Reproduce “Robust Answers, Fragile Logic: Probing the Decoupling Hypothesis in LLM Reasoning” 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

Enyi Jiang, Changming Xu, Nischay Singh, Tian Qiu, Gagandeep Singh

MATCHA tests whether a model’s explanation remains reliable when its answer stays correct. It perturbs questions while preserving initially correct answers, then checks whether the chain of thought (CoT) becomes wrong. On Llama-3-8B and GSM8K, full MATCHA has a 14.7% success rate, versus 13.0% for its CoT-loss-only variant; this measures vulnerability under the test, not ordinary-use prevalence.

Summary by alphaXiv.

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