the overall spread of the predictions. Our metrics are validated empirically by showing that the ADE/FDE is not sensitive to distribution shifts, giving a biased sense of accuracy, unlike the AMD/AMV metrics. We introduce the usage of Implicit Maximum Likelihood Estimation (IMLE) as a replacement for traditional generative models to train our model, Social-Implicit. IMLE training mechanism aligns with AMD/AMV objective of predicting trajectories that are close to the ground truth with a tight spread. Social-Implicit is a memory efficient deep model with only 5.8K parameters that runs in real time of about 580Hz and achieves competitive results. Interactive demo of the problem can be seen at https://www.abduallahmohamed.com/social-implicit-amdamv-adefde-demo . Code is available at https://github.com/abduallahmohamed/Social-Implicit .

Reinforcement learning (RL) is widely used for post-training large language models (LLMs) in code editing, where group-relative methods, such as GRPO, are popular due to their critic-free and normalized advantage estimation. However, in real-world code-editing scenarios, reward distributions are often skewed with unpredictable noise, leading to distorted advantage computation and increased rollout outliers. To address this issue, we propose Group Adaptive Policy Optimization (GAPO), which adaptively finds an interval with the highest SNR (Signal to Noise Ratio) per prompt and uses the median of that interval as an adaptive Q to replace the group mean in advantage calculation to reduce noise further. This adaptive Q robustly handles rollout noise while remaining plug-and-play and efficient. We evaluate GAPO on nine instruction-tuned LLMs (3B-14B) using a collected large dataset of 51,844 real-world, history-aware code-editing tasks spanning 10 programming languages. GAPO yields up to 4.35 in-domain (ID) and 5.30 out-of-domain (OOD) exact-match improvements over GRPO and its variant DAPO, while achieving lower clipping ratios and higher GPU throughput. Code: https://github.com/TsingZ0/verl-GAPO.

Reinforcement learning (RL) post-training has become a trending paradigm for enhancing the capabilities of large language models (LLMs). Most existing RL systems for LLMs operate in a fully synchronous manner, where training must wait for the rollout of an entire batch to complete. This design leads to severe inefficiencies, as extremely long trajectories can stall the entire rollout process and leave many GPUs idle. To address this issue, we propose Concurrency- Controlled Partial Rollout with Importance Sampling (CoPRIS), which mitigates long-tail inefficiencies by maintaining a fixed number of concurrent rollouts, early-terminating once sufficient samples are collected, and reusing unfinished trajectories in subsequent rollouts. To mitigate the impact of off-policy trajectories, we introduce Cross-stage Importance Sampling Correction, which concatenates buffered log probabilities from the previous policy with those recomputed under the current policy for importance sampling correction. Experiments on challenging mathematical reasoning benchmarks show that CoPRIS achieves up to 1.94x faster training while maintaining comparable or superior performance to synchronous RL systems. The code of CoPRIS is available at https://github.com/777pomingzi/CoPRIS.

Constructing large-scale datasets for the GitHub issue resolution task is crucial for both training and evaluating the software engineering capabilities of Large Language Models (LLMs). However, the traditional process for creating such benchmarks is notoriously challenging and labor-intensive, particularly in the stages of setting up evaluation environments, grading test outcomes, and validating task instances. In this paper, we propose SWE-Factory, an automated pipeline designed to address these challenges. To tackle these issues, our pipeline integrates three core automated components. First, we introduce SWE-Builder, a multi-agent system that automates evaluation environment construction, which employs four specialized agents that work in a collaborative, iterative loop and leverages an environment memory pool to enhance efficiency. Second, we introduce a standardized, exit-code-based grading method that eliminates the need for manually writing custom parsers. Finally, we automate the fail2pass validation process using these reliable exit code signals. Experiments on 671 issues across four programming languages show that our pipeline can effectively construct valid task instances; for example, with GPT-4.1-mini, our SWE-Builder constructs 269 valid instances at 0.045 per instance, while with Gemini-2.5-flash, it achieves comparable performance at the lowest cost of 0.024 per instance. We also demonstrate that our exit-code-based grading achieves 100% accuracy compared to manual inspection, and our automated fail2pass validation reaches a precision of 0.92 and a recall of 1.00. We hope our automated pipeline will accelerate the collection of large-scale, high-quality GitHub issue resolution datasets for both training and evaluation. Our code and datasets are released at https://github.com/DeepSoftwareAnalytics/swe-factory.

Safety benchmark scores provide incomplete evidence of deployment readiness: aligned language models often adhere to rigid rules even when a situational update flips which action is safe. We term this failure brittle safety. To diagnose it, we introduce context-flip evaluation, testing 12 models across a safety benchmark (PacifAIst) and two commonsense controls using paired variants where the nominally safe action produces harm. Three findings emerge. First, brittle safety is safety-specific: all 12 models exhibit a safety-commonsense gap (mean +17.4 pp). Baseline accuracy fails to predict brittleness: among models above 90% baseline accuracy, brittleness rates range from 13.7% to 90.0%. Second, failures stem from policy override rather than miscomprehension: despite acknowledging the context change in every case, models persist via three distinct mechanisms that vary by update type and model family. Third, on a hand-audited probe of catastrophic consequence-flip scenarios, standard action-level guardrails catch none, while a state-aware validator catches all without false alarms on correct interventions. This indicates that action-level content moderation is systematically blind to consequence-flips, motivating state-aware architectural alternatives. We release our protocol, perturbed benchmarks, and deployment probe.

Reinforcement learning (RL) is a critical stage in post-training large language models (LLMs), involving repeated interaction between rollout generation, reward evaluation, and centralized learning. Distributing rollout execution offers opportunities to leverage more cost-efficient inference resources, but introduces challenges in wide-area coordination and policy dissemination. We present ECHO-2, a distributed RL framework for post-training with remote inference workers and non-negligible dissemination latency. ECHO-2 combines centralized learning with distributed rollouts and treats bounded policy staleness as a user-controlled parameter, enabling rollout generation, dissemination, and training to overlap. We introduce an overlap-based capacity model that relates training time, dissemination latency, and rollout throughput, yielding a practical provisioning rule for sustaining learner utilization. To mitigate dissemination bottlenecks and lower cost, ECHO-2 employs peer-assisted pipelined broadcast and cost-aware activation of heterogeneous workers. Experiments on GRPO post-training of 4B and 8B models under real wide-area bandwidth regimes show that ECHO-2 significantly improves cost efficiency while preserving RL reward comparable to strong baselines.

