Research map / Large language models / Language, safety and interpretability
Towards Robust and Parameter-Efficient Knowledge Unlearning 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 LoKU, combining an Inverted Hinge Loss replacement for gradient ascent with FILA, a Fisher-weighted LoRA initialization, for efficient and robust LLM unlearning on TDEC (GPT-Neo) and TOFU (Phi-1.5B, Llama2-7B).
Problem
Gradient ascent-based LLM unlearning is unstable and causes catastrophic forgetting of retained knowledge, and combining it with LoRA gives poor trade-offs between compute cost and post-unlearning generative performance.
Approach
Replaces the negative cross-entropy (GA) term with a bounded Inverted Hinge Loss that pushes down the true token while boosting the most likely competing token, and initializes LoRA adapters via a closed-form row-weighted low-rank approximation using relative Fisher information between forget and retain sets (FILA).
Key results
- On TDEC with GPT-Neo 1.3B at 0.8% trainable params, LoKU (IHL+FILA) reaches EL10 0.5% and MA 29.6% in 13 epochs, with Reasoning 48.3 (-1.5) and Dialogue F1 12.1 (+0.6), better than LoRA GD (Dialogue -1.8, Pile PPL +20.3).
- On TOFU Forget05 with Phi-1.5B at LoRA rank 4, IHL+FILA keeps model utility at 0.51 versus 0.50 for GA-based methods that drop it, while GD+FILA and GA+FILA collapse utility to near zero.
- Across TOFU, IHL+FILA maintains model utility close to the original model (e.g., 0.51-0.52) while KL, DPO, and NPO significantly reduce utility on larger forget sets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace GA loss with IHL in GD (ablation) | LLM unlearning on TDEC · TDEC | Dialogue F1: 5.9 -> 10.3 | Helped |
| Add FILA initialization to IHL with LoRA (ablation) | GPT-Neo unlearning · TDEC | Unlearning epochs: 11.4 -> 6.0 | Helped |
| Add FILA initialization to GD with LoRA (ablation) | GPT-Neo unlearning · TDEC | Dialogue F1: 2.5 -> 6.5 | Hurt |
| Replace GD's NCE loss with IHL (ablation) | LoRA unlearning · TOFU Forget05 | Model utility: 0.48 -> 0.51 | Helped |
| Use FILA initialization with IHL (ablation) | LoRA unlearning · TOFU Forget10 | Model utility: 0.51 -> 0.51 | No clear effect |
| Use FILA initialization with GA (ablation) | LoRA unlearning · TOFU Forget01 | Model utility: 0.52 -> 0.00 | Hurt |
| Adapt FFN layers in addition to attention layers (ablation) | LoRA unlearning on GPT-Neo 125M · TDEC | Unlearning success rate | Helped |
| Adapt Q and V attention layers only (ablation) | LoRA unlearning on GPT-Neo 125M · TDEC | Unlearning success rate | Hurt |
| Increase LoRA rank (ablation) | TOFU unlearning · TOFU Forget01 | Model utility | Helped |
| Unlearn a 10% author forget set with IHL+FILA at rank 8 or 16 (ablation) | Llama2-7B · TOFU Forget10 | Forget quality | Mixed |
Surprising findings
- FILA accelerates unlearning with GD but worsens downstream performance because it amplifies GA-style divergence, while paired with bounded IHL it improves retention.
- Applying LoRA only to query and value attention layers is insufficient to meet unlearning criteria within 20 epochs, while adapting FFN layers greatly increases success rate.
Limitations
- Unlearning can overshoot: with FILA on some TOFU settings forget quality drops after reaching the optimum (Streisand effect), and early stopping needs an oracle reference model.
- Evaluation relies on reference models for forget quality, which are not available in real-world settings.
- FILA uses a row-wise approximation of the weighted low-rank problem rather than exact WLRA.
Experimental setup
Datasets: Training Data Extraction Challenge (TDEC) The Pile WikiText TOFU
Benchmarks: TDEC TOFU LAMBADA HellaSwag Winogrande COPA ARC-Easy ARC-Challenge PIQA MathQA PubmedQA Blended Skill Talk
Metrics: EL10 memorization accuracy reasoning accuracy dialogue F1 Pile perplexity forget quality model utility
Baselines: GA GD KL DPO NPO PiSSA-style initialization
Builds on: gradient ascent unlearning gradient difference LoRA PiSSA TOFU
Closest papers by meaning
- Unified Parameter-Efficient Unlearning for LLMs (ICLR 2025 · similarity 0.93)
- Catastrophic Failure of LLM Unlearning via Quantization (ICLR 2025 · similarity 0.92)
- LLM Unlearning via Loss Adjustment with Only Forget Data (ICLR 2025 · similarity 0.92)
- A Closer Look at Machine Unlearning for Large Language Models (ICLR 2025 · similarity 0.92)
- RWKU: Benchmarking Real-World Knowledge Unlearning for Large Language Models (NeurIPS 2024 · similarity 0.92)
- On Large Language Model Continual Unlearning (ICLR 2025 · similarity 0.91)
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.