frett/Live-CLKT-Bench

LiveCLKTBench Examples

Generation Config:

• Domain: sports; Entity Date range:2026-03-20 ~ 2026-04-10
• Domain: movie; Entity Date range:2026-01-01 ~ 2026-04-10

Overview

Evaluating cross-lingual knowledge transfer in large language models is challenging, as correct answers in a target language may arise either from genuine transfer or from prior exposure during pre-training. We present LiveCLKTBench, an automated pipeline for generating realistic, contamination-free, and continuously refreshable benchmarks for Cross-Lingual Knowledge Transfer. LiveCLKTBench identifies independent, time-sensitive knowledge entities from real-world domains, filters them based on temporal constraints, and verifies whether the model already possesses this knowledge. The documents corresponding to valid entities are then used to generate factual questions, which are translated into multiple languages to evaluate cross-lingual transfer. Based on the selected domain and evaluation languages, each instance generated by LiveCLKTBench consists of two components:

(i) A training set containing source-language documents that introduce previously unseen knowledge (e.g., score result of a recent MLB match) to the model.

(ii) A multilingual validation/test set composed of factual questions derived from the same documents but expressed in target languages. (The validation set is used for checkpoint selection, while the test set is reserved for final evaluation.)

Here is an example:

===== Train set instance =====
Sports: baseball
League: MLB
Match: Cincinnati Reds vs Boston Red Sox
Date: 2026-03-26
Score: 0 - 3
Venue: Great American Ball Park
Innings Breakdown:
Cincinnati Reds: 0 0 0 0 0 0 0 0 0
Boston Red Sox: 0 0 0 0 0 0 1 0 2

===== Test set instance =====
'シンシナティ・レッズ 対 ボストン・レッドソックス' の試合で、'2026-03-26' にシンシナティ・レッズは何得点しましたか？ 
(A.) 1  (B.) 3  (C.) 2  (D.) 0

在 2026-03-26 '辛辛那提紅人對波士頓紅襪'的比賽中：，辛辛那提紅人得了幾分？
(A.) 1  (B.) 3  (C.) 2  (D.) 0