Скачать или смотреть Mini-o3: Scaling Up Reasoning Patterns and Interaction Turns for Visual Search

Mini-o3 is a novel system designed to enhance Vision-Language Models (VLMs) in solving challenging visual search tasks that require extensive trial-and-error and deep reasoning, a common limitation in existing open-source approaches. To achieve this, Mini-o3 scales up tool-based interactions, enabling multi-turn reasoning that can span tens of steps. Its development relies on three core components: first, the *Visual Probe Dataset**, a collection of thousands of difficult visual search problems specifically crafted to demand exploratory reasoning; second, an **iterative data collection pipeline* that synthesizes diverse reasoning patterns like depth-first search and trial-and-error for initial supervised fine-tuning; and third, an *over-turn masking strategy* used during reinforcement learning to prevent penalizing responses that exceed the maximum number of training turns. This masking technique is crucial because it allows the model to generate trajectories that naturally scale to tens of turns during inference, significantly improving accuracy on complex problems, even though it was only trained with an upper limit of six interaction turns. As a result, Mini-o3 achieves state-of-the-art performance on difficult visual search benchmarks by generating rich reasoning patterns and deep thinking paths.

https://arxiv.org/pdf/2509.07969