Politeness
Respectful corrections across evaluated claims.
← All publications
COLM 2025
Multi-Agent Retrieval-Augmented Framework for Evidence-Based Counterspeech Against Health Misinformation
A coordinated system of specialized agents that retrieves, synthesizes, generates, and refines evidence-based counterspeech.
The problem
Why this work matters
Conventional RAG systems often depend on a narrow evidence source and provide limited control over response quality. For health misinformation, a useful correction must be current, factual, relevant, and respectful at the same time.
Methodology
How the approach works
Contributions
What this work contributes
- 01
A modular multi-agent architecture for evidence retrieval, summarization, response generation, and refinement.
- 02
Joint use of curated medical guidance and real-time web evidence for both authority and timeliness.
- 03
A public Reddit health-misinformation dataset covering COVID-19, influenza, and HIV claims.
- 04
Ablation and human-preference studies that isolate the value of each agent and evidence path.
Evaluation
Key results
Factual accuracy
Improved evidence grounding over RAG baselines.
Informativeness
More useful detail from complementary sources.
Relevance
Responses remain focused on the original claim.