Research map / Multi-agent systems and games / LLM agents
GAM-Agent: Game-Theoretic and Uncertainty-Aware Collaboration for Complex Visual Reasoning
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 GAM-Agent, a game-theoretic multi-agent debate framework in which base and critical VLM agents exchange structured claims, evidence and uncertainty estimates, and debate iteratively until consensus, improving accuracy on MMMU, MMBench, MVBench and V*Bench.
Problem
Multi-agent debate methods for vision-language models use simple averaging or voting and struggle on high-complexity visual reasoning tasks requiring multi-step reasoning and handling visual ambiguity.
Approach
Models the interaction between base perception agents and critical verifier agents as a non-zero-sum game, where uncertainty (token-level entropy and probability margin, or text-based semantic markers) controls agent weighting, claim-evidence parsing grounds claims in image regions, and multi-round debate is triggered and terminated by uncertainty and conflict thresholds.
Key results
- GAM-Agent improves Qwen2.5VL-7B on MMBench_V11_Test from 82.61 to 89.02 (+6.41) and InternVL3-14B on MMBench from 83.54 to 90.15 (+6.61).
- On V*Bench, GAM-Agent reaches 78.32% overall with GPT-4o-0513 (+3.60 over base 74.72), above the SEAL reference of 75.39%.
- Across the four models on MMMU, GAM-Agent gives gains of +2.74 to +5.17 points over base models, leading all compared multi-agent frameworks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GAM-Agent instead of DebUnc | VLM backbones · MMMU, MMBench_V11_Test, MVBench_Test | accuracy: GAM-Agent achieves the highest reported accuracy for each base model on all three benchmarks; e.g. InternVL3-14B/MMBench | Helped |
| Use GAM-Agent instead of DebUnc | VLM backbones · V*Bench Overall | accuracy: DebUnc 65.87 -> GAM-Agent 66.51 (Qwen2.5VL-7B); DebUnc 77.87 -> GAM-Agent 78.32 (GPT-4o-0513) | Helped |
| Remove iterative debate (ablation) | GAM-Agent with InternVL3-14B · MMBench_TEST_V11 | accuracy: -4.28% | Hurt |
| Remove claim-evidence grounding (P&M) (ablation) | GAM-Agent with InternVL3-14B · MMBench_TEST_V11 | accuracy: -4.72% | Hurt |
| Use a single agent without collaboration or debate (ablation) | GAM-Agent with InternVL3-14B · MMBench_TEST_V11 | accuracy | Hurt |
| Remove uncertainty quantification and use uniform weights/simple heuristics (ablation) | GAM-Agent with InternVL3-14B · MMBench_TEST_V11 | accuracy | Hurt |
| Replace dynamic agent weights with fixed weights (ablation) | GAM-Agent with InternVL3-14B · MMBench_TEST_V11 | accuracy | Hurt |
| Increase maximum debate rounds from 0 to 3 (ablation) | GAM-Agent with Qwen2.5VL-7B · MMBench_V11_Test | accuracy: 82.97% (0 rounds) -> 86.53% (3) | Helped |
Surprising findings
- The paper reports that the ablation without multi-agent collaboration (N=1) performs only slightly above the base model, while removing P&M causes a 4.72% drop.
Limitations
- Single global uncertainty threshold for all tasks may be suboptimal, as the paper itself acknowledges.
- Baseline methods were re-implemented and adapted by the authors, and the V*Bench base numbers come from the authors' own evaluations, which limits comparability.
- Uncertainty-metric evaluation (UA, CE, DA) relies on heuristic ground truth and proxies not clearly defined.
Experimental setup
Datasets: MMMU MMBench_V11_Test MVBench_Test V*Bench
Benchmarks: MMMU MMBench_V11_Test MVBench_Test V*Bench
Metrics: accuracy debate rounds inference cost
Baselines: DMAD DMPL ChatEval MAD DebUnc SEAL
Builds on: multi-agent debate DebUnc ChatEval MAD
Closest papers by meaning
- Agent-X: Evaluating Deep Multimodal Reasoning in Vision-Centric Agentic Tasks (ICLR 2026 · similarity 0.91)
- VipAct: Visual-Perception Enhancement via Specialized VLM Agent Collaboration and Tool-use (AAAI 2026 · similarity 0.90)
- Agentic Jigsaw Interaction Learning for Enhancing Visual Perception and Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.90)
- VS-Bench: Evaluating VLMs for Strategic Abilities in Multi-Agent Environments (CVPR 2026 · similarity 0.89)
- Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.89)
- MATA: A Trainable Hierarchical Automaton System for Multi-Agent Visual Reasoning (ICLR 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.