Visual Para-Thinker++: A Single-Policy Multi-Agent Framework for Visual Reasoning
About
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.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Visual Hallucination Evaluation | HallusionBench | Accuracy73.8 | 156 | |
| Visual Search | V* | Accuracy85.7 | 53 | |
| Fine-Grained Perception | MMVP | Accuracy74.9 | 41 | |
| Referring Expression Grounding | RefCOCO (testA) | -- | 41 | |
| Referring Expression Grounding | RefCOCO (testB) | -- | 41 | |
| Referring Expression Grounding | RefCOCOg (test) | -- | 37 | |
| Counting | PixMo (test) | Accuracy74 | 26 | |
| Counting | Pixmo (val) | Accuracy68.2 | 26 | |
| Referring Expression Grounding | RefCOCO+ (testA) | -- | 23 | |
| Referring Expression Grounding | RefCOCO+ (testB) | -- | 23 |