Research map / Large language models / Language, safety and interpretability
GRU: Mitigating the Trade-off between Unlearning and Retention for LLMs
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.
Proposes Gradient Rectified Unlearning (GRU), which projects unlearning gradients to remove components conflicting with retain-data gradients, plus a retain-data-free variant (TRU) based on task vectors, to reduce the unlearning-retention trade-off in LLMs.
Problem
LLM unlearning removes targeted knowledge but typically damages performance on non-targeted data; existing methods trade unlearning strength against retention.
Approach
Rectify unlearning gradients by projecting out the component that conflicts with a retain-data gradient (GEM/PCGrad-style closed-form projection), with EMA smoothing of the retain gradient and norm clipping; for retain-free settings, apply the same rectification to task vectors computed from the remaining unlearning-set examples.
Key results
- On TOFU with Llama-2-7B (5%), GA with GRU improves FQ from -16.93 to -3.52 and MU from 0.00 to 0.52.
- On WMDP with Zephyr-7B-beta, NPO with GRU raises MMLU from 0.3033 to 0.3704 while Bio accuracy drops from 0.2655 to 0.2561.
- In the retain-data-free setting on TOFU, TRU beats the task vector and WGA/NPO baselines; e.g., Llama2-7B 10% MU 0.47 vs NPO 0.29.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add GRU to GA | GA on TOFU · TOFU 5% (LLaMA2-7B) | MU: 0.0000 -> 0.5190 | Helped |
| add GRU to NPO | NPO on TOFU · TOFU 10% (LLaMA2-7B) | FQ: -8.7037 -> -2.5106 | Helped |
| add GRU to NPO | NPO on TOFU · TOFU 10% (LLaMA2-7B) | MU: 0.2876 -> 0.4570 | Helped |
| add GRU to NPO | NPO on WMDP · WMDP-Bio (Zephyr-7B-beta) | QA accuracy: 0.2655 -> 0.2561 | Helped |
| add GRU to NPO+GD | NPO+GD on WMDP · WMDP-Cyber (Zephyr-7B-beta) | QA accuracy: 0.3493 -> 0.3524 | Hurt |
| add GRU to NPO+GD | NPO+GD on WMDP · MMLU (Zephyr-7B-beta) | QA accuracy: 0.4724 -> 0.5033 | Helped |
| add GRU to NPO+GD | NPO+GD on MUSE · MUSE KnowMem-R | KnowMem: 23.5565 -> 30.7492 | Helped |
| add GRU to RMU | RMU on WMDP · WMDP-Bio (Zephyr-7B-beta) | QA accuracy: 0.26 -> 0.26 | No clear effect |
| add GRU to RMU | RMU on WMDP · WMDP-Cyber (Zephyr-7B-beta) | QA accuracy: 0.31 -> 0.28 | Helped |
| use rectified task vectors without retain data | TRU · TOFU 5% (Phi-1.5, retain-data-free) | FQ: -11.91 -> -9.04 | Helped |
| use rectified task vectors without retain data | TRU · TOFU 5% (Phi-1.5, retain-data-free) | MU: 0.36 -> 0.40 | Helped |
Surprising findings
- Plain GA, which is usually considered too weak for retention, gains substantially in both removal and retention when rectified with GRU.
- Retain-data-free TV and GA can render unlearned models completely useless on TOFU, so retain-free unlearning is especially hard.
Limitations
- Relies on quality of retain data; biased or shifted retain sets can still harm retention.
- Theoretical guarantees rely on first-order/local linearity and smoothness assumptions.
- Retain-data-free TRU results remain weak in some settings, and the choice of baseline method is task-dependent.
Experimental setup
Datasets: TOFU WMDP MUSE
Benchmarks: TOFU WMDP MUSE
Metrics: FQ MU QA accuracy on WMDP-Bio/Cyber MMLU accuracy VerbMem KnowMem
Baselines: GA GD NPO WGA NPO+GD NPO+KL SimNPO+GD UWC RMU task vector
Builds on: gradient ascent unlearning negative preference optimization gradient episodic memory PCGrad task vectors
Closest papers by meaning
- Rethinking LLM Unlearning Objectives: A Gradient Perspective and Go Beyond (ICLR 2025 · similarity 0.93)
- A Closer Look at Machine Unlearning for Large Language Models (ICLR 2025 · similarity 0.91)
- OpenUnlearning: Accelerating LLM Unlearning via Unified Benchmarking of Methods and Metrics (NeurIPS 2025 · similarity 0.91)
- Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning (NeurIPS 2025 · similarity 0.91)
- LLM Unlearning via Loss Adjustment with Only Forget Data (ICLR 2025 · similarity 0.90)
- Reversing the Forget-Retain Objectives: An Efficient LLM Unlearning Framework from Logit Difference (NeurIPS 2024 · similarity 0.90)
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.