From Traces to Agentic Worlds: Agentic Language World Models for Interactive Environment Simulation

🤗 Upvotes: 71 | cs.CL

Authors:
Quanyu Long, Xiao Chen, Jianda Chen, Haozhen Zhang, Qisheng Hu, Jianzhu Bao, Wenya Wang

Title:
From Traces to Agentic Worlds: Agentic Language World Models for Interactive Environment Simulation

Arxiv:
http://arxiv.org/abs/2610.06100v1

Abstract:
Realistic environment replicas are increasingly valuable for training and evaluating LLM agents, yet the original systems may be inaccessible or impractical to reproduce. We explore agentic language world modeling: rather than rebuilding an executable environment, a world model agent serves as the environment for a task agent and supports faithful and stateful simulation. We instantiate this paradigm with Trace2Env, a learning-free framework for settings where the original system is unavailable but historical interaction traces remain accessible. Trace2Env reconstructs these traces into a reusable environment worldbook containing environment schemas, grounded evidence, and induced behavioral knowledge. At runtime, the world model agent actively consults the worldbook together with persistent episodic state to infer each action's observation and lasting state effects. Across nine environments, Trace2Env improves both next-observation fidelity and long-horizon interaction consistency over conventional prompt-based LWMs. In multi-turn interaction, task agent actions generated against Trace2Env remain valid more often when replayed in the real environment, indicating that its simulated dynamics better preserve the consequences of earlier actions across successive turns. These results establish agentic language world modeling as an alternative direction for building realistic environment replicas without reconstructing the original executable system.

What is Daily Paper Cast?

We update every weekday to discuss highest-voted papers from Huggingface Daily Paper (https://huggingface.co/papers). Both the podcast scripts and audio are generated by AI. Feedback and suggestions are welcome! Email us: [e-mail]

Creator:
Jingwen Liang, 3D ML, https://www.linkedin.com/in/jingwen-liang/
Gengyu Wang, LLM ML, http://wanggengyu.com

Listen on:
Spotify: https://open.spotify.com/show/21nrhmdaA8qoBiH8q03NXL
Apple Podcast: https://podcasts.apple.com/us/podcast/daily-paper-cast/id1777620236

Cover Image by Kawen Kuang https://kawen.art

Evan: Welcome to Daily Paper Cast.

Ashley: Today's paper is from Hugging Face's daily paper list of October 9, 2026, and it has received 71 upvotes.

Evan: The title is: From Traces to Agentic Worlds: Agentic Language World Models for Interactive Environment Simulation.

Ashley: The first two authors are Quanyu Long and Xiao Chen, with correspondence to [e-mail] from Nanyang Technological University.

Evan: Let's dive into the Introduction section.

Interactive agents rely heavily on environments that accurately respond to their actions and maintain these effects over time.

Ashley: That's right, Evan.

Realistic replicas of such environments are extremely valuable for training, evaluating, and testing AI agents.

However, reproducing these settings can be challenging because original software systems like terminals, workspaces, or enterprise systems often require inaccessible code, data, or infrastructure.

Evan: Recent interactions might offer a solution though, right?

Historical interaction traces typically remain available and they record actions, observations, and consequences — even when the original environment cannot be reproduced.

Ashley: Exactly.

Previous work has demonstrated that trajectories can be distilled into reusable skills and structured knowledge.

Examples include methods for deducing recurring procedures from past executions or consolidating trajectory-local lessons into transferable skills.

Evan: But rebuilding a complex environment faithfully often requires unavailable implementation details and dependencies.

How does this paper handle that challenge?

Ashley: Rather than reconstructing the original system, this paper proposes recovering a behavioral specification to simulate the environment's responses to new actions.

They call this 'agentic language world modeling.'

Evan: So, we're talking about using a language model to simulate what the environment would do in response to various agent actions — but with a twist to improve continuity, right?

Ashley: Precisely.

Language world models, or LWMs, traditionally cast simulation as next-observation prediction.

Given a textual description of the environment, the interaction history, and a task agent’s action, the LWM predicts what the real environment would return.

However, achieving fidelity in simulation isn't just about next-step prediction.

Evan: Indeed, maintaining long-horizon state continuity is crucial.

Actions can change the environment in subtle ways that become relevant only many turns later.

For example, deleting a file might result in an error when trying to access it at a much later stage.

Ashley: Exactly.

This paper addresses these challenges by proposing an agentic language world model that treats environment simulation itself as an agentic task.

Instead of relying on a fixed prompt, a dedicated world model agent consults externalized environment knowledge and active reasoning.

Evan: Can the paper detail how one agent can serve as the environment for another?

And whether traces alone provide sufficient detail?

