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

Steering Black-Box Large Language Model Agents with Pitfall Anticipation World Model. By Jiongxiao Wang and six others, from the University of Wisconsin–Madison and AWS AI. Published on the twenty-first of October. No code repository is linked to it.

Frontier black-box language model agents are proficient at long-horizon reasoning and at generating precise actions. Put one into a specific agentic task, though, and it has no internalized knowledge of how that particular environment unfolds. A world model is a natural remedy. It gives the agent environment-specific knowledge by foreseeing how the environment will respond to the action the agent is about to take.

Prior world models for language model agents mostly predict the next observation at every step. The authors find two faults with that. It is costly. And it is often uninformative, because a single observation offers limited guidance relevant to the task, and not every step calls for foresight in the first place. That leaves a pair of open questions: what should a world model predict for agentic tasks, and when should it be invoked?

A pilot study points to an answer. World model guidance is most useful when it anticipates pitfalls within a bounded multi-step future. From that comes the Pitfall Anticipation World Model, PAWM. Rather than speaking at every step, it intervenes selectively, when a potential pitfall arises, and it predicts the impending failure so that the agent can be steered away from it. The prediction target changes from the next observation to the failure that lies a few steps ahead. That answers both of the open questions at once. What to predict is the pitfall, not the observation. When to predict is when a pitfall is in view, not at every step. The cost of foresight is then paid only where foresight is likely to help.

PAWM is trained in two stages. The first is supervised fine-tuning on synthetic data labeled with agent pitfalls. The second is GRPO co-training alongside a black-box language model, which further optimizes the world model in the target environment. The agent itself stays a black box throughout. It is the world model that learns to anticipate where that particular agent goes wrong in that particular setting.

The agent being steered is Claude Sonnet 4.5. With PAWM, its performance improves by twenty-four point three percent on ScienceWorld and by seven point seven percent on WebArena-Lite, outperforming a range of baselines. The authors read these results as showing that PAWM is an effective form of world model guidance for black-box agents. Two environments, one frontier agent left unchanged, and in both the selectively invoked world model adds to what the agent could do on its own.

The shift is from forecasting everything to forecasting what matters. A world model that waits for a likely pitfall, and then says what is about to go wrong, gives a black-box agent the environment knowledge it was missing without asking it to change, and without paying for a prediction at every step.