Debating the Unspoken: Role-Anchored Multi-Agent Reasoning for Half-Truth Detection

2026-04-22 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, long

Summary

RADAR (Role-Anchored multi-agent Debate for hAlf-truth Reasoning) is a novel framework designed to detect half-truths, which are factually correct but misleading claims due to omitted context. Unlike traditional fact verification systems that focus on explicit falsehoods, RADAR addresses omission-based manipulation by reasoning about both stated and unstated information. It employs a multi-agent debate structure where a "Politician" agent constructs a supportive narrative from retrieved evidence, a "Scientist" agent critically probes for missing context, and a neutral "Judge" moderates and issues a verdict. The framework incorporates a dual-threshold early termination controller to adaptively conclude debates, balancing reasoning depth with computational cost. Experiments on the PolitiFact-Hidden benchmark, using GPT-4o-mini, demonstrate that RADAR consistently outperforms strong single- and multi-agent baselines in omission detection accuracy while reducing reasoning costs.

Key takeaway

For research scientists developing advanced fact verification systems, RADAR offers a robust approach to tackle the challenging problem of half-truth detection. You should consider adopting a role-anchored multi-agent debate architecture, particularly when dealing with noisy or incomplete evidence. This framework's ability to surface omitted context through structured interaction and adaptive control can significantly improve accuracy and efficiency compared to single-pass or fixed-role methods.

Key insights

Role-anchored multi-agent debate with adaptive control effectively uncovers missing context in half-truth detection.

Principles

• Omission-aware verification requires reasoning over latent assumptions.
• Complementary reasoning roles enhance context discovery.
• Adaptive termination balances accuracy and efficiency.

Method

RADAR constructs an evidence pool, then conducts a multi-round debate between a Politician (advocacy), Scientist (critical analysis), and Judge (moderation/verdict), with adaptive early stopping based on stop margin and verdict confidence.

In practice

• Implement distinct agent roles for adversarial reasoning.
• Ground all agent arguments in shared retrieved evidence.
• Use dual thresholds for adaptive debate termination.

Topics

• Half-Truth Detection
• Multi-Agent Reasoning
• Role-Anchored Debate
• Omission-Aware Fact Verification
• Adaptive Early Stopping

Code references

• tangyixuan/RADAR

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Debating the Unspoken: Role-Anchored Multi-Agent Reasoning for Half-Truth Detection — Role-anchored multi-agent debate with adaptive control effectively uncovers missing context in fact verification.
• From Argument Components to Graphs: A Multi-Agent Debate with Confidence Gating for Argument Relations — Multi-agent debate with confidence gating improves training-free argument relation identification and offers interpretability.
• CRAFT: A Unified Counterfactual Reasoning Framework for Tabular Question Answering and Fact Verification — Counterfactual reasoning significantly improves LLM performance and robustness in complex tabular QA and fact verification.
• MARCH: Evaluating the Intersection of Ambiguity Interpretation and Multi-hop Inference — Multi-hop QA requires models to navigate layered ambiguity across complex reasoning paths.
• Multi-Sourced, Multi-Agent Evidence Retrieval for Fact-Checking — WKGFC unifies structured KGs and web evidence via an LLM agent for adaptive, self-improving fact-checking.
• More information leads to more accurate beliefs—except in echo chambers — Echo chambers, even with free information flow, can reduce group belief accuracy by reinforcing existing, potentially inaccurate, views.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.