Reproduce “Coordinating Search-Informed Reasoning and Reasoning-Guided Search in Claim Verification” 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

Qisheng Hu, Quanyu Long, Wenya Wang

HARIS, a hierarchical multi-agent framework, coordinates reasoning and information search to address multi-hop claim verification by explicitly modeling their dynamic interplay. This approach achieved superior accuracy on EX-FEVER, HOVER, and CHECKWHY datasets, including up to 62.80% for 3-hop and 61.00% for 4-hop claims on HOVER, demonstrating enhanced generalization for complex verification tasks.

Key insights

  • Decoupling complex reasoning and precise information retrieval into specialized LLM agents significantly enhances multi-hop claim verification accuracy and decision traceability.
  • Dynamic, interleaved coordination between reasoning-driven search and search-informed reasoning is crucial for uncovering implicit bridging facts and refining verification paths.
  • Reinforcement learning with group-level reward normalization (GRPO) effectively trains cooperative LLM agents for knowledge-intensive tasks, promoting stable and consistent learning without extensive intermediate supervision.

Summary by alphaXiv.

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