On-policy distillation has recently emerged as a promising alternative to standard sequence-level imitation, training a student by scoring its own rollouts with a teacher model. However, we observe ``Off-policy Teacher Decay'' problem in this paradigm: for the later tokens, with student's earlier trajectory as context that is off-policy to the teacher, the teacher's ability to produce a corrective score would decay, and may fall back to token-completion behavior learned in the pre-training stage. We empirically verify this problem, and we propose Early Stopping Rollout (ESR) to fix it: a simple yet effective distillation strategy that simply restricts the rollout generation to the first response tokens. We show that ESR both surpasses the full rollout OPD performance across model size, family, tasks and training regime, and exhibit much higher GPU efficiency and training stability, especially under cross model family scenarios. We further investigate the mechanism behind this surprising performance and discovered "Cascading Alignment" and "Sub-mode Commitment" effect of ESR that may explain why it works effectively and even sometimes exceeding the teacher model performance. Besides, we show that this position-based token selection strategy cannot be fully explainable by KL divergence and entropy signals.

Large language models (LLMs) have been widely applied to assist test generation with the source code under test provided as the context. This paper aims to answer the question: If the source code under test is incorrect, will LLMs be misguided when generating tests? The effectiveness of test cases is measured by their accuracy, coverage, and bug detection effectiveness. Our evaluation results with five open- and six closed-source LLMs on four datasets demonstrate that incorrect code can significantly mislead LLMs in generating correct, high-coverage, and bug-revealing tests. For instance, in the HumanEval dataset, LLMs achieve 80.45% test accuracy when provided with task descriptions and correct code, but only 57.12% when given task descriptions and incorrect code. For the APPS dataset, prompts with correct code yield tests that detect 39.85% of the bugs, while prompts with incorrect code detect only 19.61%. These findings have important implications for the deployment of LLM-based testing: using it on mature code may help protect against future regression, but on early-stage immature code, it may simply bake in errors. Our findings also underscore the need for further research to improve LLMs resilience against incorrect code in generating reliable and bug-revealing tests.

World models and multimodal large language models (MLLMs) provide complementary capabilities for predicting future outcomes from static visual observations. World models can generate concrete visual rollouts of possible futures, while MLLMs can reason abstractly over questions, goals, and rules. However, generated rollouts are stochastic and may be visually plausible but task-incorrect, making it necessary to determine when visual simulation is useful, whether a rollout is credible, and how it should influence the final answer. We formulate this problem as controlled concrete reasoning, where a model learns to invoke, verify, and integrate visual future simulation alongside abstract reasoning. To study this setting, we construct two human-verified benchmarks, VRQABench for controllable spatial lookahead and OpenWorldQA for open-domain physical prediction, and propose Privileged-Future On-Policy Self-Distillation (PF-OPSD). During training, PF-OPSD uses ground-truth future videos and answers only as teacher-side privileged context to evaluate on-policy concrete-reasoning trajectories, while the deployable student never observes true futures at test time. Experimental results show that PF-OPSD outperforms baseline by 10.6% and 10.9% on VRQABench and OpenWorldQA, respectively, while increasing robustness to noisy or conflicting rollouts. Our code and dataset are available at https://github.com/yczhou001/PF-OPSD.

Multi-model LLM systems such as routing, voting, cascades, fusion, and mixture-of-agents are used to beat single-model accuracy. We show that their gain is capped by a quantity the field rarely reports. For any policy whose output is one member model answer, accuracy cannot exceed one minus beta, where beta is the rate at which every model is wrong on the same query. In contrast, the usual diagnostic, average pairwise error correlation rho, cannot identify beta: error laws with identical marginals and pairwise correlations can have different all-wrong rates. A Clopper-Pearson bound on beta gives a finite-sample certificate on the largest gain any router, vote, or cascade could deliver before training a router. Across 67 models from 21 providers, a tetrachoric-calibrated single-factor model still underprices the all-wrong tail: on open-ended mathematics, observed beta is 0.052 versus 0.023 under the full 67-model Gaussian copula, about 2.5 times underpricing, with 90 percent CI 1.7 to 3.4 and k equals 17. The effect recurs on execution-graded code, where beta is 0.079. Re-asking the same GPQA-Diamond questions in free-response rather than multiple-choice form reopens the tail, with beta 0.127 and a five-judge panel with kappa 0.73 to 0.92, locating co-failure in answer format rather than subject. At matched quality, low-rho heterogeneous ensembles beat high-rho Self-MoA, but on checkable tasks in our pool, combining models rarely beats the single best model without a strong query-level routing signal. Gains come from models failing on different questions, not from adding more models.

Recent advances have demonstrated that integrating reinforcement learning with rule-based rewards can significantly enhance the reasoning capabilities of large language models, even without supervised fine-tuning. However, prevalent reinforcement learning algorithms such as GRPO and its variants like DAPO, suffer from a coarse granularity issue when computing the advantage. Specifically, they compute rollout-level advantages that assign identical values to every token within a sequence, failing to capture token-specific contributions and hindering effective learning. To address this limitation, we propose Key-token Advantage Estimation (KTAE) - a novel algorithm that estimates fine-grained, token-level advantages without introducing additional models. KTAE leverages the correctness of sampled rollouts and applies statistical analysis to quantify the importance of individual tokens within a sequence to the final outcome. This quantified token-level importance is then combined with the rollout-level advantage to obtain a more fine-grained token-level advantage estimation. Empirical results show that models trained with GRPO+KTAE and DAPO+KTAE outperform baseline methods across five mathematical reasoning benchmarks. Notably, they achieve higher accuracy with shorter responses and even surpass R1-Distill-Qwen-1.5B using the same base model.

