Renmin University Open-Sources Agent-Editing World Model (AEWM): State-Editing Paradigm for Autonomous LLM Agents

Core Background & Industry Pain Points While world models promise predictive foresight for LLM agents, existing designs attempt to predict raw environment observations—such as reconstructing noisy shell stdout or bloated HTML payloads. Reconstructing high-entropy observations provides minimal utility when real execution feedback is readily available. Concurrently, agents suffer from severe task-state contamination: outdated hypotheses and erroneous plan revisions remain trapped in context history, persistently distorting subsequent inference steps. ### Architecture Highlights & Internals RUC AI Box introduces the Agent-Editing World Model (AEWM), decoupling world modeling from observation rendering: (1) An Action Judge parses trajectories into Critical, Exploratory, and Noisy actions; (2) A State Revision module formulates surgical edits to excise misleading reasoning paths; and (3) The EditAct runtime actively rewrites the underlying conversation context upon observing execution discrepancies, directly rehabilitating the state vector prior to subsequent tool invocations. ### Authoritative Benchmarks & Measured Scores Evaluations across Search, Terminal, and Software Engineering environments demonstrate compelling robustness: AEWM scores 70.5% Macro-F1 on action verification (+10.6 points over frontier baselines). Deployed across three distinct LLM backbones, EditAct delivers consistent 3.2-6.7 point gains across six downstream benchmarks, while offline Rejection Sampling Fine-Tuning (AEWM-RFT) yields persistent out-of-domain improvements. ### Developer Hands-on Guide Engineers can clone https://github.com/RUCAIBox/Agent-Editing-World-Model to integrate EditAct into LangChain, AutoGen, or proprietary agent runtimes, accessing checkpoints directly from Hugging Face.