Internalizing the Future: A Unified Agentic Training Paradigm for World Model Planning
The paper introduces a unified training paradigm that equips large language model agents with internal world modeling for future-aware planning. It outlines a three-stage process—World Model Agentic Mid-Training, Format-Eliciting SFT, and Foresight-Conditioned Reinforcement Learning—to embed and calibrate predictive capabilities. Experiments on search and mathematical reasoning benchmarks demonstrate performance gains over existing baselines.
- ▪The authors propose training a single autoregressive model to generate prospective state rollouts and plan-conditioned success estimates, providing a textual analogue of Q-values.
- ▪A three-stage training pipeline (WM-AMT, FE-SFT, FC-RL) is designed to inject, structure, and refine latent predictive abilities within the policy.
- ▪Evaluations on search and mathematical reasoning tasks show the method consistently outperforms other training baselines.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2606.27483 (cs) [Submitted on 25 Jun 2026] Title:Internalizing the Future: A Unified Agentic Training Paradigm for World Model Planning Authors:Xuan Zhang, Zhijian Zhou, Lingfeng Qiao, Yulei Qin, Ke Li, Xing Sun, Xiaoyu Tan, Chao Qu, Yuan Qi View a PDF of the paper titled Internalizing the Future: A Unified Agentic Training Paradigm for World Model Planning, by Xuan Zhang and 8 other authors View PDF Abstract:Large language model (LLM) agents have demonstrated strong capability in sequential decision-making, yet they remains fundamentally reactive in long-horizon tasks. Unlike humans who employ "what-if" reasoning to evaluate potential plans before commitment, standard agents lack an internal world model to simulate future outcomes.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv.org.
Discussion
0 commentsMore from arXiv.org