Ensuring safety, factuality and overall quality in the generations of large language models is a critical challenge, especially as these models are increasingly deployed in real-world applications. The prevailing approach to addressing these issues involves collecting expensive, carefully curated datasets and applying multiple stages of fine-tuning and alignment. However, even this complex pipeline cannot guarantee the correction of patterns learned during pretraining. Therefore, addressing these issues during pretraining is crucial, as it shapes a model's core behaviors and prevents unsafe or hallucinated outputs from becoming deeply embedded. To tackle this issue, we introduce a new pretraining method that streams documents and uses reinforcement learning (RL) to improve the next K generated tokens at each step. A strong, post-trained model judges candidate generations -- including model rollouts, the original suffix, and a rewritten suffix -- for quality, safety, and factuality. Early in training, the process relies on the original and rewritten suffixes; as the model improves, RL rewards high-quality rollouts. This approach builds higher quality, safer, and more factual models from the ground up. In experiments, our method gives 36.2% and 18.5% relative improvements over standard pretraining in terms of factuality and safety, and up to 86.3% win rate improvements in overall generation quality.

We introduce the Agent GPA (Goal-Plan-Action) framework: an evaluation paradigm based on an agent's operational loop of setting goals, devising plans, and executing actions. The framework includes five evaluation metrics: Goal Fulfillment, Logical Consistency, Execution Efficiency, Plan Quality, and Plan Adherence. Logical Consistency checks that an agent's actions are consistent with its prior actions. Execution Efficiency checks whether the agent executes in the most efficient way to achieve its goal. Plan Quality checks whether an agent's plans are aligned with its goals; Plan Adherence checks if an agent's actions are aligned with its plan; and Goal Fulfillment checks that agent's final outcomes match the stated goals. Our experimental results on two benchmark datasets - the public TRAIL/GAIA dataset and an internal dataset for a production-grade data agent - show that this framework (a) provides a systematic way to cover a broad range of agent failures, including all agent errors on the TRAIL/GAIA benchmark dataset; (b) supports LLM-judges that exhibit strong agreement with human annotation, covering 80% to over 95% errors; and (c) localizes errors with 86% agreement to enable targeted improvement of agent performance.

Large language models (LLMs) have achieved impressive reasoning performance, with reinforcement learning with verifiable rewards (RLVR) emerging as a standard paradigm for post-training. A representative algorithm, group relative policy optimization (GRPO) (Shao et al., 2024), computes advantages by normalizing outcome rewards within response groups, but suffers from a vanishing advantage issue when all responses in a group receive identical rewards. To address this issue, we propose Adaptive Rollout and Response Reuse Policy Optimization (AR3PO), a sampling efficient RLVR algorithm that introduces two novel techniques: adaptive rollout, which dynamically allocates more responses to difficult prompts while saving computation on easier ones, and response reuse, which leverages previously generated correct responses to provide useful training signals. We compare AR3PO with strong RLVR baselines on multiple representative benchmarks using two different families of base models. Across the 7B and 8B models, AR3PO consistently outperforms GRPO and matches or surpasses DAPO (Yu et al., 2025), reducing rollout cost by up to 4.2x. On the larger 32B model, AR3PO achieves comparable performance to DAPO at similar training steps while maintaining substantially lower rollout cost.

Recent advances in Large Language Models (LLMs) and Reinforcement Learning (RL) have led to strong performance in open-domain question answering (QA). However, existing models still struggle with questions that admit multiple valid answers. Standard QA benchmarks, which typically assume a single gold answer, overlook this reality and thus produce inappropriate training signals. Existing attempts to handle ambiguity often rely on costly manual annotation, which is difficult to scale to multi-hop datasets such as HotpotQA and MuSiQue. In this paper, we present A^2Search, an annotation-free, end-to-end training framework to recognize and handle ambiguity. At its core is an automated pipeline that detects ambiguous questions and gathers alternative answers via trajectory sampling and evidence verification. The model is then optimized with RL using a carefully designed AnsF1 reward, which naturally accommodates multiple answers. Experiments on eight open-domain QA benchmarks demonstrate that A^2Search achieves new state-of-the-art performance. With only a single rollout, A^2Search-7B yields an average AnsF1@1 score of 48.4% across four multi-hop benchmarks, outperforming all strong baselines, including the substantially larger ReSearch-32B (46.2%). Extensive analyses further show that A^2Search resolves ambiguity and generalizes across benchmarks, highlighting that embracing ambiguity is essential for building more reliable QA systems. Our code, data, and model weights can be found at https://github.com/zfj1998/A2Search

We identify label errors in the test sets of 10 of the most commonly-used computer vision, natural language, and audio datasets, and subsequently study the potential for these label errors to affect benchmark results. Errors in test sets are numerous and widespread: we estimate an average of at least 3.3% errors across the 10 datasets, where for example label errors comprise at least 6% of the ImageNet validation set. Putative label errors are identified using confident learning algorithms and then human-validated via crowdsourcing (51% of the algorithmically-flagged candidates are indeed erroneously labeled, on average across the datasets). Traditionally, machine learning practitioners choose which model to deploy based on test accuracy - our findings advise caution here, proposing that judging models over correctly labeled test sets may be more useful, especially for noisy real-world datasets. Surprisingly, we find that lower capacity models may be practically more useful than higher capacity models in real-world datasets with high proportions of erroneously labeled data. For example, on ImageNet with corrected labels: ResNet-18 outperforms ResNet-50 if the prevalence of originally mislabeled test examples increases by just 6%. On CIFAR-10 with corrected labels: VGG-11 outperforms VGG-19 if the prevalence of originally mislabeled test examples increases by just 5%. Test set errors across the 10 datasets can be viewed at https://labelerrors.com and all label errors can be reproduced by https://github.com/cleanlab/label-errors.

World models -- generative models that simulate environment dynamics conditioned on past observations and actions -- are gaining prominence in planning, simulation, and embodied AI. However, evaluating their rollouts remains a fundamental challenge, requiring fine-grained, temporally grounded assessment of action alignment and semantic consistency -- capabilities not captured by existing metrics. Vision-Language Models (VLMs) have shown promise as automatic evaluators of generative content due to their strong multimodal reasoning abilities. Yet, their use in fine-grained, temporally sensitive evaluation tasks remains limited and requires targeted adaptation. We introduce a evaluation protocol targeting two recognition tasks -- action recognition and character recognition -- each assessed across binary, multiple-choice, and open-ended formats. To support this, we present UNIVERSE (UNIfied Vision-language Evaluator for Rollouts in Simulated Environments), a method for adapting VLMs to rollout evaluation under data and compute constraints. We conduct a large-scale study comparing full, partial, and parameter-efficient finetuning across task formats, context lengths, sampling strategies, and data compositions. The resulting unified evaluator matches the performance of task-specific baselines using a single checkpoint. Human studies confirm strong alignment with human judgments, establishing UNIVERSE as a scalable, semantics-aware evaluator for world models.

