Research map / Large language models / Retrieval and knowledge
DynEdit: Dynamic Entropy-Guided Sequential Editing for Large Language Models
Jinhu Fu, Yan Bai, Yihang Lou, Li Sun, Jiahong Wu, Xiangxiang Chu, Sen Su
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 DynEdit, a knowledge-editing method for LLMs that uses dynamic entropy-guided segmentation to sequentially edit long-sequence knowledge targets instead of fixed chunks.
Problem
Parameter-based knowledge editing struggles with long-sequence targets: single-point editors give insufficient coverage, while fixed chunk-based editors spread effort uniformly and fail to focus on difficult, information-dense tokens.
Approach
Uses token-level entropy to dynamically partition long editing targets into variable-length windows so that editing effort concentrates on high-entropy, information-dense spans, applied sequentially.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add entropy-guided sliding and token weighting to AnyEdit (MEMIT) | AnyEdit · UnKE benchmark Paraphrase (Llama-3-8B) | BLEU: 66.29 -> 80.3 | Improved |
| add entropy-guided sliding and token weighting to AnyEdit (MEMIT) | AnyEdit · UnKE benchmark Original (Llama-3-8B) | BLEU: 77.91 -> 91.6 | Improved |
| add entropy-guided editing to UnKE-based AnyEdit* | AnyEdit* · UnKE benchmark Paraphrase (Qwen-2.5-7B) | BLEU: 47.17 -> 66.13 | Improved |
| add entropy-guided editing to UnKE-based AnyEdit* | AnyEdit* · UnKE benchmark Original (Llama-3-8B) | BLEU: 99.54 -> 99.19 | No clear effect |
| add entropy-guided editing to UnKE-based AnyEdit* | AnyEdit* · CounterFact Original (Llama-3-8B) | BLEU: 99.93 -> 98.26 | Hurt |
| add entropy-guided editing to MEMIT-based AnyEdit | AnyEdit · MQuAKE Original (Llama-3-8B) | Rouge-L: 94.28 -> 93.11 | Hurt |
| add entropy-guided editing to AlphaEdit+AnyEdit | AlphaEdit+AnyEdit · UnKEBench Original (Qwen3-8B) | BLEU: 40.88 -> 68.16 | Improved |
| add entropy-guided editing to MEMIT+AnyEdit | MEMIT+AnyEdit · UnKEBench Original (Qwen3-8B) | BLEU: 31.29 -> 44.52 | Improved |
| add entropy-guided editing to AnyEdit | AnyEdit · EditEverything math (Llama-3-8B) | Rouge-L: 70.73 -> 97.76 | Improved |
| add dynamic entropy-guided sliding strategy (ablation) | DynEdit · UnKE, CounterFact, MQuAKE BLEU | BLEU | Improved |
| add entropy-guided token weighting (ablation) | DynEdit · UnKE, CounterFact, MQuAKE BLEU | BLEU | Improved |
| use entropy-guided editing instead of fixed chunks on MMLU and IFEval | AnyEdit · MMLU, IFEval | accuracy | No clear effect |
Experimental setup
Baselines: single-point editors fixed chunk-based editors
Closest papers by meaning
- AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models (ICLR 2025 · similarity 0.90)
- UniEdit: A Unified Knowledge Editing Benchmark for Large Language Models (NeurIPS 2025 · similarity 0.90)
- Mitigating Heterogeneous Token Overfitting in LLM Knowledge Editing (ICML 2025 · similarity 0.90)
- WISE: Rethinking the Knowledge Memory for Lifelong Model Editing of Large Language Models (NeurIPS 2024 · similarity 0.89)
- Uncovering Overfitting in Large Language Model Editing (ICLR 2025 · similarity 0.89)
- In-Context Editing: Learning Knowledge from Self-Induced Distributions (ICLR 2025 · 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.
