Visual Para-Thinker++: A Single-Policy Multi-Agent Framework for Visual Reasoning

2026-06-08

AI Summary

This work addresses the susceptibility of single-chain reasoning in visual reasoning tasks to early perceptual fixation and hallucination. To mitigate these issues, the authors propose a role-decoupled, shared-policy multi-agent framework comprising three specialized roles: a controller for task decomposition, workers for context-isolated parallel reasoning, and an aggregator for fusing complete reasoning trajectories. The framework employs role-conditioned training and tailored reward mechanisms to alleviate gradient conflicts, while leveraging KV cache reuse and a native multi-agent inference engine to enhance computational efficiency. Evaluated on multimodal large language models with a unified policy architecture, the approach significantly outperforms existing single-trajectory and inference-time parallel methods on V*, CountBench, RefCOCO, and HallusionBench, demonstrating particularly strong performance on tasks with high hallucination risk.

Abstract

Visual reasoning requires integrating evidence distributed across regions, attributes, and relations, making single-chain reasoning prone to early perceptual commitment and hallucination. We propose Visual Para-Thinker++, a single-policy multi-agent framework in which one shared MLLM policy is instantiated as role-conditioned Main, Worker, and Summary Agents. The Main Agent decomposes the task with fixed allocation patterns; Worker Agents reason in parallel under context isolation; and the Summary Agent reconciles full Worker reasoning traces rather than majority-voting on final labels. The shared policy is trained by Multi-Agent Capability Injection and Role-Decoupled Multi-Agent Optimization, which assign role-specific rewards and advantages to corresponding token segments to reduce gradient conflict among collaborative roles. A native inference engine enables efficient multi-agent rollout through shared visual prefix and KV cache reuse. Across V*, CountBench, the RefCOCO family, and HallusionBench, Visual Para-Thinker++ consistently outperforms single-trajectory and inference-time parallel baselines, with especially strong gains on hallucination-sensitive visual reasoning.

Problem

• visual reasoning
• hallucination
• multi-agent
• perceptual commitment
• evidence integration

Innovation

• multi-agent framework
• visual reasoning
• role-conditioned agents
• gradient conflict reduction
• KV cache reuse