Reproduce “Efficient Reasoning Training Does Not Always Harm CoT Faithfulness and Monitorability” 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

Samuel Lewis-Lim, Xingwei Tan, Mario Sanger, Zhixue Zhao, Nikolaos Aletras

The authors test whether training reasoning models to use fewer tokens changes two oversight uses of chain-of-thought: predicting decisions on related inputs and detecting outside cues that influence answers. Across seven faithfulness tasks, Qwen3-4B GLP shortened traces 75% and lowered NSG from 32.7% to 23.0%; monitorability was more robust, but not uniformly preserved, especially under heavy compression on medical cues.

Summary by alphaXiv.

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