Reproduce “EviGraph: Towards Verifiable Evidence Construction for Information-Seeking 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
Jiashun Chen, Yirong Mao, Wenhui Que
EviGraph introduces a framework for information-seeking agents that formalizes Web search as an explicit process of evidence construction using a typed, span-grounded evidence graph. This approach, implemented with a dual-role policy and reinforcement learning, enhances the verifiability and reliability of LLM-generated answers, achieving substantially higher accuracy on benchmarks like BrowseComp-Plus (35.9% vs. 8.2% for ReAct) and LiveVQA (78.0% zero-shot), while also optimizing token usage.
Key insights
- Explicitly structuring claims and evidence in a graph with polarity labels significantly enhances the verifiability and auditability of agent-generated responses by providing traceable provenance.
- Decoupling agent roles into executor, verifier, and evidence-proposal policy allows for clearer supervision of intermediate steps and helps expose distinct failure modes.
- Integrating a deterministic structural validator prevents structural hallucinations and duplicate entries, ensuring the integrity and auditability of the constructed evidence graph while mitigating reward exploitation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.