AI

Debate to Deceive: Poisoning LLM's Fact-Checking via Game-Theoretic Multi-Agent Debate

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 09/11/2026, 05:03:25 UTC.

Briefings

A new multi-agent debate framework called Debate-to-Deceive (D2D) was developed to attack Large Language Models by generating semantically coherent, rhetorically effective adversarial passages. The system formalizes attacks as an evolutionary game where agents adaptively evolve rhetorical strategies (Logos, Pathos, Ethos) through competition. On four fact-checking benchmarks (FEVER, HealthVer, CLIMATE-FEVER, and SciFact), D2D substantially outperforms existing attacks and evades most defenses. The research is published on arXiv:2611.08756v1.

Key points

• Debate-to-Deceive (D2D) is a multi-agent framework that formalizes fact-checking attacks as an evolutionary game where agents adaptively evolve rhetorical strategies (Logos, Pathos, Ethos) through competition.
• On four fact-checking benchmarks (FEVER, HealthVer, CLIMATE-FEVER, and SciFact), D2D substantially outperforms existing attacks and evades most defenses.

Context

Knowledge attacks on Large Language Models, such as manipulating fact-checking judgments, pose significant risks to AI reliability. Existing approaches like adversarial suffix injection and single-agent self-reflection lack robustness and are prone to confirmation bias. This research proposes D2D as a new solution grounded in human debate structure.

Limitations

The evidence is from a preprint on arXiv:2611.08756v1, meaning it is unverified and subject to potential revision. The study covers only four specific fact-checking benchmarks, limiting generalizability. The authors do not provide causal claims or specific dates for the attack's effectiveness, only that D2D substantially outperforms existing methods on the tested datasets.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 09/11/2026, 07:12:14 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeigpaovjyr6ooz33imkldl6srthiopotm7cwco6npekgfbtduvxk7e