arXivPaperNeeds Review

Forgetting Only What Matters: Layer-Selective Unlearning toward Robust LLMs

FOM-UL targets model layers that affect unwanted memories while sparing retained knowledge, and reports that forgetting survives 8-bit and 4-bit quantization better than baselines.

arXiv · Sep 9, 2026

Open Source Open Markdown Open JSON

Source Summary

FOM-UL scores transformer layers by influence on the forget set versus sensitivity to the retain set, then updates only the strongest candidates. Across **TOFU**, **KnowUnDo**, and **MUSE-style** evaluations, it reports less residual memorization while keeping utility near the original model.

Practical Implication

Teams that must remove sensitive or copyrighted training content should evaluate unlearning after every deployment transformation. The method specifically tests **8-bit and 4-bit quantization**, where low-bit rounding can otherwise undo small, diffuse parameter changes, plus adversarial prompts that try to recover forgotten content.

Agent-Ready Context

FOM-UL scores transformer layers by influence on the forget set versus sensitivity to the retain set, then updates only the strongest candidates. Across **TOFU**, **KnowUnDo**, and **MUSE-style** evaluations, it reports less residual memorization while keeping utility near the original model.

Teams that must remove sensitive or copyrighted training content should evaluate unlearning after every deployment transformation. The method specifically tests **8-bit and 4-bit quantization**, where low-bit rounding can otherwise undo small, diffuse parameter changes, plus adversarial prompts that try to recover forgotten content.

This is empirical suppression, not proof of deletion. The paper explicitly offers no formal erasure guarantee, so builders should not treat lower benchmark recovery as evidence that targeted data can never be extracted.

Connected Context · Feed7 Judgment

This narrows utility-preserving unlearning to the layers most influential for forgetting and explicitly makes post-quantization recovery testing part of the deployment workflow. It supports targeted intervention over diffuse updates for this objective, but only demonstrates resistance to tested recovery methods; it does not convert empirical suppression into a deletion guarantee.

The Structure of Quantization Damage in LLMs: Why the Next Bit Should Be Spent GloballyQuantization damage was found to be broadly distributed, whereas FOM-UL concentrates unlearning updates in selected layers; together they show that layer selectivity is objective-dependent and must be retested after quantization.p-e-w/hereticBoth modify selected model behavior while trying to preserve general utility, but their aims oppose each other: FOM-UL suppresses targeted memorization, while Heretic removes refusal behavior and can weaken safety controls.CLEAR: Continuous Latent Adapter Routing for Utility-Preserving LLM Safety AlignmentCLEAR and FOM-UL both seek limited behavioral change without rewriting the entire model, but CLEAR conditionally routes safety behavior at inference while FOM-UL permanently updates selected parameters.The Implications of Linguistic Illegibility for LLM SecurityThe limits of language-based monitoring reinforce FOM-UL’s warning that observed non-recovery cannot prove erasure; external containment and data controls remain necessary.

Context Map

modelsecurity#open-models#agent-reliability

Uncertainty

This is empirical suppression, not proof of deletion. The paper explicitly offers no formal erasure guarantee, so builders should not treat lower benchmark recovery as evidence that targeted data can never be extracted.