Ashley: The task agent decides what action to take, while the world model agent determines how the simulated environment should respond.

This agent revises trace-derived environment knowledge relevant to the current interaction, combines it with the current episode state and memory, and reasons over these resources before producing the environment’s response.

Evan: But how does it handle settings where the original system isn't directly accessible?

Ashley: For these settings, the paper introduces a learning-free framework called Trace2Env.

It reconstructs historical interaction traces into a structured worldbook, retaining concrete behavioral evidence and inducing reusable environment knowledge into schemas, grounded evidence, and behavioral abstractions.

Evan: During simulation, it seems Trace2Env plays a critical role.

How exactly does it help maintain state effects and observation fidelity?

Ashley: You're spot on.

The world model agent actively consults the worldbook along with episode state and episodic memory to determine the consequences of task agent actions.

This separation between the worldbook and the runtime environment allows maintaining long-horizon interactions consistently.

Evan: How do the experiments validate this model and its effectiveness compared to conventional approaches?

Ashley: Great question.

The paper evaluates Trace2Env across nine environments on two fronts.

Firstly, single-step next-observation fidelity measures how closely the simulated environment response matches the real one.

Secondly, multi-turn interaction consistency tests whether task agent actions in the simulated environment remain valid when replayed in the real environment.

Evan: And the results indicate better preservation of the consequences through Trace2Env compared to conventional models, improving interaction fidelity over successive turns.

Ashley: Right.

Trace2Env not only improves the next-observation fidelity but also maintains interaction consistency across turns, demonstrating that its simulated dynamics are much closer to the real environment's response.

Thus, it sets a new direction for building realistic environment replicas without reconstructing the original executable system.

Evan: That's the end of the Introduction section.

Join us next time as we delve deeper into the methodology of Trace2Env.

Evan: Thanks for staying with us.

Now, let's delve into the methodological details of Trace2Env.

Ashley: Alright, so as we discussed, the cornerstone of this approach is the concept of the agentic language world model.

At a high level, it involves reconstructing an environment worldbook from recorded traces and using a dedicated world model agent to operate it as a simulated environment.

Evan: Can you break down the process of how this worldbook is reconstructed, Ashley?

Ashley: Certainly, Evan.

The reconstruction process happens in two phases.

The first phase involves offline construction, where Trace2Env uses historical interaction traces to create a structured environment worldbook.

This worldbook comprises schemas, grounded evidence, induced abstractions, and provenance links.

Evan: And what exactly do these components entail?

Ashley: Schemas describe the action interface and the state variables that the simulator can maintain.

Grounded evidence includes recorded transitions and selected demonstrations along with their original observations.

Induced abstractions capture recurring rules, state constraints, conditional effects, observation contracts, and descriptive conventions.

Lastly, provenance records link these artifacts back to the traces and turns supporting them.

Evan: How does Trace2Env handle the separation of reusable environment knowledge from episode-specific facts?

Ashley: Great question.

Trace2Env maintains a persistent episode workspace consisting of the fixed worldbook, mutable episode-specific state, and episodic interaction memory.

The worldbook captures general knowledge about environment behavior, while represented episode state holds facts specific to the current simulation.

Episodic memory contains the details of what happened earlier in the interaction sequence.

Evan: So it essentially ensures that while the environment knowledge is transferable, what happens in each episode remains episode-specific, right?

Ashley: Exactly.

For example, the task agent might create or modify files during simulation that have never appeared in the construction traces.

Trace2Env treats these signals separately to avoid making unsupported assumptions.

Evan: How does the simulation step work in practice within Trace2Env?

Ashley: At each simulation step, Trace2Env generates the next transition through two stages: proposal and validation.

First, the world model agent proposes effects and observations based on the current episode and worldbook knowledge.

It alternates between inspecting the workspace and reasoning about the transition until ready to propose what should happen next.

Evan: And how is this proposal validated before being committed?

Ashley: The proposal undergoes validation through a gate that checks candidate state effects against declared schemas, state constraints, and supporting rules.

Only accepted proposals are committed, ensuring state continuity.

This commitment process involves resolving the proposed observation under applicable rendering constraints.

Evan: That sounds quite thorough.

How about the evaluation?

How did the authors assess the effectiveness of Trace2Env?

Ashley: Trace2Env was evaluated across nine distinct environments using two primary metrics.

The first metric, single-step next-observation fidelity, examined whether the simulated environment response matched the real environment response in terms of format, factuality, consistency, realism, and overall quality.

Evan: And the second metric?

Ashley: The second metric focused on multi-turn interaction consistency by comparing the validity of task agent behaviors in the simulated environment when replayed in the real environment.

