Reproduce “Agent Planning with World Knowledge Model” 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
Shuofei Qiao, Runnan Fang, Ningyu Zhang, Yuqi Zhu, Xiang Chen, Shumin Deng, Yong Jiang, Pengjun Xie and 2 more
A parametric World Knowledge Model (WKM) enables Large Language Model agents to self-synthesize global task knowledge and local state knowledge by analyzing both expert and rejected trajectories. This approach improves planning efficiency, reduces hallucinatory actions, and allows enhanced open-source models to outperform stronger closed-source models on various interactive environments.
Key insights
- Synthesizing task knowledge by contrasting expert and sampled (rejected) trajectories is more effective than relying solely on expert trajectories, highlighting the value of learning from less optimal paths.
- A "weak-guide-strong" paradigm is feasible, where a smaller LLM as a World Knowledge Model can effectively guide and improve the planning performance of stronger agent models like GPT-4.
- Training a single multi-task World Knowledge Model across multiple datasets can lead to performance improvements and points towards the potential for a unified world knowledge model.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.