AI models that think out loud before answering turn out to be lousy at truly forgetting things.
Researchers have proposed a new method, called GUARD, to fix that for large reasoning models, whose chain-of-thought output makes forgetting harder because existing unlearning techniques only suppress a target fact or nudge internal representations without specifying what replaces it in that trace. That gap lets models hallucinate substitute details, produce malformed reasoning, or repeat themselves once the original content is blocked. GUARD instead trains models toward what its authors call a natural forgetting trajectory, a coherent, non-disclosing chain-of-thought followed by a stable refusal, built by distilling guided behavior from a frozen version of the model using guidance tokens, and the team also introduced a new metric, the Natural Forgetting Reasoning Score, to judge whether replacement reasoning is fluent and structurally sound rather than just checking if the secret leaked.
This matters because chain-of-thought is exactly where forgetting tends to fail quietly. A model can produce a clean, safe-looking final answer while its intermediate reasoning still spells out the protected fact or the unsafe instructions it was supposed to drop. As reasoning models get used in settings involving deletion requests, privacy takedowns, or safety filtering, scrubbing only the visible answer is a thin fix if the scratch work underneath still leaks.
The tests here cover two widely used distilled reasoning models and benchmarks the authors built or adapted themselves, which is a reasonable proof of concept but not independent confirmation that CoT unlearning is a solved problem.