Unlearning in large reasoning models (LRMs) must erase forget‑set information from both the intermediate reasoning trace and the final answer. Existing methods often cause reasoning collapse: the chain‑of‑thought (CoT) becomes incoherent and the reasoning process fails to terminate properly.
To tackle this issue, we introduce REDaCT (REgenerate via Detection and Corrective Tuning), an inference‑only unlearning framework. REDaCT consists of two main components:
- A binary leak detector that scans the reasoning trace sentence by sentence and flags any sentence that leaks forget‑set data.
- An embedding optimization module that learns a corrective embedding offline; when the first leak is detected during inference, the embedding is injected to steer subsequent reasoning toward sentences that do not leak the forget‑set.
Results show that REDaCT completely avoids reasoning collapse, substantially improves reasoning preservation while keeping answer forgetting perfect and CoT forgetting near‑perfect. The implementation is publicly available at https://github.com/jmaccoby/redact.
Review