Recent large reasoning models (LRMs) can expose protected facts or unsafe rationales in their chain‑of‑thought (CoT) traces, making machine unlearning increasingly difficult. Existing unlearning objectives usually only suppress the target content or redirect internal representations, without specifying how the post‑forgetting trajectory should proceed, which often leads to hallucinated substitutes, fuzzy boundaries, or repetitive outputs. We argue that LRM unlearning should follow a natural forgetting trajectory: a coherent non‑disclosing CoT followed by a stable refusal‑style answer that replaces the original disclosure. To achieve this, we propose Guided Answer‑Reasoning Distillation (GUARD). The method first converts unsafe disclosures generated by the model into safe‑exit trajectories, then aligns a frozen LRM using guidance tokens, and finally distills the guided behavior back into the model parameters. To assess replacement quality beyond mere leakage, we introduce the Natural Forgetting Reasoning Score (NFRS), which evaluates structural stability, fluency, and unsupported substitutes in the forgotten outputs. Extensive experiments on the R‑TOFU benchmark and a STAR‑1‑derived harmful‑intent setting show that GUARD markedly reduces unsafe and privacy disclosures in two widely used distilled LRMs while preserving reasoning utility. The code is released at https://github.com/zeyu-Yan/GUARD.
Review