Research map / Large language models / Language, safety and interpretability
A Closer Look at Machine Unlearning for Large Language Models
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.
Re-examines LLM unlearning evaluation with three extra metrics (token entropy, cosine similarity, entailment score) and proposes maximizing-entropy (ME) loss for untargeted unlearning and answer-preservation (AP) regularization for targeted unlearning, evaluated on TOFU and a real-world individuals setting.
Problem
LLM unlearning evaluation relies mainly on ROUGE, ignoring output diversity, semantics and factuality; untargeted unlearning targets an unpredictable behavior that can induce hallucinations, and targeted unlearning with existing regularizers causes excessive ignorance on the retain set.
Approach
Replace gradient ascent's target with a maximum-entropy (uniform next-token distribution) forget objective for untargeted unlearning, and add an answer-preservation loss, a preference-style regularizer with no reference model, that keeps original answers' probability while suppressing rejection templates for targeted unlearning.
Key results
- On TOFU forget10, ME+GD reaches MU 0.7357 and FE 0.9489 (avg 0.8423), versus best baseline avg of 0.7469 (GA+GD) among baselines.
- On TOFU forget10, IDK+AP keeps MU 0.7471 and FE 0.7433 (avg 0.7452), while IDK+GD and DPO baselines drop MU to near zero.
- On real-world unlearning, ME+GD gets the highest average downstream score (0.5294 vs 0.5662 initial) with MU 0.4901 and FE 0.9312; IDK+AP reaches MU 0.5311 versus 0.0000 for baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use maximizing entropy with GD regularization instead of GA- or NPO-based untargeted unlearning | Llama2-chat-7B · TOFU forget05 | MU: 0.5465 -> 0.7472 | Helped |
| Use maximizing entropy with GD regularization instead of GA- or NPO-based untargeted unlearning | Llama2-chat-7B · TOFU forget05 | FE: 0.9303 -> 0.9313 | No clear effect |
| Add answer preservation loss to IDK targeted unlearning | Llama2-chat-7B · TOFU forget05 | MU: 0.0000 -> 0.7529 | Helped |
| Add answer preservation loss to IDK targeted unlearning | Llama-3-8B-Instruct · Real-world unlearning | Model Utility: 0.0000 -> 0.5311 | Helped |
| Replace GD retain regularization with KL regularization (ablation) | ME unlearning · TOFU forget05 | MU: 0.7472 -> 0.6213 | Hurt |
| Replace GD retain regularization with KL regularization (ablation) | ME unlearning · TOFU forget05 | FE: 0.9313 -> 0.9715 | Helped |
| Use AP instead of GD retain regularization for DPO (ablation) | DPO targeted unlearning · TOFU forget05 | MU: 0.0000 -> 0.7173 | Helped |
| Supplement the shrinking retain set with unrelated Alpaca samples to keep it at 2000 samples (ablation) | IDK+AP · Continual TOFU forget10 | MU | Helped |
| Keep the NPO reference model fixed to the initial model across continual tasks (ablation) | NPO-based unlearning · Continual TOFU forget01 | MU | Helped |
| Use the initial model as the fixed NPO reference across continual tasks (ablation) | NPO-based unlearning · Continual TOFU forget01 | MU | Hurt |
Surprising findings
- Surrogate retain models on TOFU produce hallucinated outputs on the forget set, with GPT-4o judging about 74% of forget-set answers as hallucinations.
- In continual unlearning, NPO's utility collapses to zero on forget01 mainly because the reference model changes between subtasks; fixing the reference to the initial model stabilizes it.
Limitations
- IDK+AP's utility declines in the last subtasks of continual forget10 because the retain set shrinks, requiring extra data to mitigate.
- Hyperparameters (alpha, learning rates, epochs) differ across scenarios and are tuned per method, with the real-world results reported as best across configurations.
Experimental setup
Datasets: TOFU Alpaca real-world individuals unlearning dataset
Benchmarks: TOFU MMLU ARC-c GSM8K TriviaQA TruthfulQA
Metrics: ROUGE probability truth ratio token entropy cosine similarity entailment score model utility forget efficacy
Baselines: GA+GD GA+KL NPO+GD NPO+KL DPO+GD DPO+KL IDK+GD
Builds on: gradient ascent NPO DPO IDK fine-tuning TOFU
Closest papers by meaning
- MUSE: Machine Unlearning Six-Way Evaluation for Language Models (ICLR 2025 · similarity 0.94)
- On Large Language Model Continual Unlearning (ICLR 2025 · similarity 0.93)
- Catastrophic Failure of LLM Unlearning via Quantization (ICLR 2025 · similarity 0.93)
- Towards Effective Evaluations and Comparisons for LLM Unlearning Methods (ICLR 2025 · similarity 0.93)
- LLM Unlearning via Loss Adjustment with Only Forget Data (ICLR 2025 · similarity 0.93)
- Wisdom is Knowing What not to Say: Hallucination-Free LLMs Unlearning via Attention Shifting (NeurIPS 2025 · similarity 0.93)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.