We introduce DS-1000, a code generation benchmark with a thousand data science problems spanning seven Python libraries, such as NumPy and Pandas. Compared to prior works, DS-1000 incorporates three core features. First, our problems reflect diverse, realistic, and practical use cases since we collected them from StackOverflow. Second, our automatic evaluation is highly specific (reliable) -- across all Codex-002-predicted solutions that our evaluation accept, only 1.8% of them are incorrect; we achieve this with multi-criteria metrics, checking both functional correctness by running test cases and surface-form constraints by restricting API usages or keywords. Finally, we proactively defend against memorization by slightly modifying our problems to be different from the original StackOverflow source; consequently, models cannot answer them correctly by memorizing the solutions from pre-training. The current best public system (Codex-002) achieves 43.3% accuracy, leaving ample room for improvement. We release our benchmark at https://ds1000-code-gen.github.io.

Wu et al. (2026) showed that most frontier large language models (LLMs) recommend a sponsored, roughly twice-as-expensive flight when their system prompt contains a soft sponsorship cue. We reproduce their evaluation on ten open-weight chat models plus the two of their twenty-three models that are still reachable today (gpt-3.5-turbo, gpt-4o). All reported rates in this paper are produced under the same judge the original paper used (gpt-4o); we additionally store every label under an open-weight (gpt-oss-120b) and a smaller proprietary (gpt-4o-mini) judge for an ablation. Three findings emerge. First, a prose description of an LLM evaluation pipeline is not, on its own, sufficient for accurate reproduction: we surfaced three silent implementation failures that each shifted a reported rate by tens of percentage points. Second, the central claims do generalise - the gpt-3.5-turbo logistic-regression intercept of alpha = 0.81 is within four points of the original alpha = 0.86, and 200 of 200 trials on gpt-3.5-turbo and gpt-4o promote a payday lender to a financially distressed user. Third, a thirty-token user prompt that asks the assistant for a neutral comparison table first cuts sponsored recommendation from 46.9% to 1.0% averaged across our ten open-source models, and from 53.0% to 0% averaged across the two OpenAI models. AI literacy and price-comparison portals are likely market-level mitigations; the harmful-product cell is bounded by neither. Raw data, labels and analysis scripts are at https://github.com/akmaier/Paper-LLM-Ads .

Current agentic AI benchmarks predominantly evaluate task completion accuracy, while overlooking critical enterprise requirements such as cost-efficiency, reliability, and operational stability. Through systematic analysis of 12 main benchmarks and empirical evaluation of state-of-the-art agents, we identify three fundamental limitations: (1) absence of cost-controlled evaluation leading to 50x cost variations for similar precision, (2) inadequate reliability assessment where agent performance drops from 60\% (single run) to 25\% (8-run consistency), and (3) missing multidimensional metrics for security, latency, and policy compliance. We propose CLEAR (Cost, Latency, Efficacy, Assurance, Reliability), a holistic evaluation framework specifically designed for enterprise deployment. Evaluation of six leading agents on 300 enterprise tasks demonstrates that optimizing for accuracy alone yields agents 4.4-10.8x more expensive than cost-aware alternatives with comparable performance. Expert evaluation (N=15) confirms that CLEAR better predicts production success (correlation ρ=0.83) compared to accuracy-only evaluation (ρ=0.41).

In personalized marketing, uplift models estimate incremental effects by modeling how customer behavior changes under alternative treatments. However, real-world data often exhibit biases - such as selection bias, spillover effects, and unobserved confounding - which adversely affect both estimation accuracy and metric validity. Despite the importance of bias-aware assessment, a lack of systematic studies persists. To bridge this gap, we design a systematic benchmarking framework. Unlike standard predictive tasks, real-world uplift datasets lack counterfactual ground truth, rendering direct metric validation infeasible. Therefore, a semi-synthetic approach serves as a critical enabler for systematic benchmarking, effectively bridging the gap by retaining real-world feature dependencies while providing the ground truth needed to isolate structural biases. Our investigations reveal that: (i) uplift targeting and prediction can manifest as distinct objectives, where proficiency in one does not ensure efficacy in the other; (ii) while many models exhibit inconsistent performance under diverse biases, TARNet shows notable robustness, providing insights for subsequent model design; (iii) evaluation metric stability is linked to mathematical alignment with the ATE, suggesting that ATE-approximating metrics yield more consistent model rankings under structural data imperfections. These findings suggest the need for more robust uplift models and metrics. Code will be released upon acceptance.

Language models are increasingly capable and are being rapidly deployed on a population-level scale. As a result, the safety of these models is increasingly high-stakes. Fortunately, advances in alignment have significantly reduced the likelihood of harmful model outputs. However, when models are queried billions of times in a day, even rare worst-case behaviors will occur. Current safety evaluations focus on capturing the distribution of inputs that yield harmful outputs. These evaluations disregard the probabilistic nature of models and their tail output behavior. To measure this tail risk, we propose a method to efficiently estimate the probability of harmful outputs for any input query. Instead of naive brute-force sampling from the target model, where harmful outputs could be rare, we operationalize importance sampling by creating unsafe versions of the target model. These unsafe versions enable sample-efficient estimation by making harmful outputs more probable. On benchmarks measuring misuse and misalignment, these estimates match brute-force Monte Carlo estimates using 10-20x fewer samples. For example, we can estimate probability of harmful outputs on the order of 10^-4 with just 500 samples. Additionally, we find that these harmfulness estimates can reveal the sensitivity of models to perturbations in model input and predict deployment risks. Our work demonstrates that accurate rare-event estimation is both critical and feasible for safety evaluations. Code is available at https://github.com/rangell/LMTailRisk

When a benchmark's accuracy saturates, it is often retired and replaced with a more challenging version. We show that this approach privileges accuracy and misses the opportunity to study six other key dimensions of agent performance: construct validity issues such as shortcuts, out-of-distribution generalizability, efficiency, reliability, the relative importance of the model versus the scaffold, and uplift from human-agent collaboration. We use CORE-Bench Hard, a benchmark for computational reproducibility of scientific code, as a case study to demonstrate that measuring agents along these dimensions yields meaningful insights into agent performance even after accuracy saturates. First, we surface threats to construct validity in CORE-Bench Hard that are difficult to anticipate with less capable agents. We introduce an improved benchmark, CORE-Bench v1.1, and an out-of-distribution task suite, CORE-Bench OOD. Second, we find that despite accuracy saturation, CORE-Bench v1.1 remains useful for measuring efficiency, reliability, model performance, and scaffold performance. Finally, we conduct a small-scale randomized experiment to measure uplift from human-agent collaboration on real-world computational reproducibility tasks. We find a statistically significant speedup by about a factor of two -- likely underestimated due to one-fifth of human-only reproductions reaching the time limit before completing -- and describe various other findings. Together, our contributions present a more rigorous alternative to the dominant accuracy-centric evaluation paradigm.

