Latent Dynamics · 2026-10-02
When Do Causal World Models Help Modular LLM Agents
Xinyuan Song, Zekun Cai
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
The authors show observational world models can incur irreducible intervention-time error in modular LLM agents, even when they accurately model observational trajectories. They propose CausalCompose, which collects intervention–response evidence locally, validates causal interfaces, and composes local mechanisms for intervention-aware prediction and optional control. Experiments and theory support that causal structure improves intervention-sensitive reasoning and planning, especially in structured tool-use settings.
Why it matters
Modular agents plan with actions that causally change downstream states across components. If a world model only learns observational correlations, it may fail at intervention time and errors can propagate across modules. CausalCompose targets intervention-aware prediction by learning which cross-module dependencies to compose, and can optionally use execution feedback for closed-loop action selection.
Method
- CausalCompose preserves local mechanisms per module and exchanges compact interface evidence (events, responses, direction, lag) rather than full trajectories.
- It validates proposed cross-module edges using repeated intervention opportunities, checking intervention support, destination response consistency, and direction/lag agreement; validated interfaces form a recovered causal graph.
- With the recovered graph, it composes local mechanisms to propagate candidate intervention effects for downstream prediction and can optionally gate actions using predicted consequences plus execution feedback.
Limitation
Causal interfaces cannot be reliably recovered if the relevant intervention is never executed or its downstream response cannot be observed.
Abstract (from arXiv)
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.
Related papers
- Beyond Prediction: Steering VLM Agents with Retrospective World Modeling
- EvoBelief: Falsifiable World Models for Self-Improving LLM Agents
- CausalDreamer: Learning Predictive World Models with Latent Disentanglement
- Causal world models under physical parameter shifts
- AnyGoal: On World Model Generalization Beyond Expert Rollouts
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-07-12 · Indexed by Pub-AI: 2026-10-02 UTC
When Do Causal World Models Help Modular LLM Agents
Preprint posted: 2026-07-12 · Original record ↗
- arxiv
- 2610.00012
Explicit source versions (1)
- v1 · 2026-07-12
- Preprint available
Source date: 2026-07-12 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
LLM agents increasingly act through modular systems, such as order, payment, inventory, and shipment services, where actions in one module change which transitions are valid in another. Standard world models usually fit observational traces, but this is not the quantity needed for intervention-time planning: a trace may show that payment precedes shipment without identifying whether payment authorizes shipment, inventory mediates the effect, or a hidden trigger explains both. We study this gap through FedCausalCompose, a causal world-model framework for modular LLM agents in which local actions provide intervention-response evidence for cross-module interfaces. We first show that observational world models incur an irreducible interventional error under unblocked back-door paths, that interface recovery improves with intervention-response coverage, and that an oracle causal composition can beat the non-causal lower bound when coverage and local mechanism errors are controlled. We then test the resulting prediction in diagnostic agent settings. Causal interfaces help most in structured tool environments, where API signatures expose preconditions and downstream effects. In contrast, dialogue and narrative environments often ignore raw edge lists unless a short attention anchor makes the causal information decision-relevant. These results identify a concrete condition for causal world models in LLM agents: causal structure helps when cross-module interfaces are both statistically identifiable and presented in a form the agent can use at action time.
