LLM is fine-tuned by means of reinforcement learning to improve its response results. The advantages of forgetting learning are that only negative samples (harmful samples) are required, the computational cost is low, and it is particularly effective when the training samples that lead to the harmful behavior of LLM are known. Experiments show that forgetting learning performs well on three application scenarios: removing harmful outputs, removing infringing protected content, and eliminating LLM illusions.
