In a debate system that shares retrieved evidence, source traceability and answer agreement require separate evaluation.
Retrieval and debate
Multiagent Debate studies a procedure in which several responses consult and revise each other. Dense Passage Retrieval retrieves relevant passages by embedding questions and passages. Combining these ideas does not automatically improve debate accuracy.
This article presents a design example that supplies retrieved documents to several agents. It does not reproduce the algorithm or performance results of a specific paper.
Information to share
Document text alone makes it hard to determine which sentence and version supported a claim. Shared evidence should include a document identifier, version, location, and source text. Link summaries to their originals, and do not automatically register agent-generated claims as new facts.
| Information | Purpose |
|---|---|
| Document ID and version | Check whether agents read the same material |
| Passage location and source text | Check whether the claim is supported by an actual sentence |
| Retrieval query | Trace why a candidate was selected |
| Claim-to-source link | Review the evidence for the agreed answer |
Processing flow
flowchart TD
Q["Question"] --> R["Retrieve documents"]
R --> E["Evidence bundle with sources"]
E --> A["Independent initial responses"]
A --> D["Compare responses and evidence"]
D --> V["Check final claims against originals"]
V --> F["Answer or identify missing evidence"]Initial responses can interpret the same evidence differently. In subsequent debate, compare what each sentence supports rather than repeating another agent's conclusion. If more material is needed, record and review the additional search results before adding them to the evidence bundle.
Evaluation
Compare a single generation with retrieval, debate without retrieval, and retrieval combined with debate. Match total token and call budgets, and score answer accuracy separately from citation support. Distinguish the effect of shared evidence from the effect of additional calls.
Higher agreement can still reflect shared dependence on an incorrect document. Record retrieval omissions, incorrect source selection, and misinterpretation separately. State when the retrieved material does not provide sufficient evidence.
Summary
Sharing evidence is a design choice for tracing the material behind multiple responses. Sharing material does not make interpretations or conclusions correct. Evaluate retrieval and debate separately, and do not automatically store generated claims as verified knowledge.