When Do Causal World Models Help Modular LLM Agents

Xinyuan Song; Zekun Cai

cs.AIJul 12, 2026

When Do Causal World Models Help Modular LLM Agents

Authors: Xinyuan Song, Zekun Cai

Organizations: Emory University, Atlanta, GA, USA · The University of Tokyo, Tokyo, Japan · LocationMind, Tokyo, Japan

Abstract

LLM agents increasingly operate in modular systems, where planning requires coordinating multiple functional components whose actions can causally change downstream states and behaviors. This makes intervention-aware prediction essential: an agent must anticipate not only what is likely to happen next, but what will happen as a consequence of its own actions. We find that purely observational world models can incur irreducible intervention-time error in such systems, even when they accurately model observational trajectories. To address this limitation, we propose CausalCompose, a causal world-model framework that equips modular LLM agents with causal structure for intervention-aware prediction and planning. We theoretically show that causal composition can overcome this intervention gap with sufficient evidence and controlled local error. Experiments across diverse modular agent environments validate these results: incorporating causal structure consistently improves intervention-sensitive reasoning and planning, with particularly clear gains in structured tool-use systems where causal dependencies directly govern action execution. Our results show when causal world models help modular LLM agents: when causal dependencies are identifiable and actionable for planning.

Figures & tables

Figure 1. Overview of CausalCompose . (a) Local evidence: each module preserves local records and mechanisms while exposing compact intervention-response summaries. (b) Interface validation: aggregated evidence validates causal dependencies, including direction and lag, to construct the causal interface graph. (c) Composition and control: the validated graph is composed with local mechanisms for downstream prediction and action guidance, with execution feedback supporting subsequent updates.

Work	Causal WM	Modular	Intervention evidence	Cross-module composition	Planning use	Theory
Gkountouras et al. ( Gkountouras et al., 2025 )	\ding 51	\ding 55	\ding 51	\ding 55	\ding 51	\ding 55
CausalPlan ( Nguyen et al., 2025 )	\ding 51	\ding 55	\ding 51	\ding 55	\ding 51	\ding 55
Dyna-Think ( Yu et al., 2025 )	\ding 55	\ding 51	\ding 55	\ding 51	\ding 51	\ding 55
PatchWorld ( Bai et al., 2026 )	\ding 55	\ding 51	\ding 55	\ding 51	\ding 51	\ding 55
WM3C ( Wang and Huang, 2025 )	\ding 51	\ding 51	\ding 51	\ding 51	\ding 51	\ding 55
PoE-World ( Piriyakulkij et al., 2025 )	\ding 55	\ding 51	\ding 55	\ding 51	\ding 51	\ding 55

Table 1. Comparison with closely related work. We compare methods by whether they support causal world modeling, modular structure, intervention-based evidence, cross-module causal composition, planning-time use, and theoretical characterization.

Figure 2. Causal composition and local control in CausalCompose . (a) Validated interfaces connect local mechanisms to propagate intervention effects and predict downstream states. (b) The composed model can optionally guide local actions through prerequisite checks, effect prediction, action gating, and post-action verification.

Method	Success (%)
No World Model	41.79
Global Sequence WM	37.31
Transition Graph WM	33.58
Correlation WM	35.07
Local WM	38.06
Temporal WM	38.81

Table 2. ALFWorld task success (%) over the full evaluation of 134 tasks.

Method	Retail	Airline
No World Model	38.10	18.19
Global Sequence WM	33.33	33.33
Transition Graph WM	42.86	23.19
Correlation WM	43.94	21.86
Local WM	33.33	18.19
Temporal WM	42.76	23.49

Table 3.

\tau

-bench task success (%) on the Retail and Airline domains.

Setting	Baseline	Treatment	Absolute $\Delta$	Relative $\Delta$
$\tau$ -bench Anchor ( Oracle-Causal )	45.00	75.00	+30.00	+66.67%
$\tau$ -bench Anchor ( Discovered-Causal )	45.00	60.00	+15.00	+33.33%
APIBank Seq-WM $\rightarrow$ Discovered-Causal	3.61	4.64	+1.03	+28.53%
APIBank Seq-WM $\rightarrow$ Oracle-Causal	3.61	5.67	+2.06	+57.06%
ScienceWorld	14.80	17.67	+2.87	+19.39%
$\tau$ -bench Model Effect	45.23	50.00	+4.77	+10.55%

Table 4. Controlled diagnostics of causal-information usability. Absolute

\Delta

is the Treatment–Baseline difference, and Relative

\Delta

normalizes it by the Baseline. Values are percentages except for ScienceWorld.

