Research map / Large language models / Language, safety and interpretability

Mechanistic Unlearning: Robust Knowledge Unlearning and Editing via Mechanistic Localization

ICML 2025 · Large language models › Language, safety and interpretability · approach: Reasoning mechanisms and failure modes · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Shows that restricting knowledge edits and unlearning to MLP components implementing a mechanistically identified fact-lookup mechanism yields more robust, less side-effectful edits than output-tracing localizations or full fine-tuning.

Problem

Knowledge editing and unlearning in LLMs often produce edits that are not robust to prompt reformatting, relearning, or latent knowledge extraction, and localization methods based on output effects may fail to remove underlying knowledge.

Approach

Localize edits to the MLP components that implement a fact-lookup mechanism (enriching the subject representation in the latent stream) rather than output-tracing components (causal tracing, attribution patching) that directly write the answer, then fine-tune only those components with an injection, retain, and Pile SFT loss.

Key results

• FLU localization gives MCQ forget error and edit accuracy superior to causal tracing and nonlocalized editing; in Sports-Athlete-Editing only FLU edits generalize meaningfully to MCQ, exceeding others by more than 40% in MCQ edit accuracy.
• Adversarial relearning recovers up to 63% of forgotten information for Causal Tracing methods versus about 20% for FLU localization.
• On Sports-Unlearning (basketball), FLU localization reaches MCQ forget accuracy 0.110 versus 0.196 for nonlocalized and 0.849 for causal tracing, while keeping MMLU at 0.613.

Measured effects reported in the paper

Surprising findings

• Output-tracing localizations (causal tracing, attribution patching) often perform equal to or worse than updating the entire model on robustness, relearning, and latent knowledge probes.
• Causal Tracing and Attribution Patching assign high scores to later-layer extraction MLPs rather than the fact-lookup MLPs, so they largely modify extraction mechanisms.

Limitations

• FLU localization is manually derived per dataset and model through probing and path patching, not automated.
• Relearning evaluation is uninformative on CounterFact tasks and Full-Sports-Editing, where all localizations relearn little or regain 100% accuracy.
• Only attention heads outside layer 2 were excluded from FLU, so the localization is an interpretive choice that may not generalize to other facts.

Experimental setup

Datasets: Sports Facts CounterFact Pile MMLU

Benchmarks: MMLU Sports Facts CounterFact

Metrics: forget accuracy edit accuracy maintain accuracy MCQ forget/edit accuracy paraphrase edit accuracy neighborhood edit error relearning accuracy MMLU accuracy probe accuracy

Baselines: Causal Tracing Attribution Patching Random All-MLPs Nonlocalized Random-MLPs

Builds on: Causal Tracing (ROME) attribution patching path patching ROME NPO

Closest papers by meaning

• Adaptive Localization of Knowledge Negation for Continual LLM Unlearning (ICML 2025 · similarity 0.89)
• Towards Robust Knowledge Unlearning: An Adversarial Framework for Assessing and Improving Unlearning Robustness in Large Language Models (AAAI 2025 · similarity 0.88)
• In-Context Unlearning: Language Models as Few-Shot Unlearners (ICML 2024 · similarity 0.88)
• RWKU: Benchmarking Real-World Knowledge Unlearning for Large Language Models (NeurIPS 2024 · similarity 0.88)
• Do LLMs Really Forget? Evaluating Unlearning with Knowledge Correlation and Confidence Awareness (NeurIPS 2025 · similarity 0.88)
• Wisdom is Knowing What not to Say: Hallucination-Free LLMs Unlearning via Attention Shifting (NeurIPS 2025 · similarity 0.88)