Reproduce “Blind Spots Travel, Knowledge Stays Home: What RL Inherits from an LLM Judge” 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

Liang Yue, Eric A. Vance

The authors test whether graders’ training feedback can cause category-specific refusals on unseen questions. In closed-book short-answer experiments, they vary verdicts on known correct training answers while holding other verdicts fixed. For Llama-3.1-8B, rejecting twelve former-state-capital answers raises refusal on unseen examples from 5% to 90% across three runs per condition; this is evidence, not a complete causal account.

Summary by alphaXiv.

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