The paper introduces a training paradigm that gives LLM agents an internal world model for future-aware planning. Standard agents are reactive in long-horizon tasks, but this approach enables them to simulate outcomes before acting. The authors identify a format-capability gap in existing fine-tuning methods and propose three stages: World Model Agentic Mid-Training to inject predictive capabilities, Format-Eliciting SFT to structure those capabilities, and Foresight-Conditioned Reinforcement Learning to refine calibration.
Evaluated on search and mathematical reasoning tasks, the approach consistently outperforms other training baselines. This is interesting because it transforms LLM agents from reactive to proactive planners, addressing a key limitation in long-horizon decision-making without requiring external simulators.