Dataset Preview
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
Job manager crashed while running this job (missing heartbeats).
Error code: JobManagerCrashedError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
text string |
|---|
**Primary Area:** foundation or frontier models, including LLMs |
{0}------------------------------------------------ |
# OPTIMAL ARCHITECTURES FOR JUDGING LLM OUTPUTS USING LLMs |
## Anonymous authors |
Paper under double-blind review |
## ABSTRACT |
This paper explores optimal architectures for evaluating the outputs of large language models (LLMs) using LLMs themselves. We propose a novel framework that interprets LLMs as advocates within an ensemble of interacting agents, allowing them to defend their answers and reach conclusions through a judge and jury system... |
## 1 INTRODUCTION |
The rapid advancement of large language models (LLMs) has revolutionized the field of natural language processing, enabling the development of increasingly sophisticated AI systems capable of generating human-like text, engaging in dialogue, and performing complex language tasks (5). As these models grow in size and ca... |
Traditional evaluation methods, such as human assessments and automated metrics, often struggle to capture the nuances and complexities of LLM outputs, leading to a gap between model performance and user expectations (7; 17; 24). Human evaluations are time-consuming, expensive, and prone to inconsistency and bias (12; ... |
To address these challenges, we propose a novel framework for evaluating LLM outputs using LLMs themselves as interacting agents in a courtroom-inspired, multi-agent system. Our approach draws inspiration from various fields, including decision theory, economics, psychology, legal theory, and voting theory, to develop ... |
 |
The diagram illustrates two architectural models for LLM evaluation. On the left, the MORE architecture shows a 'Question' being answered by 'Answer 1' and 'Answer 2'. Each answer is supported by multiple 'Advocates' (represented by small figures). These advocates present their arguments to a 'Judge' (represented by a ... |
Figure 1: Illustrations of the architectures: the MORE architecture (left) employs multiple advocates per answer, while the SAMRE architecture (right) utilizes a single advocate per answer but allows for multiple rounds of evaluation. |
Figure 1: Illustrations of the architectures: the **MORE** architecture (left) employs multiple advocates per answer, while the **SAMRE** architecture (right) utilizes a single advocate per answer but allows for multiple rounds of evaluation. |
{1}------------------------------------------------ |
### 1.1 MOTIVATION FROM DECISION THEORY AND LEGAL THEORY |
Our approach is motivated by various approaches proposed in literature on designing systems with agents of varying capabilities and incentives. In what follows, we review a few motivating frameworks. |
### DECISION THEORY AND BOUNDED RATIONALITY |
Decision theory provides a foundation for understanding how agents make choices under uncertainty and constraints (44; 21). The concept of bounded rationality, introduced by Herbert A. Simon (39; 40), acknowledges that decision-makers often operate with limited information, cognitive resources, and time, leading to sat... |
#### PSYCHOLOGICAL THEORIES OF PERSUASION AND ARGUMENTATION |
Psychological theories of persuasion and argumentation, such as the Elaboration Likelihood Model (36) and the Heuristic-Systematic Model (8), provide valuable insights into the factors that influence the effectiveness of arguments and the formation of judgments. These theories highlight the importance of central and pe... |
Our LLM advocates framework incorporates elements of persuasion and argumentation theory by encouraging LLMs to present well-structured, compelling arguments that appeal to both central and peripheral routes of persuasion. By exposing the outputs to scrutiny from opposing advocates and subjecting them to the judgment o... |
### LEGAL THEORIES OF ADVERSARIAL PROCESS AND JURISPRUDENCE |
Legal theories of adversarial process and jurisprudence emphasize the importance of structured debate, cross-examination, and impartial judgment in uncovering truth and reaching fair outcomes (46; 16; 41). The adversarial system, which lies at the heart of many legal traditions, relies on the clash of opposing argument... |
Our LLM advocates framework draws inspiration from the adversarial legal process, casting LLMs as advocates tasked with presenting and defending competing arguments, while other LLMs serve as impartial judges and juries. This structure promotes a thorough and rigorous examination of LLM outputs, exposing weaknesses and... |
Furthermore, we also draw inspiration from Voting and social choice theories, which study the design of collective decision-making systems, considering factors such as preference aggregation, strategic behavior, and fairness (2; 18; 38). Our LLM advocates framework incorporates recommendations of voting theory and soci... |
