PixelThink: Towards Efficient Chain-of-Pixel Reasoning

TL;DR

PixelThink introduces a reinforcement learning-based approach that dynamically adjusts reasoning complexity based on scene difficulty and model confidence, significantly improving efficiency and segmentation accuracy in multimodal reasoning tasks.

Contribution

It proposes a novel scheme integrating task difficulty and model uncertainty to regulate reasoning, along with a new benchmark and metrics for comprehensive evaluation.

Findings

Enhanced reasoning efficiency and segmentation performance

Effective compression of reasoning chains based on scene complexity

Improved generalization to out-of-distribution scenarios

Abstract

Existing reasoning segmentation approaches typically fine-tune multimodal large language models (MLLMs) using image-text pairs and corresponding mask labels. However, they exhibit limited generalization to out-of-distribution scenarios without an explicit reasoning process. Although recent efforts leverage reinforcement learning through group-relative policy optimization (GRPO) to enhance reasoning ability, they often suffer from overthinking - producing uniformly verbose reasoning chains irrespective of task complexity. This results in elevated computational costs and limited control over reasoning quality. To address this problem, we propose PixelThink, a simple yet effective scheme that integrates externally estimated task difficulty and internally measured model uncertainty to regulate reasoning generation within a reinforcement learning paradigm. The model learns to compress reasoning length in accordance with scene complexity and predictive confidence. To support comprehensive evaluation, we introduce ReasonSeg-Diff, an extended benchmark with annotated reasoning references and difficulty scores, along with a suite of metrics designed to assess segmentation accuracy, reasoning quality, and efficiency jointly. Experimental results demonstrate that the proposed approach improves both reasoning efficiency and overall segmentation performance. Our work contributes novel perspectives towards efficient and interpretable multimodal understanding. The code and model will be publicly available.