RoleWorld: Knowledge Transfer in World Modeling with Functional Role Abstractions

Boyuan Xiao; Pu Wang; Ji Feng; Yao-Xiang Ding

cs.AIOct 9, 2026

RoleWorld: Knowledge Transfer in World Modeling with Functional Role Abstractions

Authors: Boyuan Xiao, Pu Wang, Ji Feng, Yao-Xiang Ding

Organizations: State Key Lab of CAD&CG Zhejiang University, China · Baiont Quant, China

Abstract

World models enable large language model (LLM) agents to anticipate action effects and make informed decisions. For generalization, it is desirable for such world models to transfer efficiently to new environments with similar interaction structure and underlying dynamics. Recent approaches encode world dynamics as symbolic rules or executable programs, whose compositional structure in principle affords transfer, but still organize such knowledge around environment-specific objects. To address this limitation, we introduce RoleWorld, a symbolic world modeling approach which uses reusable prototypes and environment-specific parameter values as functional role abstractions to represent how objects behave in interactions. From interaction trajectories, a teacher LLM extracts role prototypes and executable transition rules defined over role instances, together with a skill that guides the actor LLM to use role information in decision making. In a target world, without the teacher, the actor uses this skill together with its current role assignments and parameter estimates, and actively tests a role instance hypothesis through interaction when its uncertainty affects the current decision. Feedback from each action refines role assignments and estimates parameter values online, with no LLM parameter updates and the transition rules held fixed. Experiments on MARS benchmark show that RoleWorld enables robust generalization across environments with significantly altered object behavior, suggesting that functional role abstractions can serve as an effective component for knowledge transfer in world modeling. Project Page: https://future-item.github.io/RoleWorld.

Figures & tables

Figure 1: Overview of RoleWorld . A teacher LLM induces role prototypes, transition rules, and an actor skill from source trajectories. These remain fixed in the target world, where the actor refines role estimates online and tests hypotheses when uncertainty could affect its decision.

Stage	ReAct	IFR	PoE-World	WALL-E	WorldCoder	RoleWorld
Source learning	$\times$	$\times$	$\times$	$\times$	$\checkmark$	$\checkmark$
Target learning	$\times$	$\checkmark$	$\checkmark$	$\checkmark$	$\times$	$\times$
Online refinement	$\times$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$

Table 1: Evaluation stages (

\checkmark

: yes;

\times

: no). Source learning: at most five episodes in Default; target learning: five per target world before evaluation; online refinement: within each evaluation episode.

Metrics	Mod. Type	Qwen3-8B						GPT-5.6 Luna
Metrics	Mod. Type	ReAct	IFR	PoE- World	WALL-E	World Coder	Ours	ReAct	IFR	PoE- World	WALL-E	World Coder	Ours
Reward	Terrain	$2.8\pm 1.6$	$2.0\pm 0.9$	$0.4\pm 1.2$	$2.0\pm 2.2$	$5.3\pm 1.7$	$\bm{9.4\pm 3.7}$	$4.0\pm 2.5$	$10.8\pm 1.2$	$0.8\pm 0.7$	$8.2\pm 3.1$	$5.8\pm 1.7$	$\bm{11.4\pm 1.1}$
	Survival	$2.0\pm 1.2$	$3.0\pm 2.0$	$0.0\pm 1.1$	$1.3\pm 0.8$	$2.3\pm 1.4$	$\bm{8.0\pm 2.6}$	$4.9\pm 2.5$	$9.3\pm 1.5$	$1.7\pm 2.4$	$10.2\pm 4.0$	$3.1\pm 1.7$	$\bm{10.8\pm 2.4}$
	Task Dep.	$1.9\pm 1.7$	$2.2\pm 0.7$	$1.5\pm 1.6$	$2.2\pm 1.4$	$0.8\pm 0.8$	$\bm{6.8\pm 0.8}$	$1.7\pm 1.1$	$7.7\pm 1.0$	$1.4\pm 0.8$	$4.8\pm 2.5$	$1.2\pm 0.6$	$\bm{8.5\pm 1.5}$
	Terr. Surv.	$1.4\pm 1.0$	$1.7\pm 0.8$	$0.1\pm 0.7$	$1.7\pm 3.2$	$4.4\pm 1.6$	$\bm{8.1\pm 3.3}$	$4.1\pm 1.5$	$7.8\pm 0.7$	$2.1\pm 0.9$	$5.7\pm 2.4$	$4.3\pm 1.4$	$\bm{11.0\pm 1.7}$
	Terr. Task.	$2.1\pm 1.2$	$3.0\pm 0.7$	$1.4\pm 1.1$	$1.7\pm 1.4$	$3.7\pm 1.5$	$\bm{6.2\pm 2.1}$	$3.8\pm 1.9$	$6.4\pm 0.8$	$3.8\pm 0.8$	$3.8\pm 1.3$	$4.2\pm 1.4$	$\bm{7.0\pm 1.0}$

