RUC-AIBOX/AEWM

Hugging Face
TEXT GENERATIONPricing:Input $0.4 / Cached $0.07 / Output $4Concurrent Unit Cost:2Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 23, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Warm

RUC-AIBOX/AEWM is a 35.1 billion parameter world model developed by RUC-AIBOX, built upon Qwen3.5-35B-A3B, designed to enhance agent reasoning and tool use. It integrates an Action Judge (AJ) to classify agent decisions and State Revision (SR) to edit noisy actions, operating within the EditAct inference framework. This model excels in long-horizon tasks across Search, Terminal, and Software Engineering domains by actively judging and revising agent decisions rather than merely predicting tool responses.

Loading preview...

AEWM: Agent-Editing World Model

AEWM (Agent-Editing World Model) is a 35.1 billion parameter model developed by RUC-AIBOX, based on Qwen3.5-35B-A3B. Unlike traditional world models that predict tool responses, AEWM focuses on judging and editing an agent's decisions to improve long-horizon reasoning and tool use. It addresses "task-state contamination" where agents make locally reasonable but globally incorrect decisions.

Key Capabilities

  • Action Judge (AJ): Classifies proposed agent decisions as Critical (advances task), Exploratory (gathers info), or Noisy (redundant/irrelevant) before execution.
  • State Revision (SR): For noisy decisions, SR generates revised reasoning and actions, directly editing the agent's continuation.
  • EditAct Framework: Integrates AJ and SR into a four-step inference process (Propose, Judge, Edit, Execute) to guide agent interaction with real environments.
  • Cross-Domain Expertise: Optimized for Search, Terminal, and Software Engineering (SWE) tasks.
  • Performance Improvement: Achieves 70.5% overall macro-F1 on the AEWM Action Judge Benchmark, outperforming baselines. It also demonstrates significant end-to-end performance gains for agents using EditAct, with up to +6.7 absolute percentage points over strong baselines.

Good for

  • Enhancing Agent Reliability: Ideal for developers building agents that require robust, long-horizon reasoning and precise tool use in complex environments.
  • Reducing Task-State Contamination: Useful for scenarios where agents might otherwise make incorrect assumptions or persist with outdated plans.
  • Software Engineering Tasks: Particularly strong in SWE, Terminal, and Search domains, making it suitable for automated development, command-line operations, and information retrieval agents.
  • Integrating with Existing Agents: Designed to work as a world model alongside a separate agent, providing online judgment and revision capabilities.

Finetunes
1 models