Forecasting has become a natural benchmark for reasoning under uncertainty. Yet existing evaluations of large language models remain limited to judgmental tasks in simple formats, such as binary or multiple-choice questions. In practice, however, forecasting spans a far broader scope. Across domains such as economics, public health, and social demographics, decisions hinge on numerical estimates over continuous quantities, a capability that current benchmarks do not capture. Evaluating such estimates requires a format that makes uncertainty explicit and testable. We propose prediction intervals as a natural and rigorous interface for this purpose. They demand scale awareness, internal consistency across confidence levels, and calibration over a continuum of outcomes, making them a more suitable evaluation format than point estimates for numerical forecasting. To assess this capability, we introduce a new benchmark QuantSightBench, and evaluate frontier models under multiple settings, assessing both empirical coverage and interval sharpness. Our results show that none of the 11 evaluated frontier and open-weight models achieves the 90\% coverage target, with the top performers Gemini 3.1 Pro (79.1\%), Grok 4 (76.4\%), and GPT-5.4 (75.3\%) all falling at least 10 percentage points short. Calibration degrades sharply at extreme magnitudes, revealing systematic overconfidence across all evaluated models.

Financial transaction processing requires extracting structured merchant information from noisy, abbreviated bank transaction strings at scale. Our current production system, a LoRA-fine-tuned LLaMA 3.1-8B, achieves 96.95% F1 on this task, but deploying 8-billion-parameter models imposes prohibitive memory, latency, and cost constraints. To identify more efficient alternatives, we conduct a deployment-focused study of 24 model variants spanning four model families: Gemma 3 (270M, 1B, 4B), Qwen 3.5 (0.8B, 2B, 4B), Aya (3.35B), and LLaMA 3.1-8B, systematically evaluating accuracy, inference throughput, training cost, and hardware behavior to assess production suitability. Our findings show that: (1) reproducing the LLaMA 3.1-8B fine-tune with a LoRA rank of 8 achieves 96.75% F1, only 0.20 points below the rank-32 baseline; (2) Qwen 3.5 4B with JSON-only prompting reaches 96.60% F1, within 0.35 points of the 8B baseline while using roughly half the parameters; (3) the 0.8B Qwen 3.5 model achieves 94.75% F1, matching models 2.5-4x larger and offering an attractive latency-accuracy trade-off; (4) chain-of-thought fine-tuning generally improves F1 by 0.3-1.8 points across most models, although Qwen 3.5 4B performs best with direct JSON-only prompting; and (5) Qwen 3.5 Think and Nothink training templates produce nearly identical results (F1 differences <0.004), indicating that explicit reasoning supervision is unnecessary for structured extraction tasks. We further deploy all 14 fine-tuned sub-8B models as Databricks Model Serving endpoints and observe that benchmark performance transfers reliably to production, with an average F1 change of only 0.8 points. Aya 3.35B, based on the Cohere2 architecture, is the sole exception, exhibiting a 3-5 point decline under serving conditions. Based on these results, we provide deployment recommendations across accuracy and latency requirements, ...

Leveraging the models' outputs, specifically the logits, is a common approach to estimating the test accuracy of a pre-trained neural network on out-of-distribution (OOD) samples without requiring access to the corresponding ground truth labels. Despite their ease of implementation and computational efficiency, current logit-based methods are vulnerable to overconfidence issues, leading to prediction bias, especially under the natural shift. In this work, we first study the relationship between logits and generalization performance from the view of low-density separation assumption. Our findings motivate our proposed method MaNo which (1) applies a data-dependent normalization on the logits to reduce prediction bias, and (2) takes the L_p norm of the matrix of normalized logits as the estimation score. Our theoretical analysis highlights the connection between the provided score and the model's uncertainty. We conduct an extensive empirical study on common unsupervised accuracy estimation benchmarks and demonstrate that MaNo achieves state-of-the-art performance across various architectures in the presence of synthetic, natural, or subpopulation shifts.

The increased use of large language models (LLMs) across a variety of real-world applications calls for automatic tools to check the factual accuracy of their outputs, as LLMs often hallucinate. This is difficult as it requires assessing the factuality of free-form open-domain responses. While there has been a lot of research on this topic, different papers use different evaluation benchmarks and measures, which makes them hard to compare and hampers future progress. To mitigate these issues, we developed OpenFactCheck, a unified framework, with three modules: (i) RESPONSEEVAL, which allows users to easily customize an automatic fact-checking system and to assess the factuality of all claims in an input document using that system, (ii) LLMEVAL, which assesses the overall factuality of an LLM, and (iii) CHECKEREVAL, a module to evaluate automatic fact-checking systems. OpenFactCheck is open-sourced (https://github.com/hasaniqbal777/openfactcheck) and publicly released as a Python library (https://pypi.org/project/openfactcheck/) and also as a web service (https://huggingface.co/spaces/hasaniqbal777/OpenFactCheck). A video describing the system is available at https://youtu.be/-i9VKL0HleI.

Recent machine learning papers often report 1-2 percentage point improvements from a single run on a benchmark. These gains are highly sensitive to random seeds, data ordering, and implementation details, yet are rarely accompanied by uncertainty estimates or significance tests. It is therefore unclear when a reported +1-2% reflects a real algorithmic advance versus noise. We revisit this problem under realistic compute budgets, where only a few runs are affordable. We propose a simple, PC-friendly evaluation protocol based on paired multi-seed runs, bias-corrected and accelerated (BCa) bootstrap confidence intervals, and a sign-flip permutation test on per-seed deltas. The protocol is intentionally conservative and is meant as a guardrail against over-claiming. We instantiate it on CIFAR-10, CIFAR-10N, and AG News using synthetic no-improvement, small-gain, and medium-gain scenarios. Single runs and unpaired t-tests often suggest significant gains for 0.6-2.0 point improvements, especially on text. With only three seeds, our paired protocol never declares significance in these settings. We argue that such conservative evaluation is a safer default for small gains under tight budgets.

