Paper review · October 10, 2026
From Traces to Agent Worlds: Agentic Language World Models That Simulate Interactive Environments
From Traces to Agentic Worlds: Agentic Language World Models for Interactive Environment Simulation
By restructuring past interaction traces into an environment 'worldbook' and simulating the environment with a world model agent that consults it, the method produced more faithful and consistent simulations than existing prompt-based approaches.
Problem
Replicas of real environments are useful for training and evaluating LLM agents, but the original system's code, data and infrastructure are often unavailable. Existing language world models put the environment description and traces into a single prompt and predict the next observation, which makes it hard to maintain state changes that only surface later, such as file deletion, across multiple turns.
Method
The paper proposes Trace2Env, which requires no training. In an offline stage, traces are turned into a worldbook made up of a schema (definitions of actions and states), grounding traces, derived action rules and provenance information. At execution time, a world model agent consults the worldbook, the current episode state and episode memory as needed, and proposes observations and state changes. These pass through a gate that judges whether the retrieved traces apply to the current situation and a gate that validates proposed state changes against the schema and rules, after which a shared harness commits them.
Results
- Across 9 environments, the authors report improvements over existing prompt-based language world models on all five criteria: format, factuality, consistency, realism and overall quality.
- Next-observation prediction was evaluated on Terminal, SWE, Android and Web from AgentWorldBench and Food, Shopping and Benefits from EnvScaler; multi-turn evaluation used ALFWorld and SciWorld.
- In multi-turn interaction, the authors report that agent actions generated inside Trace2Env were more often valid when replayed in the real environment.
Limitations
The provided text was truncated at the experimental setup, so specific numbers and comparison baselines could not be confirmed. The worldbook is not an executable reconstruction, so behavior not present in the traces relies on inference. Multi-turn evaluation is limited to two text games.
Why it matters
If environments for agent training and testing can be built from interaction traces alone, without the original system, agents could be practiced safely even in environments that are hard to replicate, such as internal company systems.
Key terms
- Language World Model (LWM)
- A model that imitates an environment by using a language model to predict 'what observation the environment will return if this action is taken.'
- Worldbook
- A bundle of environment knowledge built from past traces. It contains the schema, grounding from real observations, recurring rules and provenance information, and it does not change during execution.
- Episode State
- A mutable state that records what has been established as fact in the current episode (e.g., a certain file was deleted). It is kept separate from the general knowledge in the worldbook.
Original
Authors: Quanyu Long, Xiao Chen, Jianda Chen, Haozhen Zhang, Qisheng Hu, Jianzhu Bao, Wenya Wang · Posted 2026-10-05 · License arXiv 비독점 배포 라이선스 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.