LLM Unlearning with LLM Beliefs
This addresses a critical safety issue for AI developers and users by improving unlearning to prevent harmful content resurfacing, though it is an incremental advance over prior gradient-based methods.
The paper tackles the problem of unlearning sensitive content in large language models, where existing methods cause a 'squeezing effect' that redistributes probability to related rephrasings, leading to spurious unlearning. The proposed bootstrapping framework uses model beliefs to counter this effect, achieving more thorough forgetting while preserving utility, as confirmed by extensive experiments across diverse benchmarks.
Large language models trained on vast corpora inherently risk memorizing sensitive or harmful content, which may later resurface in their outputs. Prevailing unlearning methods generally rely on gradient ascent and its variants to lower the probability of specific target responses. However, we find that this strategy induces a critical side effect: probability mass is redistributed into high-likelihood regions, often corresponding to semantically related rephrasings of the targets. We refer to this as the squeezing effect, which explains why many methods yield merely spurious unlearning, a problem further obscured by automated metrics (e.g., ROUGE, truth ratio) that misreport actual success. To address this, we propose a bootstrapping (BS) framework that explicitly links the squeezing effect with the model's own high-confidence generations, namely its model beliefs. Since model beliefs inherently capture the very high-likelihood regions where probability mass is squeezed, incorporating them into the unlearning objective directly counters the squeezing effect. By jointly suppressing both target responses and model beliefs, BS-T (token) attenuates high-probability tokens, whereas BS-S (sequence) removes entire high-confidence generations, together achieving more thorough forgetting while preserving utility. Extensive experiments across diverse benchmarks with various model families confirm the effectiveness of our approach.