DynEdit uses token entropy to guide long-sequence LLM edits

Token entropy, the measure of how uncertain a language model is at each position, tracks two things at once in long passages: how hard a factual edit will be, and where the important content sits. That is the claim behind DynEdit, a paper from Jinhu Fu, Yan Bai, Yihang Lou, Li Sun, Jiahong Wu, Xiangxiang Chu and Sen Su, accepted as a poster at NeurIPS 2026.

The authors argue that current parameter-editing methods leave long-sequence knowledge updates underserved. 'Single-point editors provide insufficient coverage for extended targets, while fixed chunk-based editors distribute editing effort uniformly across rigid predefined windows, failing to prioritize difficult, information-dense tokens,' they write. DynEdit instead uses token-level entropy to drive a non-uniform sliding window across the target, then weights the optimization inside each window toward the high-entropy tokens.

Across three benchmarks and two backbone models, the paper reports gains of 'up to +14.0 BLEU points on challenging paraphrase queries' and 'near-saturated semantic scores on several diverse knowledge domains.' The abstract names no specific benchmarks, backbones, or baselines beaten. Two researchers in our Who's Who directory circulated the paper after it went up.