01
What the paper asks
Can hidden chain-of-thought — reasoning a model performs but does not verbalize — be detected from the outside? The paper introduces the Hidden CoT Detection Score (HCDS), which combines linguistic, behavioral, and mechanistic indicators, and tests anchor-suppression interventions on Qwen3-4B across GSM8K and StrategyQA.
02
Why it sits in this notebook
What a system verbalizes is not necessarily what it computes. That is the plan-versus-implementation gap from Factor(U,T) in another form, and it is why unverbalized reasoning is a monitoring problem rather than a curiosity: a monitor that reads only the stated chain of thought is reading a narrated artifact, not the computation.
03
Status, and what this page deliberately is not
Accepted to an ICML 2026 workshop; I am a co-author. This digest is minimal by design: it records the acceptance and the paper's scope, and nothing further. The full digest — with the standard question-to-falsifier ledger and measured results — lands when the public artifact does. Until then, this page makes no claims the artifact cannot yet back.