SketchThinker-R1: Towards Efficient Sketch-Style Reasoning in Large Multimodal Models

TL;DR

The paper tackles the inefficiency of long chain-of-thought reasoning in large multimodal systems by introducing SketchThinker-R1, a three-stage reinforcement learning framework that encourages sketch-style, concise reasoning. It creates SketchColdStart-20K data by converting long reasoning into sketch-like traces, trains a SketchJudge reward model to encourage sketch-style thinking, and optimizes with GRPO to generalize this style across diverse tasks. Across four multimodal reasoning benchmarks, SketchThinker-R1 achieves over 64% reduction in thinking token cost while preserving accuracy, demonstrating improved efficiency and interpretability. The approach integrates data-driven cold-start training, reward-shaped RL, and cross-domain evaluation, offering a practical path to faster, more transparent multimodal reasoning.

Abstract

Despite the empirical success of extensive, step-by-step reasoning in large multimodal models, long reasoning processes inevitably incur substantial computational overhead, i.e., in terms of higher token costs and increased response time, which undermines inference efficiency. In contrast, humans often employ sketch-style reasoning: a concise, goal-directed cognitive process that prioritizes salient information and enables efficient problem-solving. Inspired by this cognitive efficiency, we propose SketchThinker-R1, which incentivizes sketch-style reasoning ability in large multimodal models. Our method consists of three primary stages. In the Sketch-Mode Cold Start stage, we convert standard long reasoning process into sketch-style reasoning and finetune base multimodal model, instilling initial sketch-style reasoning capability. Next, we train SketchJudge Reward Model, which explicitly evaluates thinking process of model and assigns higher scores to sketch-style reasoning. Finally, we conduct Sketch-Thinking Reinforcement Learning under supervision of SketchJudge to further generalize sketch-style reasoning ability. Experimental evaluation on four benchmarks reveals that our SketchThinker-R1 achieves over 64% reduction in reasoning token cost without compromising final answer accuracy. Qualitative analysis further shows that sketch-style reasoning focuses more on key cues during problem solving.

Figures (11)

• Figure 1: Our SketchThinker-R1 significantly reduces the thinking cost without compromising final answer accuracy. Vanilla-R1 serves as the baseline, representing the standard R1-style trained model. Evaluation across four benchmarks from diverse domains shows that our model achieves comparable or even superior performance (see (a)). At the same time, it reduces thinking token cost by more than 64% (see (b)). During RL training, sketch-style reasoning consistently yields higher accuracy rewards (see (c)) while maintaining a much shorter reponse length (see (d)).
• Figure 2: Overview of our SketchThinker-R1 pipeline. (1) In the Sketch-Mode Cold Start stage, we convert long reasoning processes from existing multimodal reasoning datasets into sketch-style, and fine-tune the base multimodal model to instill initial sketch-style reasoning ability. (2) Next, we train a SketchJudge Reward Model, which favors sketch-style reasoning and penalizes overly verbose reasoning. (3) Finally, we perform Sketch-Thinking Reinforcement Learning on the cold-started multimodal model under the supervision of the trained SketchJudge reward model, further enhancing the sketch-thinking ability.
• Figure 3: Qualitative analysis of our SketchThinker-R1. SketchThinker-R1 conducts a highly efficient yet effective sketch-style reasoning process. By focusing on key cues in problem-solving, our model arrives at the correct answer. The samples are from MathVision wang2024measuring.
• Figure 4: Examples of our SketchColdStart-20K data. Our sketch-style reasoning process focuses on key cues for solving questions. The thinking process effectively contributes to obtaining correct answers. At the same time, the thinking process is very concise, which significantly reduces the thinking cost.
• Figure 5: Qualitative case of SketchThinker-R1.