Reproduce “Benchmarking Reasoning Robustness in Large Language Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Tong Yu, Yongcheng Jing, Xikun Zhang, Wentao Jiang, Wenjie Wu, Yingjie Wang, Wenbin Hu, Bo Du and 1 more

Yu et al. introduce Math-RoB, a novel benchmark designed to systematically assess the robustness of Large Language Models (LLMs) in mathematical reasoning. Evaluations reveal consistent vulnerabilities across models, including positional bias, instruction sensitivity, numerical fragility, and a tendency to hallucinate missing information rather than acknowledge incompleteness.

Key insights

  • Current Large Language Models often rely on heuristic recall and probabilistic pattern matching from memorized data rather than demonstrating systematic logical inference.
  • The reasoning capabilities of LLMs are fragile, exhibiting significant vulnerabilities to even subtle changes in input structure, instructions, or numerical representations.
  • Models frequently hallucinate missing information instead of recognizing unsolvable problems, indicating a lack of true understanding or self-awareness regarding data completeness.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.