Recent advancements in Large Reasoning Models (LRMs), exemplified by DeepSeek-R1, have underscored the potential of scaling inference-time compute through Group Relative Policy Optimization (GRPO). However, GRPO frequently suffers from gradient signal attenuation when encountering problems that are either too trivial or overly complex. In these scenarios, the disappearance of inter-group advantages makes the gradient signal susceptible to noise, thereby jeopardizing convergence stability. While variants like DAPO attempt to rectify gradient vanishing, they do not alleviate the substantial computational overhead incurred by exhaustive rollouts on low-utility samples. In this paper, we propose Difficulty-Estimated Policy Optimization (DEPO), a novel framework designed to optimize the efficiency and robustness of reasoning alignment. DEPO integrates an online Difficulty Estimator that dynamically assesses and filters training data before the rollout phase. This mechanism ensures that computational resources are prioritized for samples with high learning potential. Empirical results demonstrate that DEPO achieves up to a 2x reduction in rollout costs without compromising model performance. Our approach significantly lowers the computational barrier for training high-performance reasoning models, offering a more sustainable path for reasoning scaling. Code and data will be released upon acceptance.

Machine learning systems in fraud detection, credit scoring, and clinical risk assessment operate under delayed ground truth: outcome labels arrive days to months after the decision they evaluate. During this blind period, governance evidence degrades through mechanisms that neither drift detection methods nor governance frameworks adequately address. This paper formalizes an evidence sufficiency model with four dimensions (completeness, freshness, reliability, representativeness) and a decision-readiness gate that quantifies how label latency degrades evidence quality. The model maps three drift types to dimension-specific degradation trajectories. A complementary proxy indicator framework comprising seven measurement categories estimates sufficiency degradation without labels, with explicit coverage mapping and characterized blind spots per drift type. Evaluation on the IEEE-CIS Fraud Detection dataset (~590K transactions) with controlled drift injection shows that composite proxy monitoring detects covariate and mixed drift with 100% detection rate, while concept drift without feature change remains undetected -- consistent with the theoretical impossibility of unsupervised detection when P(X) is unchanged. Blind period simulation confirms monotone sufficiency degradation, with concept drift degrading fastest (S=0.242 at day 60 vs 0.418 for no-drift). The framework contributes a governance sufficiency monitoring instrument; its value lies in translating drift signals into auditable sufficiency assessments with characterized blind spots. Mapping sufficiency levels to governance actions requires deployment-specific calibration beyond this study's scope.

Despite algorithm-level innovations for multi-agent reinforcement learning (MARL), the underlying networked infrastructure for large-scale MARL training remains underexplored. Existing training frameworks primarily optimize for single-agent scenarios and fail to address the unique system-level challenges of MARL, including rollout-training synchronization barriers, rollout load imbalance, and training resource underutilization. To bridge this gap, we propose FlexMARL, the first end-to-end training framework that holistically optimizes rollout, training, and their orchestration for large-scale LLM-based MARL. Specifically, FlexMARL introduces the joint orchestrator to manage data flow under the rollout-training disaggregated architecture. Building upon the experience store, a novel micro-batch driven asynchronous pipeline eliminates the synchronization barriers while providing strong consistency guarantees. Rollout engine adopts a parallel sampling scheme combined with hierarchical load balancing, which adapts to skewed inter/intra-agent request patterns. Training engine achieves on-demand hardware binding through agent-centric resource allocation. The training states of different agents are swapped via unified and location-agnostic communication. Empirical results on a large-scale production cluster demonstrate that FlexMARL achieves up to 7.3x speedup and improves hardware utilization by up to 5.6x compared to existing frameworks.

Given a black-box AI system and a task, at what confidence level can a practitioner trust the system's output? We answer with a reliability level -- a single number per system-task pair, derived from self-consistency sampling and conformal calibration, that serves as a black-box deployment gate with exact, finite-sample, distribution-free guarantees. Self-consistency sampling reduces uncertainty exponentially; conformal calibration guarantees correctness within 1/(n+1) of the target level, regardless of the system's errors -- made transparently visible through larger answer sets for harder questions. Weaker models earn lower reliability levels (not accuracy -- see Definition 2.4): GPT-4.1 earns 94.6% on GSM8K and 96.8% on TruthfulQA, while GPT-4.1-nano earns 89.8% on GSM8K and 66.5% on MMLU. We validate across five benchmarks, five models from three families, and both synthetic and real data. Conditional coverage on solvable items exceeds 0.93 across all configurations; sequential stopping reduces API costs by around 50%.

Benchmarks are essential for consistent evaluation and reproducibility. The integration of Artificial Intelligence into Software Engineering (AI4SE) has given rise to numerous benchmarks for tasks such as code generation and bug fixing. However, this surge presents challenges: (1) scattered benchmark knowledge across tasks, (2) difficulty in selecting relevant benchmarks, (3) the absence of a uniform standard for benchmark development, and (4) limitations of existing benchmarks. In this paper, we review 173 studies and identify 204 AI4SE benchmarks. We classify these benchmarks, analyze their limitations, and expose gaps in practices. Based on our review, we created BenchScout, a semantic search tool to find relevant benchmarks, using automated clustering of the contexts from associated studies. We conducted a user study with 22 participants to evaluate BenchScout's usability, effectiveness, and intuitiveness which resulted in average scores of 4.5, 4.0, and 4.1 out of 5. To advance benchmarking standards, we propose BenchFrame, a unified method to enhance benchmark quality. As a case study, we applied BenchFrame to the HumanEval benchmark and addressed its main limitations. This led to HumanEvalNext, featuring (1) corrected errors, (2) improved language conversion, (3) expanded test coverage, and (4) increased difficulty. We then evaluated ten state-of-the-art code language models on HumanEval, HumanEvalPlus, and HumanEvalNext. On HumanEvalNext, models showed a pass@1 score reduction of 31.22% and 19.94% compared to HumanEval and HumanEvalPlus, respectively.

