A new unlearning method keeps AI reasoning models coherent even after scrubbing specific data from their memory.
Researchers built REDaCT, a technique for making large reasoning models (the kind that write out step-by-step chains of thought before answering, like OpenAI's o1 or DeepSeek's R1) forget specific information without breaking their ability to reason. The system works in two stages. A detector scans the reasoning trace in real time to spot sentences that leak information from the targeted forget-set data. The moment a leak is caught, a separately trained correction is injected, steering the model's reasoning away from the forbidden material instead of letting it ramble or never terminate. The researchers report the method erased targeted information from both the reasoning trace and the final answer while keeping the model's broader reasoning intact, and they have published the code on GitHub.
Unlearning matters because retraining a large model from scratch every time it needs to drop a dataset, say after a copyright dispute or a privacy request, is too expensive to do routinely. Reasoning models made the problem worse. Earlier fixes that silenced forbidden information also tended to silence the model's ability to think clearly, leaving it stuck rambling or never reaching an answer. REDaCT's inference-time patch avoids a full retrain while, according to the paper's own tests, also avoiding that collapse.
That said, this is a single preprint benchmarked by its own authors, and near-perfect chain-of-thought forgetting still leaves a gap, which is exactly what regulators and plaintiffs tend to care about.