This involves measuring the success rate of a task agent in the simulated environment and re-evaluating the same agent actions in the real environment.

Evan: How did Trace2Env perform on these metrics?

Ashley: Trace2Env improved next-observation fidelity across all five dimensions compared to conventional language world models.

It also demonstrated higher multi-turn interaction consistency.

This means that task agent actions generated in the Trace2Env simulation were more likely to remain valid when replayed in the real environment, reflecting better simulation fidelity and long-term state continuation.

Evan: What do the authors suggest about the implications of these results for future work?

Ashley: The authors propose that agentic language world modeling offers a promising direction for building realistic environment replicas without reconstructing the original executable systems.

Future work could enhance active inspection and inference over the worldbook, expand reconstruction capabilities beyond observed behavior, and explore more ways agentic simulated environments can support task agent training and evaluation.

Evan: It's certainly an intriguing approach with significant potential.

That covers the Method section of the paper.

Evan: Let's move on to the Experiment and Results section to understand how Trace2Env was evaluated.

Ashley: The paper evaluates Trace2Env across nine environments in two different setups.

The first setup focuses on single-step next-observation prediction, using environments such as Terminal, SWE, Android, and Web from AgentWorldBench, along with three others from EnvScaler: Food, Shopping, and Benefits.

Evan: And what about the second setup?

Ashley: The second setup evaluates multi-turn interaction consistency using two text-game environments: ALFWorld and SciWorld.

Both setups aim to measure how faithful the simulated environment's response is to the real environment, and whether interactions remain valid across successive turns.

Evan: What metrics did they use to assess single-step prediction?

Ashley: For single-step next-observation prediction, they used the official AgentWorldBench five-dimensional judge score.

This score encompasses Format, Factuality, Consistency, Realism, and Quality, each rescaled to 0-100 and averaged.

Evan: Can you explain how they conducted these evaluations?

Ashley: Sure.

Each record in AgentWorldBench consists of a real trajectory prefix and the current action, with the world model predicting the next observation.

They scored each prediction on the five dimensions we mentioned.

Similarly, for EnvScaler, they applied the same five-dimensional evaluation.

Evan: How did the different configurations of Trace2Env and conventional prompting methods fare in these tests?

Ashley: Trace2Env achieved the highest average score with both backbones, scoring 75.94 with GPT and 75.75 with DeepSeek, improving substantially over Direct Prompting by 6.43 and 7.04 points, respectively.

It was best in 11 out of 14 environment-backbone pairs.

Evan: What about the contributions from the worldbook and runtime itself?

Ashley: The worldbook and runtime provide complementary gains.

Trace RAG Prompting already improves over Direct Prompting, but Worldbook Prompting raises the average further.

Combining both components through Trace2Env yields the highest fidelity.

Evan: What were their findings on multi-turn interaction consistency?

Ashley: In multi-turn evaluations, they looked at task success rates in the real environment compared to the simulated world model.

For example, in ALFWorld, Direct Prompting had a 97% success rate in simulation, but when actions were replayed in the real environment, the success rate dropped to 3%.

In contrast, Trace2Env maintained a success rate of 85%, showing much higher long-term consistency.

Evan: That's quite impressive in maintaining real-environment consistency.

Did they perform any ablations to understand the impact of different components of Trace2Env?

Ashley: Yes, they did.

For example, on the Terminal environment, they tested the impact of adding various components like abstraction and evidence to the schema-only configuration.

The schema benefited more when paired with evidence, improving over just schema alone.

Abstractions added a smaller gain, but together with evidence, they achieved the highest score.

Evan: Interesting.

How about efficiency?

Was there a trade-off between accuracy and computational cost?

Ashley: Indeed, there was a trade-off.

Trace2Env’s comprehensive validation and state tracking incur higher computational costs.

For instance, the full system's inference cost on Terminal was $0.187 per record with a latency of 38.1 seconds, which is more than Direct Prompting.

However, removing components such as the adaptive loop reduced costs significantly while retaining a large portion of the gain in fidelity.

Evan: So, what about the implications of these results for future developments?

Ashley: The authors suggest that these results point toward the potential of agentic language world modeling as a robust method for creating realistic environment replicas, especially in the absence of original systems.

Future work could focus on enhancing active inspection and reasoning capabilities and expanding reconstruction techniques beyond observed behavior.

Evan: That sums up the Experiment section nicely.

It's clear that Trace2Env sets a new benchmark for simulating interactive environments.

Evan: Let's now turn our attention to the Related Work section to understand how Trace2Env fits into the broader research landscape.

Ashley: Trace2Env builds on a rich history of research in language world models and trace-grounded reconstruction techniques.

Language world models, or LWMs, have evolved to predict the consequences of agent actions.

