AI models can now forget secrets without losing reasoning ability

Researchers built a method to make large reasoning models unlearn sensitive information (copyrighted text, private data) without degrading their ability to think through complex problems. The approach identifies which parts of a model's reasoning chain contain the sensitive stuff and replaces them with harmless placeholders that keep the logic intact.

Large reasoning models show their work — they output the chain of thought before the answer, which means sensitive information leaks through the intermediate steps, not just the final output. Existing unlearning methods either only target final answers or break the model's reasoning altogether. This matters because companies deploying these models now have a way to remove copyrighted material or private data without crippling the system. The practical effect: a deployed model can be patched to forget something without retraining from scratch.