Research map / Large language models / Language, safety and interpretability
Wisdom is Knowing What not to Say: Hallucination-Free LLMs Unlearning via Attention Shifting
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 Attention-Shifting (AS), an adapter-based LLM unlearning method that suppresses attention to fact-bearing tokens in the forget set and reinforces attention to important tokens in retained data, aiming for hallucination-free unlearning.
Problem
LLM unlearning faces a dilemma: aggressive methods degrade model utility, while conservative logit-level methods preserve utility but cause hallucinated responses about unlearned content.
Approach
Instead of manipulating logits or replacing outputs, shift attention away from important (fact-bearing) tokens in the unlearning set via KL-based attention targets trained on lightweight adapters, while reinforcing attention to semantically important tokens on retained data through a dual-loss objective.
Key results
- On ToFU, AS reaches NEK ROUGE-L 0.73 and GEK accuracy 0.80 on the 5-author forget set, versus 0.32 GEK for GA, while keeping TUD ROUGE-L at 0.16 (paper claims up to 15% accuracy gains).
- AS reports 0% reproduction and 0% hallucination rate on ToFU per GPT-4 judging, while IHL and ULD show elevated reproduction and hallucination.
- AS has a Top-50 exclusion rate of 0.93 on original TUD queries and remains above 0.92 under rephrased, perturbed, and noised prompts.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use attention shifting for unlearning instead of the strongest reported baseline | tofu_ft_llama2-7b · ToFU NEK | accuracy: 0.70 -> 0.76 | Helped |
| Use attention shifting rather than ULD | tofu_ft_llama2-7b · ToFU TUD (100 samples) | FQ: 0.89 -> 0.17 | Hurt |
| Use attention shifting rather than ULD | tofu_ft_llama2-7b · ToFU TUD (100 samples) | TR: 0.47 -> 0.97 | Helped |
| Use attention shifting rather than IHL + CE | tofu_ft_llama2-7b · ToFU NEK | accuracy: 0.73 -> 0.76 | Helped |
| Use attention shifting for target unlearning | tofu_ft_llama2-7b · ToFU TUD | hallucination rate: 0% | Helped |
| Use ASP alone by removing AKL (ablation) | AS · ToFU retained knowledge | model utility | Hurt |
| Replace AKL with CE or KL retention losses (ablation) | AS · ToFU retained knowledge | model utility | Mixed |
| Place adapters after Q and K rather than after O (ablation) | AS · ToFU adapter placement | Unlearning ΔAcc (%): -58.2 -> -98.8 | Helped |
| Place adapter after V rather than after O (ablation) | AS · ToFU adapter placement | Unlearning ΔAcc (%): -58.2 -> -97.9 | Helped |
| Raise token-importance threshold from 20% to 60% (ablation) | AS · ToFU | Unlearn ΔAcc (%): -67.3 -> -94.6 | Helped |
| Raise suppression strength λ from 0.20 to 0.99 (ablation) | AS · ToFU TUD | TUD ΔROUGE-L (%): -69.3 -> -98.8 | Helped |
| Use dynamic α instead of fixed α (ablation) | AS · ToFU NEK | NEK ΔROUGE-L (%): 0.66 -> 0.74 | Helped |
Surprising findings
- Strong suppression (lambda=0.99) yields maximal forgetting with zero hallucination, which the authors attribute to the model lacking content-word cues to reconstruct the target.
- Forget Quality for AS (0.17) is lower than for IHL and ULD, the standard metric disfavors refusal-style behavior that the authors argue is preferable.
Limitations
- Performance is weaker on TDEC's noisy web data, where token-importance estimation is a bottleneck.
- Behavioral suppression does not remove latent representations; traces may remain and retraining can reintroduce knowledge.
- Suitable mainly for refusal-oriented unlearning, less so for partial updates or creative generation.
Experimental setup
Datasets: ToFU TDEC Wikitext LAMBADA PubMedQA
Benchmarks: ToFU TDEC Wikitext LAMBADA PubMedQA
Metrics: ROUGE-L Forget Quality Top-k exclusion rate accuracy hallucination rate reproduction rate
Baselines: GA NPO IHL ULD
Builds on: Gradient Ascent unlearning Negative Preference Optimization ULD SAR (Shifting Attention to Relevance) LoRA
Closest papers by meaning
- A Closer Look at Machine Unlearning for Large Language Models (ICLR 2025 · similarity 0.93)
- Do LLMs Really Forget? Evaluating Unlearning with Knowledge Correlation and Confidence Awareness (NeurIPS 2025 · similarity 0.92)
- Towards Effective Evaluations and Comparisons for LLM Unlearning Methods (ICLR 2025 · similarity 0.92)
- Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy (NeurIPS 2025 · similarity 0.92)
- In-Context Unlearning: Language Models as Few-Shot Unlearners (ICML 2024 · similarity 0.92)
- Catastrophic Failure of LLM Unlearning via Quantization (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.