Reproduce “CoMAP: Co-Evolving World Models and Agent Policies for LLM Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Youwei Liu, Jian Wang, Hanlin Wang, Wenjie Li
The COMAP framework introduces a co-evolutionary process where a textual world model continuously adapts to an LLM agent's evolving policy, using its future predictions as a self-improving signal to refine agent decision-making. This approach reduces reliance on external feedback, achieving an average 2.58% absolute improvement in task success rates on the Qwen3-8B backbone over state-of-the-art baselines and enhancing world model prediction quality.
Key insights
- Continuous co-evolution between a world model and an agent policy is critical for maintaining accurate future predictions as agent behavior evolves, especially in long-horizon tasks.
- Internal future predictions from a dynamic world model can serve as a robust, self-improving signal for policy refinement, reducing the need for external feedback.
- The combination of future-aware reflection and on-policy self-distillation enables LLM agents to achieve more robust decision-making and better capture environmental dynamics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.