Source-linked AI summary
Current Agents Fail to Leverage World Model as Tool for Foresight
Cheng Qian, Emre Can Acikgoz, Bingxuan Li, Xiusi Chen, Yuji Zhang, Bingxiang He, Qinyu Luo, Dilek Hakkani-Tür, Gokhan Tur, Yunzhu Li, Heng Ji
TL;DR
Current agents need foresight for long-horizon decisions, but it is unclear whether they can strategically use external world models. This paper evaluates agents that optionally simulate candidate actions before acting and finds that tool use, interpretation, and integration remain unreliable. The results motivate more explicit mechanisms for calibrated interaction with simulations.
Problem
Long-horizon tasks require agents to anticipate evolving environments, yet current agents lack robust foresight and may not know when to use external world models.
Method
The paper evaluates agents that use world models as optional tools for simulating candidate actions and compares this with unavailable or mandatory access modes.
Results
Agents show inconsistent benefits from simulation, with difficulties deciding when to simulate, interpreting predicted outcomes, and integrating foresight into downstream reasoning.
Takeaways & Limitations
Reliable foresight requires calibrated, strategic interaction with world models rather than simply providing agents with predictive tools.
Takeaways & Limitations
The study covers a limited set of model families and uses different simulators across task types, introducing mild cross-task incomparability.
Abstract
from arXiv · showhide
Agents built on vision-language models increasingly face tasks that demand anticipating future states rather than relying on short-horizon reasoning. Generative world models offer a promising remedy: agents could use them as external simulators to foresee outcomes before acting. This paper empirically examines whether current agents can leverage such world models as tools to enhance their cognition. Across diverse agentic and visual question answering tasks, we observe that some agents rarely invoke simulation (fewer than 1%), frequently misuse predicted rollouts (approximately 15%), and often exhibit inconsistent or even degraded performance (up to 5%) when simulation is available or enforced. Attribution analysis further indicates that the primary bottleneck lies in the agents' capacity to decide when to simulate, how to interpret predicted outcomes, and how to integrate foresight into downstream reasoning. These findings underscore the need for mechanisms that foster calibrated, strategic interaction with world models, paving the way toward more reliable anticipatory cognition in future agent systems.
1 Introduction
Long-horizon agents need to anticipate evolving environments, but current agents often lack robust foresight. This paper studies world models as optional external simulation tools and finds that agents struggle to invoke, interpret, and integrate them reliably.
- Long-horizon tasks require agents to anticipate how environments evolve across interdependent decisions and potentially irreversible consequences.
- Foresight is defined here as anticipating environmental evolution rather than generating goal-directed plans from the agent’s own perspective.
- World models can serve as optional tools that agents query to simulate candidate actions before committing to real-world behavior.
- Agents often hesitate to invoke world-model tools even when simulation could improve environmental foresight.
- When agents do invoke simulations, they may produce one deterministic future, override predicted outcomes with confident internal reasoning, or fail to evaluate counterfactual branches.
- The paper frames progress as calibrating when agents consult simulations, how they interpret them, and how they integrate simulated evidence with reasoning.
2 Related Work
Prior work studies foresight through internal world modeling, imagined trajectories, and learned simulators embedded in decision loops. These approaches extend mental simulation ideas beyond text to code and other environments.
- Cognitive science describes planning as mentally simulating actions and outcomes with internal world models.
- Recent agent research uses LLMs or VLMs as dynamics models, searches imagined trajectories, or embeds learned simulators into decision loops.
3 Preliminaries
The framework lets an agent choose whether to execute an action in the real environment or simulate it with an independent world model. Simulated outcomes are retained as hypothetical trajectory observations for later reasoning.
- The protocol evaluates whether a decision-making agent can invoke a learned world model to simulate candidate actions before committing to them.
- The agent’s trajectory records prior actions and observations while the real environment contains the current state and the world model approximates its transition dynamics independently.