Large language model agents now act on codebases, browsers, operating systems, calendars, files, and tool ecosystems, but the benchmarks used to evaluate them are fragmented: each emphasizes a different unit of measurement (final task success, tool-call validity, repeated-pass consistency, trajectory safety, or attack robustness). A line of 2024-2025 work has converged on the diagnosis that a single accuracy column is no longer the right unit of comparison for deployable agents. AgentAtlas extends this line of work with four components: (i) a six-state control-decision taxonomy (Act / Ask / Refuse / Stop / Confirm / Recover); (ii) a nine-category trajectory-failure taxonomy with two orthogonal hierarchical labels (primary_error_source, impact); (iii) a taxonomy-aware vs. taxonomy-blind methodology that measures how much of a model's apparent capability comes from the supervision in the prompt; and (iv) a benchmark-coverage audit mapping fifteen agent benchmarks against six behavioral axes. To demonstrate the methodology we run a small fixed eight-model set (1,342 generated items, four frontier closed and four open-weight) under both prompt modes. Removing the explicit label menu drops every model's trajectory accuracy by 14-40 pp to a tight 0.54-0.62 floor regardless of family, and no single model wins on all three of control accuracy, trajectory diagnosis, and tool-context utility retention. We treat the synthetic run as a measurement-protocol demonstration, not a benchmark release.

Training a model to predict the next step in a concurrent program is harder than it looks: two runs of the same program from the same trace prefix can produce different next events, both valid, because the scheduler is nondeterministic. A model trained against a single label is learning to guess one outcome of a random process. We turn this around and use the nondeterminism as a training signal. We run each program many times, aggregate the observed next events into an empirical distribution, and fine-tune a 7B model to match that distribution with a KL objective. On 798 held-out predictions drawn from real production Go bugs (CockroachDB, Kubernetes, gRPC, etcd), fine-tuning on fewer than a thousand traces reaches 36.2% accuracy, ahead of Gemini 3.5 Flash used zero-shot (34.8%) and the same model without fine-tuning (28.6%). Distribution training matches cross-entropy on accuracy (35.8% vs. 36.2%) while reducing Expected Calibration Error from 0.205 to 0.169. We also derive a formal goroutine-leak signature for a class of select-blocked goroutines where P(GoUnblock)=0 holds by scheduler semantics, not by learning. We release the dataset, trained adapters, and all tooling.

Recent benchmarks have probed factual consistency and rhetorical robustness in Large Language Models (LLMs). However, a knowledge gap exists regarding how directional framing of factually true statements influences model agreement, a common scenario for LLM users. AssertBench addresses this by sampling evidence-supported facts from FEVEROUS, a fact verification dataset. For each (evidence-backed) fact, we construct two framing prompts: one where the user claims the statement is factually correct, and another where the user claims it is incorrect. We then record the model's agreement and reasoning. The desired outcome is that the model asserts itself, maintaining consistent truth evaluation across both framings, rather than switching its evaluation to agree with the user. AssertBench isolates framing-induced variability from the model's underlying factual knowledge by stratifying results based on the model's accuracy on the same claims when presented neutrally. In doing so, this benchmark aims to measure an LLM's ability to "stick to its guns" when presented with contradictory user assertions about the same fact. The complete source code is available at https://github.com/achowd32/assert-bench.

Skill-augmented reinforcement learning improves language agents by storing reusable procedural knowledge acquired from past experience. Existing methods typically use strong language models to analyze trajectories, generate skills, and update a retrievable skill bank during online training. However, they rarely assess whether a newly generated skill is useful before it is stored and reused. We find that this assumption is unreliable: even skills generated by proprietary frontier LLMs exhibit highly mixed utility, with many providing little benefit or even degrading performance. Once such skills enter the bank, their effects are difficult to identify, because subsequent rollout feedback is delayed and usually reflects the combined effect of multiple retrieved skills rather than the marginal contribution of any individual skill. We propose an online reinforcement learning framework for pre-storage skill validation. The framework estimates whether a candidate skill contributes useful information beyond the skills already retrieved for the current task. It uses the standard rollout budget to form two matched groups under the same task and retrieval context: base rollouts conditioned on the currently retrieved skills, and skill-augmented rollouts conditioned on the same skills plus one candidate skill induced from the base trajectories. The reward gap between these two groups estimates the candidate skill's context-dependent marginal utility, enabling the framework to promote useful skills while filtering ineffective or harmful ones without additional rollout overhead. The framework further uses this marginal-utility signal to train the policy itself as a skill generator, reducing reliance on repeated calls to proprietary models. The learned skill-generation likelihood serves as a context-dependent score for retrieval-time reranking and outdated-skill pruning as the policy evolves.

Reinforcement Learning with Verifiable Rewards (RLVR) effectively trains reasoning models that rely on abundant perfect labels, but its vulnerability to unavoidable noisy labels due to expert scarcity remains critically underexplored. In this work, we take the first step toward a systematic analysis of noisy label mechanisms in RLVR. In contrast to supervised classification, most RLVR algorithms incorporate a rollout-based condition: a label's influence on training is contingent on whether the current policy can generate rollouts that realize it, a property that naturally extends to noisy labels. Based on this observation, we distinguish two types of noise: inactive noisy labels, which reduce data efficiency, and active noisy labels, which are reinforced and risk skewing the model toward incorrect distributions. From experiments on training with noisy samples, we identify an Early Correctness Coherence phenomenon: although noisy samples begin to lag behind in later stages, accuracy on both clean and noisy samples increases similarly in early training. Motivated by this dynamic, we propose Online Label Refinement (OLR), which progressively corrects potentially noisy labels with majority-voted answers when two conditions hold: a positive slope in the majority answer's rollout pass rate and stable historical consistency across updates, enabling gradual self-correction as the policy improves. We evaluate OLR on six in-distribution mathematical reasoning benchmarks (AIME24/25, AMC, MATH-500, Minerva, and Olympiad) and three out-of-distribution tasks (ARC-c, GPQA-diamond, and MMLU-pro). Across noise ratios from 0.1 to 0.9, OLR consistently improves robustness under both inactive and active noisy-label settings, achieving average gains of 3.6% to 3.9% on in-distribution benchmarks and 3.3% to 4.6% on out-of-distribution evaluations.

