Reproduce “D$^2$Plan: Dual-Agent Dynamic Global Planning for Complex Retrieval-Augmented Reasoning” 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

Kangcheng Luo, Tinglang Wu, Yansong Feng

Peking University researchers developed D²PLAN, a dual-agent dynamic global planning framework that enhances search-augmented Large Language Models (LLMs) by enabling explicit planning and robust resistance to irrelevant information. The system achieves an average absolute LasJ improvement of 3.83% over strong baselines and significantly reduces error rates on complex multi-hop reasoning tasks.

Key insights

  • Explicit dynamic global planning, guided by specific RL rewards, is crucial for improving reasoning coherence and the ability of LLMs to self-correct and adapt plans during complex multi-hop reasoning.
  • The dual-agent architecture, particularly the Purifier, effectively mitigates information overload and distraction by actively assessing relevance and condensing retrieved evidence, significantly reducing context length and improving efficiency.
  • The design of plan-oriented rewards (Initial Planning and Plan Adaptation Rewards) provides fine-grained supervision, teaching LLMs not just to answer correctly but to internalize robust planning and failure-reflection-correction patterns.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.