AnyEdit++: Adaptive Long-Form Knowledge Editing via Bayesian Surprise

2026-05-31

AI Summary

This work addresses the challenge that large language models often produce incoherent or factually inconsistent edits when modifying complex long-form texts, primarily due to fixed-window chunking that disregards logical structure. To overcome this limitation, the authors propose AnyEdit++, which introduces Bayes-Chunk—a Bayesian surprise-driven adaptive segmentation mechanism that dynamically identifies semantic boundaries to enable structure-aware editing. Grounded in two theoretical principles, structural independence and causal locality, the method incorporates geometrically orthogonal anchor points to minimize cross-chunk interference and enhance edit controllability. Experimental results demonstrate that AnyEdit++ significantly outperforms existing approaches across diverse tasks, including mathematical reasoning, code generation, and narrative editing, thereby validating the efficacy and robustness of structure-aware mechanisms for knowledge editing in long texts.

Abstract

Editing complex, long-form knowledge in Large Language Models remains a significant challenge due to the difficulty of maintaining generation coherence. Existing autoregressive methods like AnyEdit alleviate length constraints but rely on Fixed-window Chunking, which disregards logical structure and compromises consistency. To address this, we present AnyEdit++, a structure-aware framework incorporating Bayes-Chunk, an adaptive segmentation mechanism that dynamically identifies semantic boundaries based on Bayesian Surprise. We underpin this approach with a theoretical framework establishing two key principles: (1) Structural Independence: we prove that cross-segment interference is minimized when anchor keys are geometrically orthogonal (a condition naturally satisfied by our surprisal-based boundaries but violated by fixed windows), and (2) Causal Locality: we demonstrate that updates injected at these semantic peaks yield strictly superior control compared to arbitrary split points. Extensive experiments across mathematical reasoning, code generation, and narrative tasks demonstrate that AnyEdit++ achieves superior performance and robustness compared to state-of-the-art baselines, validating that structural awareness is critical for effective long-form knowledge editing.

Problem

• knowledge editing
• long-form generation
• coherence
• consistency
• Large Language Models

Innovation

• Bayesian Surprise
• adaptive segmentation
• structure-aware editing
• long-form knowledge editing
• causal locality