- A room-escape example uses simulation to compare breaking a reinforced door with crawling through a vent before selecting the viable exit.
- The policy maps the current trajectory to an action and an execution context, choosing either the real environment or the world model.
- Simulation predicts a next state without changing reality, and the simulated outcome is appended as a hypothetical observation for future reasoning.
16: return Tt+1
The framework evaluates optional, unavailable, and mandatory world-model access to separate strategic tool use from mere tool availability. It measures whether agents can decide when simulation is worthwhile and integrate predictions with real observations.
- The interaction process continues until the goal is reached or the maximum horizon is exceeded, retaining real and simulated outcomes in one trajectory.
- Interaction Modes: Normal Mode makes the world model optional, allowing agents to query it or act directly in the real environment.
- Interaction Modes: WM Invisible Mode removes access to the world model, so all actions execute directly in the real environment as a baseline.
- Interaction Modes: WM Force Mode requires agents to simulate actions before executing them, testing whether compulsory planning helps or hinders performance.
- The evaluation examines when simulation is worthwhile, how predictions are interpreted, and how simulated and real experience guide subsequent actions.
4 Experiments
The experiments test optional and mandatory world-model use across interactive agent tasks and non-embodied visual reasoning, finding limited or negative practical benefit. Results also show that invocation behavior varies by model family and that agents struggle to use simulations reliably.
- Experimental Tasks: The evaluation covers four interactive agent tasks and four visual question-answering benchmarks requiring spatial reasoning, visual grounding, or hypothetical transformations.The study compares embodied control with non-embodied perceptual reasoning to assess whether simulation helps across task types.
- Test Models: Nine vision-language models from GPT, Llama, and Qwen families are evaluated across multiple model sizes.All tested models have basic visual reasoning abilities, allowing differences to be attributed to world-model usage rather than insufficient perception.
- Evaluation Setup: Agent tasks use cloned environments to simulate hypothetical actions, while evaluations compare optional, unavailable, and—in agentic tasks—mandatory world-model access.Performance is measured by task success for agents and answer correctness for VQA.
- Performance Results: Across agent and VQA settings, world-model access does not reliably improve performance; agent results often worsen, while VQA gains remain marginal.Table 2 reports that almost all models fail to achieve higher agent-task performance with access, whereas Table 3 shows nearly unchanged VQA accuracy.
- Calling Behavior: World-model usage is consistently low, especially in VQA, where rates remain below 0.1 for all but the Llama family.The findings indicate that models lack a clear strategy for deciding when and why simulation would improve predictions.
- Calling Behavior: Calling behavior differs across families: Llama models query most proactively with little measurable benefit, while larger GPT models often bypass external help.Qwen models show a similar size-related pattern on agent tasks, with Qwen2.5-VL-7B as an unusually unwilling caller.
- Fine-Grained Outcomes: VQA cases show nearly equal rates of helpful and harmful world-model use, while agent tasks experience harmful rollouts more frequently than helpful ones.The case-level analysis concludes that current world-model access offers limited practical advantage and can be counterproductive.
5 Analysis
The analysis finds that world-model benefits are fragile because agents invoke simulations inconsistently, misinterpret them, and lack stable policies for integrating foresight into action. These governance failures differ across agent and VQA tasks, while forced or repeated use can worsen outcomes.
- Attribution Analysis: Models can extract value from simulations, but poorly calibrated use produces fragile gains and frequent regressions.The analysis motivates taxonomies of governance successes and failures across the cognitive pipeline.
- Agent Tasks: Agent-task rollouts mainly improve planning and state understanding, but action loops, over-planning, and misinterpretation destabilize execution.Agents may repeatedly re-plan without meaningful progress, making foresight beneficial in isolated cases but detrimental over long horizons.
- VQA Tasks: VQA rollouts mainly aid state understanding and visual confirmation, while unclear requests and ambiguous scenarios can amplify uncertainty.These errors reflect failures in governance of meaning, where ambiguity propagates rather than resolves uncertainty.
