Reproduce “Not All LLM Reasoning is Visible in the Chain-of-Thought” 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
Vatsal Baherwani, Tom Goldstein, Ashwinee Panda
The authors test whether fixed filler sequences improve language models’ answers without readable chain of thought. They compare tasks and contexts, using activation interventions and probes to examine task-related information. In a synthetic hidden-rule task, Claude Opus 4.5’s correct “N/A” rate rises from 33.5% to 44.5%. This shows capability, not real-world monitor evasion, and does not isolate extra computation alone.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.