Reproduce “Procedure without Judgment: What Small Language Models Learn under an Exact Verifier” 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
Arunkumar V, Manoranjan Gandhudi, Gangadharan G. R., Arun Prakash, S. Senthilkumar
The authors test whether language models can identify minimal request features that guarantee an agent’s authorization decision without seeing its policy. Models can query an exact checker, while an algorithm teaches a deletion procedure. On 64 GlassGate episodes, a Qwen3-4B student trained on 2,000 algorithm traces is exact on 92.2%; on unseen policies, exactness falls to 47.9%, despite 93.8% sufficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.