Reproduce “The belief-action gap in language model agents” 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

Ayse Sila Okcu, Weisheng Wang, Tongmeng Xie, Niall Dalton, Angelos Nalmpantis, Raghu Arghal, Mario Giulianelli

The authors assess how a language-model agent’s actions relate to decoded state information and inferred preferences in a deterministic DoorKey grid task. They compare probe-estimated beliefs with emitted moves and infer preferences from choices. Across 3,194 decisions from gpt-oss-20b, the belief–action gap falls from 0.467 to 0.268, while decoded entropy rises. These model-derived diagnostics do not identify causes of failures.

Summary by alphaXiv.

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