Research map / Large language models / Retrieval and knowledge
Mitigating Heterogeneous Token Overfitting in LLM Knowledge Editing
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.
Identifies heterogeneous token overfitting in LLM knowledge editing and proposes OVERTONE, a token-level adaptive target smoothing loss that improves portability and locality across four editing methods on LLaMA 2 and LLaMA 3.
Problem
Knowledge editing of LLMs with single-instance updates overfits tokens at different rates, degrading portability (multi-hop reasoning about new knowledge) and generality while locality is also affected.
Approach
Replace the one-hot CE target per token with an adaptive smoothed target that mixes the one-hot label with the model's own top-nσ-filtered predicted distribution, applied only when the label is already the argmax, plus a clipped forward-KL loss acting as token-level early stopping; shown to generalize CE and to relate to an implicit DPO objective without preference data.
Key results
- On ZsRE with LLaMA 2, LoRA + OVERTONE raises average score from 63.45 to 85.67, with portability rising from 23.34 to 61.16.
- FT-M with OVERTONE on Wiki_counterfact (LLaMA 2) improves average from 76.52 to 82.99, and MELO on Wiki_recent rises from 64.71 to 75.58.
- On MQuAKE-2002 with LLaMA 2, LoRA average improves from 78.0 to 80.25 and FT-M from 71.0 to 75.29.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply OVERTONE to LoRA | LoRA on LLaMA 2-7B-Chat · ZsRE | portability: 23.34 -> 61.16 | Helped |
| Apply OVERTONE to LoRA | LoRA on LLaMA 2-7B-Chat · WikiData counterfact | locality: 9.99 -> 57.85 | Helped |
| Apply OVERTONE to LoRA | LoRA on LLaMA 3-8B-Instruct · ZsRE | portability: 26.55 -> 51.57 | Helped |
| Apply OVERTONE to LoRA | LoRA on LLaMA 3-8B-Instruct · WikiData counterfact | locality: 9.02 -> 57.54 | Helped |
| Apply OVERTONE to FT-M | FT-M on LLaMA 2-7B-Chat · MQuAKE-2002 | Avg.: 71.0 -> 75.29 | Helped |
| Apply OVERTONE to LoRA | LoRA on LLaMA 2-7B-Chat · MQuAKE-2002 | Avg.: 78.0 -> 80.25 | Helped |
| Apply OVERTONE to FT-M | FT-M on LLaMA 3-8B-Instruct · MQuAKE-2002 | Avg.: 68.67 -> 71.67 | Helped |
| Apply OVERTONE to LoRA | LoRA on LLaMA 3-8B-Instruct · MQuAKE-2002 | Avg.: 77.67 -> 78.33 | Helped |
| Add token-level clipping (ablation) | OVERTONE with LoRA on LLaMA 2-7B-Chat · ZsRE | portability: 26.6 -> 61.16 | Helped |
| Use dynamic filtered prediction rather than unedited prediction (ablation) | OVERTONE with LoRA on LLaMA 2-7B-Chat · ZsRE | Avg.: 71.18 -> 85.67 | Helped |
| Check that the label has the largest probability before mixing (ablation) | OVERTONE with LoRA on LLaMA 2-7B-Chat · ZsRE | Avg.: 82.47 -> 85.67 | Helped |
| Filter noisy tail regions from the predicted distribution (ablation) | OVERTONE with LoRA on LLaMA 2-7B-Chat · ZsRE | generality: 83.93 -> 94.31 | Helped |
Surprising findings
- Plain token-level smoothing without clipping raises portability and locality, but the full method with mixing and skip mechanism is needed for best results.
- Using the unedited model's own predicted distribution without dynamic update causes a significant drop, attributed to outdated internal knowledge conflicts.
Limitations
- Evaluation averages over editing samples without multiple runs per sample, which the authors acknowledge limits reliability.
- Only applied to knowledge editing with teacher-forced CE-based objectives; generalization to free-form editing data is left to future work.
- Gains on some configurations are mixed, e.g., reliability slightly drops for some methods and WISE gains are modest.
Experimental setup
Datasets: ZsRE WikiData_recent WikiData_counterfact WikiBio MQuAKE-2002
Benchmarks: ZsRE WikiData_recent WikiData_counterfact WikiBio MQuAKE-2002
Metrics: reliability generality portability locality average single-hop accuracy multi-hop accuracy
Baselines: ROME MEMIT FT-M LoRA MELO WISE
Builds on: ROME MEMIT WISE MELO DPO top-nσ sampling
Closest papers by meaning
- Uncovering Overfitting in Large Language Model Editing (ICLR 2025 · similarity 0.92)
- UniEdit: A Unified Knowledge Editing Benchmark for Large Language Models (NeurIPS 2025 · similarity 0.91)
- Unveiling the Pitfalls of Knowledge Editing for Large Language Models (ICLR 2024 · similarity 0.90)
- AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models (ICLR 2025 · similarity 0.89)
- AnyEdit: Edit Any Knowledge Encoded in Language Models (ICML 2025 · similarity 0.89)
- Golden Layers and Where to Find Them: Improved Knowledge Editing for Large Language Models Via Layer Gradient Analysis (NeurIPS 2026 · similarity 0.89)
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.