Getting an AI model to forget something turns out to be surprisingly hard to do cleanly.
Researchers have proposed DareU, an unlearning framework that takes a different approach to scrubbing data from large language models. Most existing methods work by cranking up the prediction loss on whatever the model is supposed to forget — essentially penalizing the model for remembering. The problem is that this tends to overshoot, degrading the model's usefulness on unrelated tasks. DareU sidesteps that by reframing the goal: instead of maximizing loss, it uses reinforcement learning to reduce how much the model's outputs can be traced back to specific training data. The team calls this "de-attribution."
This matters because LLM unlearning is quickly becoming a compliance problem, not just a research curiosity. As regulators push for the right to be forgotten and copyright disputes over training data pile up, AI labs need reliable tools to surgically remove specific content from deployed models without retraining from scratch. DareU's approach — targeting data attribution scores rather than raw prediction error — offers a more precise lever.
Whether attribution-based scoring scales to the size of models actually in production is the next question nobody has answered yet.