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

Why Keep Your Doubts to Yourself? Trading Visual Uncertainties in Multi-Agent Bandit Systems

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 0 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 Agora, a market-based multi-agent VLM framework that treats epistemic uncertainty as a tradable asset, with profitability-driven trades and a Thompson-Sampling-style market-aware broker, reporting gains on five multimodal benchmarks at lower cost.

Problem

Coordinating multiple heterogeneous VLM agents is economically costly; existing heuristic coordinators (Mixture-of-Agents, knowledge-based routers) ignore costs and collapse uncertainty structure.

Approach

Reframes multi-agent coordination as a decentralized market in which multi-dimensional (perceptual, semantic, inferential) uncertainty is traded between agents only when a cost-delta is negative, with a market-aware Thompson Sampling broker choosing the initial agent.

Key results

• Agora reaches 79.2% on MMMU (Val), +8.5 points over the best baseline, per the main table.
• On MMBench_V11_Test, Agora achieves 89.5% accuracy vs. 87.12% for KABB-VLM and 86.65% for MoA, with MoA at 3.11x relative cost.
• Ablation: removing the strategic uncertainty factor drops MMBench accuracy from 89.50% to 86.42%; 'Only Net Return' drops to 82.15%.

Surprising findings

• A single agent configuration (N=1) reportedly reaches 87.5% on MMBench at a very low cost ratio, outperforming gemini-2.5-pro and InternVL3-78B on that metric.

Limitations

• Proposition 1 and Theorem 1 are asserted informally with no rigorous proof in the main text; the greedy descent only guarantees a local equilibrium.
• Main results use API-accessed models with greedy decoding and no variance or seed reporting.
• Uncertainty definitions include many unspecified functions and hyperparameters (gamma, lambda, weights) left to the appendix.

Experimental setup

Datasets: MMMU MMBench MathVision InfoVQA CC-OCR

Benchmarks: MMMU MMBench_V11_Test MathVision InfoVQA CC-OCR

Metrics: accuracy relative cost inference time collaboration overhead index final epistemic uncertainty UAPS cost-performance ratio

Baselines: Mixture-of-Agents KABB-VLM FrugalGPT RouteLLM EmbedLLM HybridLLM PPO selector MCTS selector A2C selector DQN selector gpt-4o-2024-08-06 gemini-2.5-pro-exp-03-25

Builds on: Thompson Sampling Mixture-of-Agents KABB

Closest papers by meaning

• GAM-Agent: Game-Theoretic and Uncertainty-Aware Collaboration for Complex Visual Reasoning (NeurIPS 2025 · similarity 0.88)
• VS-Bench: Evaluating VLMs for Strategic Abilities in Multi-Agent Environments (CVPR 2026 · similarity 0.86)
• VAGEN: Reinforcing World Model Reasoning for Multi-Turn VLM Agents (NeurIPS 2025 · similarity 0.85)
• CertainlyUncertain: A Benchmark and Metric for Multimodal Epistemic and Aleatoric Awareness (ICLR 2025 · similarity 0.84)
• Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.84)
• From Debate to Equilibrium: Belief‑Driven Multi‑Agent LLM Reasoning via Bayesian Nash Equilibrium (ICML 2025 · similarity 0.84)