Automated Code Review (ACR) is crucial for software quality, yet existing benchmarks often fail to reflect real-world complexities, hindering the evaluation of modern Large Language Models (LLMs). Current benchmarks frequently focus on fine-grained code units, lack complete project context, and use inadequate evaluation metrics. To address these limitations, we introduce SWRBench , a new benchmark comprising 1000 manually verified Pull Requests (PRs) from GitHub, offering PR-centric review with full project context. SWRBench employs an objective LLM-based evaluation method that aligns strongly with human judgment (~90 agreement) by verifying if issues from a structured ground truth are covered in generated reviews. Our systematic evaluation of mainstream ACR tools and LLMs on SWRBench reveals that current systems underperform, and ACR tools are more adept at detecting functional errors. Subsequently, we propose and validate a simple multi-review aggregation strategy that significantly boosts ACR performance, increasing F1 scores by up to 43.67%. Our contributions include the SWRBench benchmark, its objective evaluation method, a comprehensive study of current ACR capabilities, and an effective enhancement approach, offering valuable insights for advancing ACR research.

Machine learning models often experience performance degradation post-deployment due to shifts in data distribution. It is challenging to assess model's performance accurately when labels are missing or delayed. Existing proxy methods, such as drift detection, fail to measure the effects of these shifts adequately. To address this, we introduce a new method, Probabilistic Adaptive Performance Estimation (PAPE), for evaluating classification models on unlabeled data that accurately quantifies the impact of covariate shift on model performance. It is model and data-type agnostic and works for various performance metrics. Crucially, PAPE operates independently of the original model, relying only on its predictions and probability estimates, and does not need any assumptions about the nature of the covariate shift, learning directly from data instead. We tested PAPE on tabular data using over 900 dataset-model combinations created from US census data, assessing its performance against multiple benchmarks. Overall, PAPE provided more accurate performance estimates than other evaluated methodologies.

Language models can use verifiable rewards to improve at a wide variety of reasoning tasks. However, both parametric (e.g. RLVR) and non-parametric (e.g. prompt optimization) approaches to doing so typically require hundreds of training samples and thousands of model rollouts, making them expensive in the best case and intractable in the worst. To address this challenge, we introduce Contrastive Reflection (CORE), a non-parametric learning algorithm that compares past reasoning traces to generate insights: short natural-language descriptions of reasoning strategies and constraints that capture differences between successful and unsuccessful problem attempts. Across four reasoning tasks, we demonstrate that CORE enables more rapid improvement than both parametric (GRPO) and non-parametric (GEPA, episodic RAG, and MemRL) methods, while using fewer rollouts. Under fixed rollout budgets with as few as five training samples, we then show that CORE also achieves comparable or greater performance gains than each baseline. Finally, we highlight how CORE is also substantially more context-efficient than non-parametric baselines, requiring fewer prompt tokens while storing learned knowledge as compact, interpretable natural-language insights. Our results therefore suggest that distilling contrasts between successful and unsuccessful reasoning traces into abstract and useful insights can provide a more efficient and interpretable route to model self-improvement than weight updates, prompt optimization, or direct reuse of stored reasoning traces.

As AI-generated images proliferate across digital platforms, reliable detection methods have become critical for combating misinformation and maintaining content authenticity. While numerous deepfake detection methods have been proposed, existing benchmarks predominantly evaluate fine-tuned models, leaving a critical gap in understanding out-of-the-box performance -- the most common deployment scenario for practitioners. We present the first comprehensive zero-shot evaluation of 16 state-of-the-art detection methods, comprising 23 pretrained detector variants (due to multiple released versions of certain detectors), across 12 diverse datasets, comprising 2.6~million image samples spanning 291 unique generators including modern diffusion models. Our systematic analysis reveals striking findings: (1)~no universal winner exists, with detector rankings exhibiting substantial instability (Spearman~ρ: 0.01 -- 0.87 across dataset pairs); (2)~a 37~percentage-point performance gap separates the best detector (75.0\% mean accuracy) from the worst (37.5\%); (3)~training data alignment critically impacts generalization, causing up to 20--60\% performance variance within architecturally identical detector families; (4)~modern commercial generators (Flux~Dev, Firefly~v4, Midjourney~v7) defeat most detectors, achieving only 18--30\% average accuracy; and (5)~we identify three systematic failure patterns affecting cross-dataset generalization. Statistical analysis confirms significant performance differences between detectors (Friedman test: χ^2=121.01, p<10^{-16}, Kendall~W=0.524). Our findings challenge the ``one-size-fits-all'' detector paradigm and provide actionable deployment guidelines, demonstrating that practitioners must carefully select detectors based on their specific threat landscape rather than relying on published benchmark performance.

How much do large language models actually hallucinate when answering questions grounded in provided documents? Despite the critical importance of this question for enterprise AI deployments, reliable measurement has been hampered by benchmarks that rely on static datasets vulnerable to contamination, LLM-based judges with documented biases, or evaluation scales too small for statistical confidence. We address this gap using RIKER, a ground-truth-first evaluation methodology that enables deterministic scoring without human annotation. Across 35 open-weight models, three context lengths (32K, 128K, and 200K tokens), four temperature settings, and three hardware platforms (NVIDIA H200, AMD MI300X, and Intel Gaudi 3), we conducted over 172 billion tokens of evaluation - an order of magnitude beyond prior work. Our findings reveal that: (1) even the best-performing models fabricate answers at a non-trivial rate - 1.19% at best at 32K, with top-tier models at 5 - 7% - and fabrication rises steeply with context length, nearly tripling at 128K and exceeding 10% for all models at 200K; (2) model selection dominates all other factors, with overall accuracy spanning a 72-percentage-point range and model family predicting fabrication resistance better than model size; (3) temperature effects are nuanced - T=0.0 yields the best overall accuracy in roughly 60% of cases, but higher temperatures reduce fabrication for the majority of models and dramatically reduce coherence loss (infinite generation loops), which can reach 48x higher rates at T=0.0 versus T=1.0; (4) grounding ability and fabrication resistance are distinct capabilities - models that excel at finding facts may still fabricate facts that do not exist; and (5) results are consistent across hardware platforms, confirming that deployment decisions need not be hardware-dependent.