Recent studies have focused on improving their fidelity and utility across diverse environments.

Evan: Can you give us a brief overview of some of these recent advancements in LWMs?

Ashley: Of course.

LWMs have been explored for their predictive capabilities in various interactive environments.

Studies by Li et al., and Zuo et al., in 2026, have scaled LWMs across different agentic environments, while research by Chae et al., and Liu et al., in 2025 and 2026, utilized simulated transitions to enhance decision-making and policy learning.

Evan: Interesting.

What are the key challenges these studies aim to address?

Ashley: The main challenges revolve around maintaining state continuity over longer interactions and ensuring that simulated dynamics preserve behaviorally or decision-relevant information.

For example, Cai et al., and Huang et al., in 2026, emphasized objectives beyond literal next-observation accuracy, such as preserving dynamically relevant information.

Evan: But relying entirely on parametric knowledge has its limitations.

Are there ways to augment LWMs with external experience?

Ashley: Approaches like Retrieval-Augmented World Models, or RAWM, retrieve historical transitions as in-context evidence for prediction.

A notable example is WorldEvolver by Zhang et al., in 2026, which augments a frozen world model with episodic memories and semantic rules to improve predictions.

Evan: How does Trace2Env compare to these methods?

Ashley: Trace2Env differentiates itself by focusing on a trace-only setting where the original environment is unavailable.

Instead of relying on continuous updates from the original environment, Trace2Env organizes historical traces into a detailed environment worldbook, which is actively inspected by a world model agent during simulation.

Evan: That makes sense.

But LWMs aren't the only related field—trace-grounded reconstruction from agent trajectories is crucial, right?

Ashley: Definitely.

A complementary line of research treats agent trajectories as evidence for distilling reusable structure.

For instance, Agent Workflow Memory by Wang et al., in 2025, developed recurring procedures from past executions.

Similarly, Trace2Skill by Ni et al., in 2026, crystallizes trajectory-local lessons into transferable skills directories.

Evan: Can you explain how specific works relate to Trace2Env in this context?

Ashley: Trace2Env shares similarities with these methods in how it leverages past agent interactions but stands out in its detailed environmental reconstruction approach.

For example, Terminal-Universe by Wu et al., in 2026, reconstructs terminal workspaces from agent trajectories but is evaluated through real tool feedback rather than through language model predictions.

Evan: And how does Trace2Env utilize this reconstructed knowledge within its framework?

Ashley: Trace2Env uses a non-executable environment worldbook containing schemas, grounded evidence, and induced behavioral knowledge.

A world model agent inspects this knowledge to infer state changes and simulate subsequent observations actively.

This approach ensures sustained interaction consistency, addressing the challenge of long-term state fidelity.

Evan: So, the depth and fidelity of the worldbook are key to Trace2Env's success?

Ashley: Exactly.

By externalizing environment knowledge, Trace2Env can provide consistent and realistic simulations without needing the original environment.

It's a strategic blend of capturing historical interactions and applying modern world model techniques.

Evan: That's a comprehensive look at how Trace2Env fits into and extends existing research.

That wraps up our exploration of the Related Work section.

Evan: As we wrap up, let's summarize the key contributions and takeaways from this paper.

Ashley: Sure, Evan.

The paper introduces Trace2Env, an innovative framework for simulating realistic environmental replicas through agentic language world modeling.

The authors propose leveraging historical interaction traces, even when the original systems are inaccessible, to reconstruct a detailed environment worldbook.

Evan: Right.

This worldbook includes schemas, grounded evidence, induced abstractions, and provenance, effectively creating a resource that a world model agent consults actively during simulations.

Ashley: These components ensure that the simulation maintains state continuity and interaction fidelity over long-term tasks.

By separating reusable environment knowledge from episode-specific facts, Trace2Env avoids making unsupported assumptions and keeps the simulation faithful to real-world dynamics.

Evan: The paper’s evaluations demonstrate significant improvements over conventional language world models.

Trace2Env not only enhances next-observation fidelity but also better preserves task agent behaviors in simulated environments when replayed in real environments.

Ashley: Moreover, the authors highlight the potential of agentic language world modeling to support future task agent training and evaluation, suggesting enhancements to active inspection and inference capabilities, as well as expanding reconstruction techniques beyond observed behavior.

Evan: This exploration sets a promising direction for creating realistic environment replicas without needing the original executable systems.

Well, that covers our discussion for today.

Ashley: We hope you found this episode insightful.

Remember to join us next time as we dive into another fascinating paper from Hugging Face’s daily list.

Evan: Thanks for listening to Daily Paper Cast.

Until next time, stay curious and keep exploring!