Table 2: Reward and score (%) on the seven MARS variants under the protocols in Table 1 . Column groups indicate the backbone used for all target-world LLM calls.

Metrics	Method	Terrain	Survival	Task Dep.	Terr. Surv.	Terr. Task.	Surv. Task.	All Three	All
Reward	Direct LLM	$0.2\pm 1.0$	$2.0\pm 1.7$	$0.9\pm 1.4$	$1.3\pm 1.2$	$2.0\pm 1.1$	$0.8\pm 1.4$	$0.0\pm 1.1$	1.0
	w/o Skill	$0.9\pm 0.8$	$2.2\pm 1.6$	$1.3\pm 1.0$	$2.3\pm 1.5$	$1.1\pm 0.9$	$1.2\pm 0.7$	$-0.3\pm 0.5$	1.2
	w/o Adaptation	$9.7\pm 1.8$	$8.4\pm 3.1$	$0.9\pm 0.4$	$6.2\pm 1.9$	$4.3\pm 2.4$	$4.8\pm 2.6$	$0.2\pm 1.4$	4.9
	Fixed Role Assignments	$6.8\pm 3.1$	$6.5\pm 3.3$	$1.4\pm 1.3$	$7.8\pm 4.1$	$5.4\pm 1.4$	$3.7\pm 3.9$	$2.5\pm 1.7$	4.9
	w/o Active Verification	$8.0\pm 1.9$	$6.0\pm 4.3$	$1.2\pm 0.7$	$7.7\pm 3.1$	$1.7\pm 1.5$	$0.9\pm 0.8$	$-0.5\pm 0.7$	3.6
	RoleWorld (full)	$9.4\pm 3.7$	$8.0\pm 2.6$	$6.8\pm 0.8$	$8.1\pm 3.3$	$6.2\pm 2.1$	$4.3\pm 0.8$	$3.0\pm 1.5$	6.5

Table 3: Ablation results and decision-policy comparison on the seven MARS variants. LLM-based variants use Qwen3-8B. All is the arithmetic mean of the per-world means for each metric. Bold indicates the highest mean in each column within each metric.

Figure 2: Controlled validation of role factorization. Role assignments and role parameter values are corrupted separately, while the environment, role prototypes, and transition rules remain fixed. Shaded bands indicate one standard deviation; see Appendix A.5 for details.

Figure 3: Correction of role estimates with fixed role prototypes and transition rules: (a) correcting a role assignment; (b) correcting a role parameter.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Corruption type	Level (%)	Role parameter accuracy		Prediction discrepancy
		100	300	0	100	300
Role parameters	30	$87.27\pm 3.77$	$87.27\pm 4.23$	$34.19\pm 4.22$	$6.81\pm 2.09$	$6.00\pm 1.93$
	50	$79.09\pm 6.88$	$79.55\pm 7.01$	$45.15\pm 3.84$	$10.95\pm 2.83$	$9.68\pm 2.80$
	70	$70.91\pm 3.50$	$74.55\pm 4.78$	$48.79\pm 2.59$	$15.33\pm 3.84$	$12.30\pm 3.56$
	90	$58.79\pm 7.10$	$65.45\pm 5.75$	$51.29\pm 1.25$	$19.85\pm 3.53$	$14.98\pm 2.81$
Role assignments	30	$92.27\pm 3.93$	$96.52\pm 2.11$	$29.89\pm 4.26$	$4.69\pm 2.18$	$1.88\pm 1.92$

Appendix

Table 4: Controlled validation results (mean

\pm

sample standard deviation across six random seeds). Column subheadings give the active-interaction budget; both metrics are reported as percentages.