## 1.2 NOVEL CONTRIBUTIONS AND PAPER STRUCTURE |
Building on the insights from these diverse fields, our paper makes several novel contributions to the problem of LLM evaluation: |
{2}------------------------------------------------ |
1. We propose a dynamic, multi-agent framework that casts LLMs as interacting advocates, judges, and juries, enabling a more comprehensive and contextual assessment of LLM outputs. |
2. We introduce a courtroom-inspired architecture that leverages the power of structured debate, cross-examination, and impartial judgment to uncover strengths, weaknesses, and inconsistencies in LLM responses. |
3. We draw on theories of bounded rationality, incentive design, persuasion, argumentation, and adversarial process to inform the design of our LLM advocates framework, ensuring that the system promotes accurate, unbiased, and trustworthy evaluations. |
4. We explore the use of voting theory and social choice principles to design effective jury systems for aggregating LLM judgments, promoting fair and representative assessments while mitigating the influence of strategic behavior and individual biases. |
The remainder of this paper is structured as follows: Section 2 reviews related work on LLM evaluation, highlighting the limitations of existing approaches and the need for more sophisticated assessment frameworks. Section 3 introduces our LLM advocates framework, detailing its courtroom-inspired architecture, the role... |
## 2 RELATED WORK |
The evaluation of language models has been a longstanding challenge in the field of natural language processing, with the rapid growth of LLMs in recent years bringing this issue to the forefront. As these models have increased in size and capability, the need for robust, comprehensive, and theoretically grounded evalu... |
### 2.1 HUMAN-BASED EVALUATIONS AND THE CHALLENGE OF SUBJECTIVITY |
Human judgments have long been considered the gold standard for evaluating the quality of language model outputs, with platforms like LMSYS Chatbot Arena (10) and others (47; 23) providing structured environments for collecting human ratings and preferences. However, the subjectivity and variability inherent in human e... |
Research in human-computer interaction and cognitive psychology has shown that factors such as individual differences, task framing, and cognitive biases can significantly influence human judgments of AI systems (20; 14; 6). For example, the anchoring effect (42) and the halo effect (33) can lead to over- or under-esti... |
Moreover, the reliance on reinforcement learning from human feedback (RLHF) (4; 11; 50) for aligning LLMs with user expectations introduces additional challenges, as the pool of reinforcers may not be representative of the general user population (31). This can lead to models that are optimized for the preferences of a... |
## 2.2 AUTOMATED METRICS AND THE LIMITS OF REFERENCE-BASED EVALUATION |
To address the scalability and consistency issues of human evaluations, researchers have developed various automated metrics for assessing LLM performance across different tasks, such as BLEU |
{3}------------------------------------------------ |
(35) for machine translation, ROUGE (28) for summarization, and exact match (EM) and F1 scores (37) for question answering. These metrics provide a standardized and efficient means of evaluating LLMs, enabling the comparison of different models and the tracking of progress over time. |
However, the reliance of these metrics on reference-based evaluation, where model outputs are compared against a fixed set of ground-truth answers, has been shown to have significant limitations (29; 13; 22). In open-ended generation tasks, such as dialogue and creative writing, there may be a wide range of acceptable ... |
### 2.3 LLM-BASED EVALUATION AND THE PROMISE OF MULTI-AGENT FRAMEWORKS |
To overcome the limitations of human evaluations and automated metrics, recent research has explored the use of LLMs themselves as evaluators. This approach leverages the linguistic knowledge and reasoning capabilities of LLMs to provide more nuanced and contextually aware assessments of model outputs. Initial studies ... |
However, the use of single LLMs as evaluators has been shown to suffer from issues of bias and limited generalizability (34). To address these concerns, researchers have proposed using multiple LLMs as evaluators, drawing on insights from multi-agent systems and ensemble learning (43; 26). |
Our work builds on these ideas by proposing a novel LLM advocates framework that interprets LLMs as interacting agents within a courtroom-inspired setting. This framework draws on principles from adversarial legal systems (46; 16; 41), where the clash of opposing arguments and the judgment of impartial decision-makers ... |
### 2.4 SCORING, RANKING, AND AGGREGATION METHODS IN MULTI-AGENT EVALUATION |
End of preview.
No dataset card yet
- Downloads last month
- 1,660