Imagine asking an AI to analyze a complex image, only for it to confidently hallucinate details that aren’t there. This happens because traditional models tend to lock into early perceptions, struggling to weigh evidence from different parts of a scene simultaneously. A new framework called Visual Para-Thinker++ aims to fix this by mimicking how human teams collaborate, turning a single model into a coordinated group of specialized agents.

Instead of relying on one chain of thought, this system uses a shared large multimodal language model instantiated as three distinct roles: a Main Agent, Worker Agents, and a Summary Agent. The Main Agent breaks down the problem, while multiple Worker Agents tackle different aspects in parallel, isolated from each other to prevent confusion. Crucially, the Summary Agent does not just pick the most popular answer; it reviews the full reasoning traces of all workers to reconcile the final conclusion, reducing the risk of errors.

The real innovation lies in how these agents learn to work together. Researchers developed a training method that assigns specific rewards to different roles, preventing their learning signals from conflicting. This allows the system to optimize each agent’s contribution without interfering with the others. An efficient inference engine further boosts performance by reusing visual data and memory caches, making the process faster and more resource-friendly.

In tests across V*, CountBench, the RefCOCO family, and HallusionBench, Visual Para-Thinker++ consistently outperformed single-trajectory models and other parallel baselines. The gains were particularly striking in tasks prone to hallucination, proving that delegating tasks to specialized, parallel agents yields more accurate and reliable results. This approach suggests that the future of visual AI lies not in bigger single models, but in smarter, collaborative architectures.