Research map / Large language models / Retrieval and knowledge
Large Scale Knowledge Washing
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 LaW, an MEMIT-inspired method that washes large sets of factual knowledge from decoder-only LLM MLP layers by optimizing a norm-constrained objective that maximizes output change on target keys while bounding change on general keys.
Problem
Unlearning very large sets of factual knowledge from LLMs without degrading fluency or reasoning ability, where gradient-based reverse-loss unlearning can destroy the model at scale.
Approach
Instead of editing facts to an end-of-sequence token, LaW removes knowledge by maximizing the perturbation of MLP outputs on target keys K_w subject to a norm bound on perturbation over general keys K (estimated from WikiText), initialized from MEMIT's delta and applied layer by layer with successive elimination of already-washed facts.
Key results
- On Wiki-Latest with GPT2-XL, LaW reduces knowledge accuracy to 0.1926 (QA-F1 0.1735) while keeping reasoning accuracy at 0.4832 vs 0.5105 original; MEMIT leaves 0.2972 accuracy, and most fine-tuning baselines collapse (perplexity >10^10).
- On zsRE with GPT2-XL, LaW reaches 0.0050 accuracy with reasoning average 0.5105, matching the original model.
- LaW keeps fluency close to MEMIT (log perplexity 2.75 vs 2.72 on Wiki-Latest for GPT2-XL) while SeUL rises to 96.95.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LaW instead of MEMIT | GPT2-XL · zsRE | accuracy: 0.0462 -> 0.0050 | Helped |
| Use LaW instead of MEMIT | GPT-J-6B · CounterFactual | QA-F1: 0.0759 -> 0.0125 | Helped |
| Use LaW instead of MEMIT | GPT2-XL · Wiki-Latest | reasoning accuracy: 0.5029 -> 0.4832 | Hurt |
| Initialize the update from MEMIT instead of random initialization (ablation) | LaW on GPT2-XL · zsRE | accuracy: 0.8845 -> 0.0008 | Helped |
| Initialize the update from MEMIT instead of random initialization (ablation) | LaW on GPT2-XL · CounterFactual | accuracy: 0.9158 -> 0.1258 | Helped |
| Increase β from 1.05β₀ to 1.5β₀ (ablation) | LaW on GPT2-XL · zsRE | accuracy: 0.0074 -> 0.0000 | Helped |
| Increase β from 1.05β₀ to 1.5β₀ (ablation) | LaW on GPT2-XL · zsRE | Avg_Acc: 0.5112 -> 0.5010 | Hurt |
| Increase β from 1.05β₀ to 1.5β₀ (ablation) | LaW on GPT2-XL · CounterFactual | accuracy: 0.1266 -> 0.0655 | Helped |
| Increase β from 1.05β₀ to 1.5β₀ (ablation) | LaW on GPT2-XL · CounterFactual | Avg_Acc: 0.4910 -> 0.4602 | Hurt |
| Add successive elimination of already-forgotten facts between layer updates (ablation) | LaW on GPT2-XL · Wiki-Latest | accuracy: 0.2398 -> 0.1926 | Helped |
| Add successive elimination of already-forgotten facts between layer updates (ablation) | LaW on GPT2-XL · Wiki-Latest | reasoning accuracy: 0.4788 -> 0.4657 | Hurt |
Surprising findings
- Random initialization of the delta matrix leads to very poor washing (accuracy 0.8845 on zsRE), so MEMIT initialization is essential.
- Several baselines (FT-UL, ME-FT, SeUL) destroy the model at large scale, with perplexity exceeding 10^10 on Wiki-Latest.
Limitations
- Only applicable to knowledge expressed as triplets convertible to sentences, not free text.
- Experiments limited to GPT2-XL and GPT-J-6B; MEMIT adaptation to LLaMA2-7B reached efficacy below 0.75, so the method was not validated on newer LLMs.
- Washed model outputs random or null answers, which may cause hallucination without further instruction tuning.
Experimental setup
Datasets: zsRE CounterFactual Wiki-Latest WikiText fineweb-edu
Benchmarks: zsRE CounterFactual Wiki-Latest LAMBADA (lambda_openai) HellaSwag ARC-Easy GSM8K fineweb-edu perplexity
Metrics: accuracy QA-F1 average reasoning accuracy perplexity
Baselines: FT MEMIT ME-FT FT-UL WOH SeUL
Builds on: MEMIT ROME Who is Harry Potter unlearning SeUL ME-FT
Closest papers by meaning
- RWKU: Benchmarking Real-World Knowledge Unlearning for Large Language Models (NeurIPS 2024 · similarity 0.90)
- A Closer Look at Machine Unlearning for Large Language Models (ICLR 2025 · similarity 0.89)
- Large Language Model Unlearning via Embedding-Corrupted Prompts (NeurIPS 2024 · similarity 0.88)
- Towards Reasoning-Preserving Unlearning in Multimodal Large Language Models (CVPR 2026 · similarity 0.88)
- SHRED: Retain-Set-Free Unlearning via Self-Distillation with Logit Demotion (NeurIPS 2026 · similarity 0.88)
- Towards Robust and Parameter-Efficient Knowledge Unlearning for LLMs (ICLR 2025 · similarity 0.88)
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.