Reproduce “Training LLMs to Verbalize Evaluation Awareness” 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

Usman Anwar, Sahar Abdelnabi, David Krueger

Verbalization training uses a model’s spontaneous evaluation-awareness reports to identify prefixes where awareness has just appeared, then reinforces only the report span with masking and KL regularization. Qwen3.6-35B-A3B’s judge-identified VEA rate rose from 5.8% to 16.7% across 17 EAB conditions, but this measures reporting, not full awareness, and relies on faithful seed reports.

Summary by alphaXiv.

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