Research map / Large language models / Reasoning and chain of thought

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

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SketchThinker-R1, a three-stage pipeline (sketch-style SFT cold start, a SketchJudge LLM reward model, and GRPO with a style reward) that cuts reasoning tokens by over 64% on Qwen2.5-VL backbones while keeping accuracy on four multimodal benchmarks.

Problem

Long step-by-step reasoning in large multimodal models incurs high token cost and latency, and overthinking can hurt accuracy. The paper asks how to get concise, sketch-style reasoning without losing answer accuracy.

Approach

Convert long CoT traces into concise numbered sketch-style reasoning for SFT cold start, train an LLM judge (SketchJudge) to classify reasoning style, and use that judge's binary score as a reward term in GRPO alongside accuracy and format rewards.

Key results

• SketchThinker-R1-7B reaches 62.8 MMMU accuracy with 64.3 tokens versus Vanilla-R1's 61.0 accuracy and 182.2 tokens (EoT 0.977 vs 0.335).
• SketchThinker-R1-3B reduces tokens by more than 50% relative to Vanilla-R1-3B on MMMU while keeping comparable accuracy (55.9 vs 54.8).
• Ablation shows cold start alone yields 114.5 tokens and RL alone yields 152.2 tokens on MMMU, versus 64.3 for the combination.

Surprising findings

• Binary SketchJudge reward outperformed a dense floating-point judge reward on MMMU (62.8 vs 62.6 accuracy, 64.3 vs 65.4 tokens).
• RL directly from the base model without sketch cold start gave only marginal token reduction (152.2 tokens), showing cold start is necessary.

Limitations

• Reliance on GPT-5 for generating sketch-style training data, which is costly and closed-source.
• Efficiency gains are smaller for the 3B model, which already produces shorter chains.
• Interpretability evaluated with only 20 human samples and an LVLM judge rather than a broad human study.

Experimental setup

Datasets: LLaVA-CoT-100K Vision-R1-cold MMStar MathVista LogicVista SeePhys MMMU MathVision VisuLogic PhyX

Benchmarks: MMMU MathVision VisuLogic PhyX

Metrics: Acc. #Token EoT (Efficiency of Thinking)

Baselines: Vanilla-R1 Constrained CoT Chain-of-Draft C3oT VeriThinker L1 ThinkPrune

Builds on: GRPO Easy-R1 LLaMA-Factory Vision-R1 LLaVA-CoT

Closest papers by meaning

• Visual Sketchpad: Sketching as a Visual Chain of Thought for Multimodal Language Models (NeurIPS 2024 · similarity 0.89)
• Reinforcing Spatial Reasoning in Vision-Language Models with Interwoven Thinking and Visual Drawing (NeurIPS 2025 · similarity 0.88)
• GThinker: Towards General Multimodal Reasoning via Cue-Guided Rethinking (CVPR 2026 · similarity 0.88)
• Zebra-CoT: A Dataset for Interleaved Vision-Language Reasoning (ICLR 2026 · similarity 0.88)
• Open Vision Reasoner: Transferring Linguistic Cognitive Behavior for Visual Reasoning (NeurIPS 2025 · similarity 0.88)
• AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.87)