Research map / Multi-agent systems and games / LLM agents

Multi-Agent Undercover Gaming: Hallucination Removal Through Counterfactual Test for Multimodal Reasoning

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Multi-agent Undercover Gaming (MUG), a multi-agent debate protocol where one agent sees an counterfactually edited image and must be detected via voting, improving multimodal reasoning and hallucination benchmarks on Qwen2.5VL-7B and InternVL3-14B.

Problem

Multi-agent debate assumes rational, reflective debaters, but MLLM agents themselves hallucinate, so consensus can be unreliable for multimodal reasoning.

Approach

Reframes multi-agent debate as an social-deduction game: one agent gets a counterfactually edited image (generated by Step1X-Edit with question-guided edits), and agents detect the "undercover" agent via voting, giving ground truth for hallucinating agents before a summarization phase over the remaining agents.

Key results

• Qwen2.5VL-7B (MUG) reaches 50.3% on MMMU vs 45.0% base and 44.7% for MAD-Vote.
• InternVL3-14B (MUG) reaches 58.0% HallusionBench average and 91.1 F1 on POPE, vs 55.1 and 89.5 base.
• Removing the undercover mechanism drops HallusionBench by 4.49 points for Qwen2.5VL-7B (MUG).

Surprising findings

• Self-Refine and MAD-Vote often underperform the single-model baseline on several benchmarks (e.g., MAD-Vote on HallusionBench 37.8 avg vs 46.4 for Qwen2.5VL-7B).

Limitations

• Counterfactual edit generation relies on an external image editor and LLM prompting; quality of edits not fully validated.
• Gains on MMMU for InternVL3-14B are small (0.9 points over baseline).

Experimental setup

Datasets: MMStar MMMU HallusionBench POPE

Benchmarks: MMMU MMStar HallusionBench POPE

Metrics: accuracy average accuracy figure accuracy question accuracy precision recall F1

Baselines: Chain-of-thought Self-Refine MAD-Vote MAD-Judge DeepSeek-VL-7B LLaVA-OneVision-7B LLaVA-Next-Llama3-8B InternVL2-8B Gemini-1.5-Pro GPT-4o mini GPT-4V Claude-3.5-Sonnet

Builds on: Multi-Agent Debate Self-Refine Step1X-Edit chain-of-thought

Closest papers by meaning

• InEx: Hallucination Mitigation via Introspection and Cross-Modal Multi-Agent Collaboration (AAAI 2026 · similarity 0.90)
• Debate with Images: Detecting Deceptive Behaviors in Multimodal Large Language Models (ICML 2026 · similarity 0.89)
• Visual Multi-Agent System: Mitigating Hallucination Snowballing via Visual Flow (ICLR 2026 · similarity 0.88)
• Tracing the Cascade: A Topology-Aware Evaluation Framework for Scientific Agent Hallucinations (NeurIPS 2026 · similarity 0.88)
• MM-Snowball: Evaluating and Mitigating Hallucination Snowballing in Multimodal Multi-Turn Dialogue (ICML 2026 · similarity 0.88)
• Auditing Meta-Cognitive Hallucinations in Reasoning Large Language Models (NeurIPS 2025 · similarity 0.88)