Diagnostic	Start	End	Change	$\rho$	Control
Observational MSE under confounding	0.043	16.309	$379.3\times$	–	$\gamma$
Causal MSE under confounding	0.040	0.040	$1.0\times$	–	$\gamma$
Observational–causal MSE gap over horizon	1.259	4.764	$+3.505$	–	$H$
Causal-edge recovery F1	0.000	0.889	$+0.889$	$+1.000$	$N$
Causal advantage with stronger proxy	0.254	0.132	$-48.0\%$	$-1.000$	$c$

Table 5. Synthetic SCM diagnostics for theoretical predictions. Start and End denote the two endpoints of each sweep; Change summarizes the difference, and

\rho

is the rank correlation over the full sweep.

Environment	Condition	Performance	Calls	Cost/Run	$\Delta$ Perf.	$\Delta$ Calls
Retail	Sequence	33.33	15.8	$0.0112	–	–
	Causal	54.86	16.8	$0.0119	+21.53	+1.0
ALFWorld	Sequence	37.31	33.1	$0.0252	–	–
	Causal	49.55	32.6	$0.0248	+12.24	-0.5
ScienceWorld	Sequence	14.80	12.5	$0.0095	–	–
	Causal	17.67	7.2	$0.0055	+2.87	-5.3

Table 6. Execution efficiency across three environments. Performance is task success (%) for Retail and ALFWorld and environment score for ScienceWorld; Calls and Cost/Run report average API usage and cost.

\Delta

values are relative to Sequence.

Metric	Measurement
Pipeline stages	6
Processed traces	140
Construction cost	$0.045
Raw-trace communication	90.8 KB
Federated-DAG communication	1.4 KB
Communication reduction	$64.9\times$

Table 7. End-to-end CausalCompose construction and communication measurements.

Depth	Sequence MSE	Oracle MSE	Gap
1	3.432	0.040	3.392
2	3.273	0.040	3.234
3	3.447	0.040	3.407
4	4.155	0.040	4.115
6	3.354	0.040	3.314
8	4.279	0.040	4.239

Table 8. Synthetic dependency-depth sweep. Gap is the Global Sequence WM MSE minus Oracle Causal WM MSE.

Causal World Models

Causal world models represent how entities and environments change under actions, enabling prediction, planning, and counterfactual reasoning.

See all 47 papers

Steering Black-Box Large Language Model Agents with Pitfall Anticipation World Model
Oct 21, 2026Jiongxiao Wang, Han Ding, Youzhi Luo +4LLM World Models LLM Agents
University of Wisconsin–Madison · AWS AI
Learning Implicit Causal World Models from Multi-Agent Demonstrations
Jul 28, 2026Jasorsi GhoshCausal Representation Learning Causal World Models
Department of Computer Science Purdue University West Lafayette, Indiana, USA
CausaLab: A Scalable Environment for Interactive Causal Discovery Toward AI Scientists
May 25, 2026Junlin Yang, Dylan Zhang, Xiangchen Song +7Structural Causal Models LLM Agent Evaluation
Tsinghua University · University of Illinois Urbana-Champaign · Carnegie Mellon University +2

Oct 21, 2026J/K move · Enter open · S save

Steering Black-Box Large Language Model Agents with Pitfall Anticipation World Model

Jiongxiao Wang, Han Ding, Youzhi Luo, Yawei Wang+3

University of Wisconsin–Madison · AWS AI

LLM World Models LLM Agents Long-Horizon LLM Agents LLM Agent Reliability RL for LLM Agents

Open paper AskPDF arXiv

Frontier black-box Large Language Model (LLM) agents are proficient at long-horizon reasoning and precise action generation. However, when it comes to a specific agentic task, they have no internalized knowledge of how that particular environment unfolds. A world model offers a natural remedy, providing the agent with environment specific knowledge by foreseeing how the environment will respond to the action it is about to take. Prior world models for LLM agents mainly predict next observations at each step, which is costly and often uninformative because individual observations provide limited task-relevant guidance and not every step requires call for foresight. This raises a key question: what a world model should predict and when should it be invoked for agentic tasks? Our pilot study suggests that world model guidance is most useful when it anticipates pitfalls within a bounded multi-step future. Motivated by this, we introduce Pitfall Anticipation World Model (PAWM), which selectively intervenes when a potential pitfall arises and predicts the impending failure to steer the agent away from it. PAWM is trained in two stages: Supervised Fine-Tuning (SFT) on synthetic data labeled with agent pitfalls, followed by GRPO co-training with a black-box LLM to further optimize it in the target environment. PAWM improves Claude Sonnet 4.5 performance by 24.3% on ScienceWorld and 7.7% on WebArena-Lite, outperforming a range of baselines. These results demonstrate that PAWM provides an effective form of world model guidance for black-box LLM agents.

Abstract

Causal World Models

Explore similar work

Steering Black-Box Large Language Model Agents with Pitfall Anticipation World Model

Learning Implicit Causal World Models from Multi-Agent Demonstrations

CausaLab: A Scalable Environment for Interactive Causal Discovery Toward AI Scientists