Prior	Deterministic mismatches	Missing predictions	Case mean TV	Assignment mean TV
Reference	0	0	0.0410	0.0421
Remove tree/R1	19	19	3.2889	3.3754
Require tool for tree	12	0	2.0923	2.1474
Change wood cost	7	0	1.2376	2.1633
Remove workstation requirement	4	0	0.7248	1.0677
Allow planting on sand	4	0	0.7248	0.2802

Appendix

Table 5: Environment-transition validation of the frozen reference prior and independent single-error controls. Deterministic mismatches are counted over 566 cases and include missing predictions. Both TV averages cover all 585 cases and are reported as percentages.

World Models

World models represent and predict how environments change, helping agents plan, act, and reason about possible futures.

See all 366 papers

Qwen-AgentWorld: Language World Models for General Agents
Jun 23, 2026Yuxin Zuo, Zikai Xiao, Li Sheng +30LLM World Models Agentic RL
Qwen Team
How do World Models and Policies Compose in LLM Agents? A Joint Spectral and Behavioral Account
Aug 30, 2026Ruize Xu, Xiao Yu, Yujin Tang +2World Model Learning World Models
Dartmouth College · Columbia University
From Traces to Agentic Worlds: Agentic Language World Models for Interactive Environment Simulation
Oct 5, 2026Quanyu Long, Xiao Chen, Jianda Chen +4LLM World Models Interactive World Models
Nanyang Technological University · The Hong Kong Polytechnic University

Jun 23, 2026J/K move · Enter open · S save

Qwen-AgentWorld: Language World Models for General Agents

Yuxin Zuo, Zikai Xiao, Li Sheng, Fei Huang+29

Qwen Team

LLM World Models Agentic RL World Model Learning World Model Evaluation

Open paper AskPDF arXiv

A world model predicts environment dynamics based on current observations and actions, serving as a core cognitive mechanism for reasoning and planning. In this work, we investigate how world modeling based on language models can further push the boundaries of general agents. (i) We first focus on building foundation models for agentic environment simulation. We introduce Qwen-AgentWorld-35B-A3B and Qwen-AgentWorld-397B-A17B, the first language world models capable of simulating agentic environments covering 7 domains via long chain-of-thought reasoning. Leveraging more than 10M environment interaction trajectories of 7 domains in real-world environments, we develop Qwen-AgentWorld through a three-stage training pipeline: CPT injects general-purpose world modeling capabilities from the state transition dynamics and augmented professional corpora, SFT activates next-state-prediction reasoning, and RL sharpens simulation fidelity through a tailored framework with hybrid rubric-and-rule rewards. To evaluate language world models, we present AgentWorldBench, a comprehensive benchmark constructed from real-world interactions of 5 frontier models on 9 established benchmarks. Empirical results demonstrate that Qwen-AgentWorld significantly outperforms existing frontier models. (ii) Beyond foundation models, we further investigate two complementary paradigms through which world modeling enhances general agents. First, as a decoupled environment simulator, Qwen-AgentWorld supports scalable and controllable simulation of thousands of real-world environments for agentic RL, yielding gains that surpass real-environment training alone. Second, as a unified agent foundation model, world-model training acts as a highly effective warm-up that improves downstream performance across 7 agentic benchmarks. Code: https://github.com/QwenLM/Qwen-AgentWorld

Abstract

World Models

Explore similar work

Qwen-AgentWorld: Language World Models for General Agents

How do World Models and Policies Compose in LLM Agents? A Joint Spectral and Behavioral Account

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