- Governance: Foresight governance requires deciding what to simulate, how to interpret simulations, and when to act on them; stable use depends more on governance than simulation.The proposed success pattern combines calibrated queries, unambiguous verification, and stable integration into actionable plans.
- Task Perspective: Agent tasks show higher world-model usage and benefit rates than VQA tasks, whose targeted disambiguation makes imprecise simulations more distracting.Current invocation patterns are better suited to dynamic, stateful environments than static perception queries.
- Call Perspective: Repeated world-model calls correlate negatively with task success because models often re-query instead of integrating prior rollouts into stable plans.Excessive calls frequently coincide with action loops and degraded execution, especially in agent tasks.
- Usage Perspective: Forced world-model invocation causes a steeper performance drop across agent tasks, increasing both the frequency and severity of harmful cases.Average performance declines for every task under the forced-use condition.
6 Discussions
The discussion proposes treating world-model interaction as a governed reasoning process rather than simple context expansion. It recommends structured modules and training objectives that encourage selective, diverse, and information-seeking simulation use.
- Hypothesis Testing: Confirmation-style queries can reinforce an incorrect hypothesis, whereas simulating several plausible hypotheses supports structured discrimination.This shifts simulation from affirming one guess toward testing alternatives against predicted cues.
- Integration Modules: Dedicated Decider, Reflector, and Memory modules could organize when to simulate, how to evaluate outcomes, and how to maintain task state.The proposed loop explicitly separates invocation decisions, feedback evaluation, and long- and short-term memory.
- Agent Training: Training with online multi-turn rollouts could reward appropriate invocation, diverse queries, and information gain while penalizing excessive repetition.Supervised cold-start data and indirect objectives such as reductions in hypothesis entropy are proposed to improve credit assignment.
7 Conclusion and Future Work
The conclusion characterizes world-model access as a source of new cognitive pressures, not a straightforward enhancement. Future work should develop finer-grained governance for maintaining coherent reasoning across simulated and real trajectories.
- Conclusion: World-model access introduces hesitation, over-analysis, and misaligned interpretation because agents must manage hypothetical branches alongside real experience.The observed difficulties motivate more fine-grained governance of foresight.
- Future Work: Promising directions include stable internal state across simulated and real trajectories, exploration-driven simulation policies, and hypothesis-refining interfaces.These directions aim to integrate simulation more directly into decision-making and support anticipatory cognition.
Limitations
The study’s scope is limited by model-family coverage and simulator differences across task types, including occasional artifacts and cross-task incomparability.
- Scope and Evaluation: The evaluation excludes Gemini and Claude, uses different simulators for agentic and VQA tasks, and observes occasional WAN2.1 frame artifacts.These choices may introduce mild cross-task incomparability and affect downstream performance.
- Scope and Evaluation: Many failures trace to underspecified or ambiguous agent instructions that degrade simulation quality and downstream performance.Future work will expand model coverage and use stronger, more consistent simulators to better isolate agent effectiveness.
Ethical Statement
The paper evaluates world-model use in controlled experiments rather than real-world deployment and warns that simulated futures may be overtrusted or misread in high-stakes settings.
- Controlled experiments, rather than real-world deployment, evaluate world models as tools for VLM agents.
- Simulated futures can be overtrusted or misread, potentially leading to unsafe decisions when transferred to high-stakes domains.
- The paper recommends robust oversight, evaluation, and privacy protections before safety-critical deployment involving user data.
Moon, Hiroki Furuta, Gopala Anumanchipalli, Kurt
The supplied passages consist primarily of related-work references spanning planning, world models, multimodal reasoning, robotics, benchmarks, and cognitive science.
- Prior work covers planning and long-horizon reasoning for agents operating in complex environments.
