Reproduce “ThinkBench: Dynamic Out-of-Distribution Evaluation for Robust LLM Reasoning” on OpenResearch

About this paper

Shulin Huang, Linyi Yang, Yan Song, Shuang Chen, Leyang Cui, Ziyu Wan, Qingcheng Zeng, Ying Wen and 4 more

A collaborative effort including researchers from Zhejiang University, Westlake University, UCL, Shanghai Jiao Tong University, Northwestern University, and Huawei Noah’s Ark Lab created ThinkBench, a dynamic evaluation framework that robustly assesses LLMs' reasoning capabilities by generating out-of-distribution (OOD) data. The framework, designed to combat data contamination, revealed a universal performance decline, averaging 24.9% on AIME-500, when models shifted from in-distribution to OOD test samples.

Key insights

• LLMs universally exhibit a noticeable performance degradation when evaluated on ThinkBench's OOD data, demonstrating a reliance on specific phrasing or patterns encountered during training rather than pure reasoning.
• Analysis revealed strong evidence of data leakage in older, static benchmarks (e.g., AIME-500 questions from 1983-2023) due to significantly larger ID-OOD performance gaps compared to newer, unseen data like AIME 2024.
• ThinkBench's OOD datasets effectively validate test-time scaling laws for Process Reward Models (PRMs), showing consistent performance improvements with increased computational budgets at inference time, even for unseen problems.

Summary by alphaXiv.