State‑of‑the‑art black‑box Large Language Model (LLM) agents excel at long‑horizon reasoning and precise action generation, yet they lack internalized knowledge of how a specific environment unfolds. A world model can fill this gap by forecasting how the environment will react to the agent's next action. Existing world models for LLM agents typically predict the next observation at every step, which is costly and often uninformative because single observations provide limited task‑relevant guidance and not every step requires foresight.
The core question is: what should a world model predict, and when should it be invoked to benefit agentic tasks? Our pilot study indicates that guidance is most valuable when the model anticipates pitfalls within a bounded multi‑step future. Motivated by this insight, we introduce the Pitfall Anticipation World Model (PAWM), which intervenes selectively when a potential pitfall emerges and predicts the impending failure to steer the agent away.
PAWM is trained in two stages. First, supervised fine‑tuning (SFT) on synthetic data labeled with agent pitfalls equips the model with a basic notion of failure modes. Second, GRPO co‑training with a black‑box LLM refines the model in the target environment, optimizing its predictions for real‑world interaction.
Empirical results show that PAWM boosts Claude Sonnet 4.5 performance by 24.3% on ScienceWorld and 7.7% on WebArena‑Lite, outperforming a range of baselines. These findings demonstrate that PAWM offers an effective form of world‑model guidance for black‑box LLM agents.
Review: By focusing on critical failure points rather than exhaustive observation prediction, PAWM delivers a more efficient and task‑aware foresight mechanism, opening a promising direction for enhancing the robustness of black‑box LLM agents.