- Other cited efforts study spatial reasoning, multimodal intelligence, robotic manipulation, and dynamic agent benchmarks.
- The references also connect foresight to cognitive theories of mental models, prospection, and imagining the future.
- Related research includes learned, pretrained, and generative world models for simulating environments and dynamics.
A Evaluation Settings Details
Experiments evaluate agents under fixed inference settings, using environment simulations for agentic tasks and WAN2.1-generated frames for VQA tasks.
- Agentic tasks use frames produced by the environment while executing the model’s predicted simulation actions, with at most three frames incorporated per step.
- VQA tasks use WAN2.1 as the world model and Qwen2.5-3B-VL for prompt augmentation.
- Frames are filtered for consecutive similarity below 0.95, uniformly interpolated to three representatives, and always include the final generated frame.
- Accuracy is averaged across data points, using success rate for agent tasks and multiple-choice accuracy for VQA tasks.
B Additional Experiment Results
Additional experiments show that forcing world-model use often worsens performance, compliance remains incomplete, and benefits do not consistently exceed degradations across tasks and models.
- Forced world-model usage: Forcing world-model usage generally performs worse than allowing optional access, providing no reliable effectiveness improvement.
- Usage compliance: GPT-5’s world-model usage rate remains below 1.0 even when explicit instructions require simulation before each action.
- Usage compliance: SAT tasks show particular reluctance to use the world model, including under forced-use conditions.
- Agent behavior: Many models frequently reach the 15-step episode limit, while step count and action validity show no meaningful differences across settings.
- Comparative outcomes: Across evaluated tasks and models, performance improvements from world models do not consistently exceed performance degradations.
C Attribution Analysis Details
The attribution analysis combines manual case review with GPT-4o-assisted labeling to identify why world-model access helps or fails. Human–model agreement is then checked on reviewed cases to assess annotation reliability.
- Attribution Analysis: 60 cases were manually analyzed to identify core categories explaining when world-model assistance succeeds or fails.The analysis used these findings to curate eight attribution categories.
- Automated Attribution: GPT-4o compared trajectories with and without world-model access, assigning one primary attribution reason and any additional contributing reasons.The compared trajectories covered both optional and forced world-model settings.
- Reliability Check: 81.67% primary-attribution agreement with human judgments supports the reliability of GPT-4o’s case-level annotations.The agreement was measured on the 60 cases previously reviewed and annotated by humans.
D Additional Analysis Results
Additional analyses show that world-model invocation varies by model family, scale, and task, but more frequent or mandatory use rarely improves outcomes. Across agentic and VQA settings, repeated simulation is associated with degradation or limited corrective value.
- Usage and Performance: Higher world-model invocation does not consistently yield positive returns across model families, scales, agent tasks, and VQA settings.Smaller agent models sometimes obtain modest gains, whereas larger models often reduce usage and experience neutral or negative effects; VQA changes remain marginal.
- Call Frequency: Agent tasks induce substantially higher world-model call frequencies than VQA tasks, often by an order of magnitude.Most VQA models make fewer than one call per question on average.
- Call Frequency: Higher world-model call counts are associated with lower success rates in both agent and VQA tasks, with sharper degradation in agent tasks.The analyses report negative trends and binned degradation, linking repeated querying to uncertainty and poor foresight integration.
- Forced Usage: Mandatory world-model invocation in VQA consistently fails to improve performance and often causes modest, systematic degradation.Cases where forced use hurts outnumber cases where it helps, while nearly half of questions remain unaffected.
- Comparative Analysis: Tables 14–17 compare both-correct, world-model-helps, world-model-hurts, and both-wrong outcomes across normal and forced settings.These percentages are reported separately for agentic and VQA tasks, with each model-task combination summing to 100%.
- Evaluation Setup: The appendix presents evaluation prompts using system, initial-task, and feedback templates for each mode and task.These templates include placeholders for action limits, separators, inserted states, and simulated images.