Agent-Editing World Model: Rethinking World Modeling for LLM Agents

Hugging Face Daily Papers Papers
world-modeling llm-agents action-judgment state-revision fine-tuning

Summary

This paper proposes the Agent-Editing World Model (AEWM) to enhance LLM agents by modeling how reasoning and actions shape future task progress, achieving significant performance improvements on benchmarks across domains like search and software engineering.

Recent advances in large language models (LLMs) have enabled agents to tackle long-horizon tasks across diverse environments. To further improve agent performance, existing language world models typically predict environment observations, yet reconstructing high-entropy, execution-dependent tool responses offers limited value when real feedback is available. Meanwhile, agents suffer from task-state contamination, where unsupported assumptions and outdated plans persist in history and distort subsequent decisions. We propose the Agent-Editing World Model (AEWM), which models how reasoning and actions shape future task progress rather than simulating tool responses. AEWM combines Action Judge to distinguish Critical, Exploratory, and Noisy decisions with State Revision to edit noisy reasoning--action continuations from the same observed history. EditAct integrates these capabilities with real execution, directly changing the state underlying subsequent decisions rather than merely providing critiques. We train AEWM across Search, Terminal, and Software Engineering through mid-training and supervised fine-tuning. AEWM achieves 70.5\% macro-F1 on our Action Judge benchmark, exceeding the strongest frontier baseline by 10.6 points. Across six benchmarks and three agent backbones, EditAct improves average scores by 3.2--6.7 points over the strongest baseline. Furthermore, rejection sampling fine-tuning on verified EditAct trajectories, termed AEWM-RFT, improves over Self-RFT by 2.2--2.6 points across three domains without online AEWM guidance.
Original Article
View Cached Full Text

Cached at: 09/25/26, 03:45 AM

Paper page - Agent-Editing World Model: Rethinking World Modeling for LLM Agents

Source: https://huggingface.co/papers/2609.28416

Abstract

Recentadvancesinlargelanguagemodels(LLMs)haveenabledagentstotacklelong-horizontasksacrossdiverseenvironments.Tofurtherimproveagentperformance,existinglanguageworldmodelstypicallypredictenvironmentobservations,yetreconstructinghigh-entropy,execution-dependenttoolresponsesofferslimitedvaluewhenrealfeedbackisavailable.Meanwhile,agentssufferfromtask-statecontamination,whereunsupportedassumptionsandoutdatedplanspersistinhistoryanddistortsubsequentdecisions.WeproposetheAgent-EditingWorldModel(AEWM),whichmodelshowreasoningandactionsshapefuturetaskprogressratherthansimulatingtoolresponses.AEWMcombinesActionJudgetodistinguishCritical,Exploratory,andNoisydecisionswithStateRevisiontoeditnoisyreasoning--actioncontinuationsfromthesameobservedhistory.EditActintegratesthesecapabilitieswithrealexecution,directlychangingthestateunderlyingsubsequentdecisionsratherthanmerelyprovidingcritiques.WetrainAEWMacrossSearch,Terminal,andSoftwareEngineeringthroughmid-trainingandsupervisedfine-tuning.AEWMachieves70.5\%macro-F1onourActionJudgebenchmark,exceedingthestrongestfrontierbaselineby10.6points.Acrosssixbenchmarksandthreeagentbackbones,EditActimprovesaveragescoresby3.2--6.7pointsoverthestrongestbaseline.Furthermore,rejectionsamplingfine-tuningonverifiedEditActtrajectories,termedAEWM-RFT,improvesoverSelf-RFTby2.2--2.6pointsacrossthreedomainswithoutonlineAEWMguidance.

View arXiv pageView PDFProject pageGitHub3Add to collection

Get this paper in your agent:

hf papers read 2609\.28416

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2609.28416 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2609.28416 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2609.28416 in a Space README.md to link it from this page.

Collections including this paper1

Similar Articles

Towards a Belief-Based World Model for LLM Agents

arXiv cs.AI

This paper introduces Belief-Based World Models (BB-WMs) to enhance LLM agents' decision-making under partial observability by providing direct access to beliefs about uncertain states, showing improved task performance.

Qwen-AgentWorld: Language World Models for General Agents

Hacker News Top

Qwen-AgentWorld introduces language world models for agentic environments, covering seven domains with long chain-of-thought reasoning. The work includes a new benchmark, AgentWorldBench, and shows that world modeling improves downstream agent performance.