Multi-Agent Claim Verification

A multi-agent framework for automated fact-checking and claim verification at scale.

Venue: ICWSM 2026

CARP is a multi-agent framework for automated fact-checking that addresses two key failure modes in existing systems: confirmation bias in single-agent approaches and groupthink in homogeneous multi-agent systems. It introduces heterogeneous agent teams, powered by different language models, organized as adversarial support and refutation sides, combined with structured claim decomposition and multi-hop evidence retrieval.

CARP framework overview

Key contributions:

  • Modular multi-agent architecture for end-to-end claim verification
  • Evaluation on benchmark fact-checking datasets
  • Analysis of agent collaboration and failure modes