Current Agents Fail to Leverage World Model as Tool for Foresight

Chen Qian, Emre Can Acikgoz, Bingxuan Li, Xiusi Chen, Yuji Zhang, Bingxiang He, Qinyu Luo, Dilek Hakkani-Tür, Gokhan Tur, Yunzhu Li, Heng JiarXiv 2026 · arXiv · OpenReview

TL;DR

Modern AI agents struggle to use "world models" (simulators of the future) effectively for planning. Despite having access to these powerful tools, agents often fail to invoke them, misuse their predictions, or even perform worse when they are available. This suggests the core challenge isn't building better simulators, but teaching agents how to strategically and reliably use them for foresight.

What Problem Does This Paper Address?

As AI agents tackle increasingly complex, long-horizon tasks, they need to anticipate future consequences of their actions, not just react to immediate stimuli. Generative world models offer a potential solution by allowing agents to "simulate" potential futures before acting in the real world. However, it's unclear whether current agents can effectively leverage these world models as tools to improve their foresight and decision-making. This paper investigates this gap: can agents strategically use world models, and if not, why?

Methods

The researchers framed world models as external tools that agents can optionally call upon. They designed an evaluation framework to systematically study:

  • Agentic tasks: Tasks requiring visual grounding, perception-to-action mapping, and multi-step reasoning in interactive environments (e.g., Frozen-Lake, Navigation, PrimitiveSkill, Sokoban).
  • Pure VLM reasoning tasks: Tasks where agents interact through perceptual or spatial reasoning without embodied control (e.g., 3DSRBench, MMSI, SAT, Spatial-MM Object).

They tested agents in three modes:

  1. Normal Mode: Agents can choose to use the world model or act directly.
  2. WM Invisible Mode: Agents are unaware of the world model (baseline).
  3. WM Force Mode: Agents are required to simulate before acting.

They analyzed agent behavior by measuring world model usage rates, performance gains/losses, and performing detailed attribution analysis to understand why agents succeeded or failed.

Datasets

The paper evaluated agents on existing benchmarks across two categories:

  • Agentic Tasks: Frozen-Lake, Navigation, PrimitiveSkill, and Sokoban. These tasks cover long-horizon interaction, 2D symbolic and 3D photorealistic environments, and various embodiments.
  • Pure VLM Reasoning Tasks: 3DSRBench, MMSI Bench, SAT, and Spatial-MM Object. These tasks require rich spatial reasoning and visual grounding, where simulated outcomes could theoretically aid perceptual reasoning.

For agentic tasks, the "world model" was often a cloned environment state or a generative dynamics model. For VQA tasks, they used WAN2.1 as the simulator.

Key Insights and Ideas

  • World models aren't magic bullets: Simply having access to a world model doesn't guarantee better performance. In fact, it can sometimes degrade it.
  • The core problem is governance, not generation: The real challenge is teaching agents when to simulate, how to interpret the results, and how to integrate that foresight into their decision-making process.
  • Agents are hesitant to use world models: Even when beneficial, agents often rarely invoke simulation tools, suggesting they don't recognize their utility or lack a strategy for when to use them.
  • Forcing usage backfires: Mandating world model use often leads to worse performance, highlighting that strategic, optional use is key.
  • Model family matters: Different LLM families (GPT, Llama, Qwen) exhibit distinct patterns in their willingness to use world models, with some showing higher usage but little benefit.
  • More calls don't mean better results: In agentic tasks, excessive world model calls often correlate with degraded performance, indicating confusion or over-analysis rather than effective foresight.

Key Results

  • Performance degradation: In agent tasks, models often performed worse (up to 5% degradation) when world models were available or enforced. VQA tasks showed marginal gains or no significant difference.
  • Low usage rates: Agents rarely invoked simulation tools (fewer than 1% of the time for VQA tasks).
  • Misuse of predictions: When agents did use world models, they frequently misused the predicted rollouts (approx. 15% of the time), leading to inconsistent or degraded performance.
  • Agentic tasks vs. VQA: Agents used world models significantly more in agentic tasks (order of magnitude higher) than in VQA tasks, suggesting foresight is more naturally integrated into dynamic, sequential decision-making.
  • Forced usage impact: Forcing world model use led to a steeper drop in performance than normal mode, reinforcing that strategic use is crucial.

Limitations and Future Work

  • Limited model families: The study focused on a few LLM families (GPT, Llama, Qwen) and excluded others like Gemini and Claude due to cost.
  • Simulator quality: The use of different simulators for agentic and VQA tasks might introduce mild cross-task incomparability. WAN2.1-generated frames sometimes had quality artifacts that confused agents.
  • Ambiguous instructions: Many failure cases stemmed from underspecified or ambiguous agent instructions, degrading simulation quality.

Future work should:

  • Expand coverage to more model families.
  • Employ stronger, more consistent simulators.
  • Develop mechanisms for "foresight governance": teaching agents when to simulate (input governance), how to interpret results (meaning governance), and when to act (action governance).
  • Explore architectures that maintain stable internal states across simulated and real trajectories.
  • Develop interfaces that encourage agents to articulate and refine hypotheses before querying world models.
  • Investigate how agents build and anchor beliefs across possible futures.

Summary updated

Abstract

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.
BibTeX
@misc{qian2026current,
  author = {Chen Qian and Emre Can Acikgoz and Bingxuan Li and Xiusi Chen and Yuji Zhang and Bingxiang He and Qinyu Luo and Dilek Hakkani-Tür and Gokhan Tur and Yunzhu Li and Heng Ji},
  title = {{Current Agents Fail to Leverage World Model as Tool for Foresight}},
  year = {2026},
  eprint = {2601.03905},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.03905},
}
Figure