LLM Unlearning with LLM Beliefs

2025-10-22

AI Summary

Large language models (LLMs) often memorize and regurgitate sensitive or harmful content; existing unlearning methods frequently induce a “squeezing effect”—where probability mass shifts toward semantically similar, high-likelihood paraphrases—resulting in illusory forgetting and misleading evaluation metrics. This work formally characterizes the squeezing effect for the first time and proposes a belief-steering-based unlearning framework: by jointly suppressing both target responses and the model’s own high-confidence generations via gradient ascent, we design two algorithms—BS-T (token-level) and BS-S (sequence-level). Extensive experiments across multiple LLMs and benchmarks demonstrate that our approach significantly reduces target content re-generation while preserving overall model utility. It achieves more thorough, verifiable, and genuine forgetting—advancing trustworthy model editing with a novel, principled paradigm.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

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.

Problem

• Preventing harmful content memorization in large language models
• Addressing semantic rephrasing leakage during unlearning processes
• Correcting misleading automated metrics for true unlearning evaluation

Innovation

• Bootstrapping framework counters squeezing effect
• Suppresses target responses and model beliefs
• Achieves thorough forgetting while preserving utility