SketchThinker-R1: Efficient Sketch-Style Reasoning

• The paper introduces a three-stage pipeline—sketch-mode cold start, supervised fine-tuning, and reinforcement learning—to achieve efficient, concise reasoning in LMMs.
• It reduces reasoning token consumption by over 64%, improving Efficiency of Thinking (EoT) 2–6× compared to traditional chain-of-thought approaches.
• Empirical results across benchmarks demonstrate maintained accuracy with dramatic token compression, paving the way for edge device deployment and cost-effective inference.

SketchThinker-R1 is a training and inference methodology for large multimodal models (LMMs) that enables efficient sketch-style reasoning. Unlike traditional chain-of-thought (CoT) approaches that produce verbose reasoning traces, SketchThinker-R1 prioritizes concise, goal-directed enumerations of logical steps. Through a three-stage pipeline—comprising sketch-style supervised fine-tuning, reward modeling, and reinforcement learning—SketchThinker-R1 achieves over 64% reduction in reasoning token consumption without loss of accuracy. By focusing on saliency and outline brevity, the method aligns computational reasoning in LMMs with the efficient, outline-driven heuristics observed in human problem solving (Zhang et al., 6 Jan 2026).

1. Motivation for Sketch-Style Reasoning

CoT prompting and SFT typically yield correct answers but induce excessively long reasoning traces (hundreds of tokens). This redundancy results in increased API/hardware billing, slower inference, and higher likelihood of error propagation due to irrelevant or misleading substeps (Zhang et al., 6 Jan 2026). Empirical analysis shows that many CoT steps are not logically necessary for final answer derivation, and excessive detail may degrade performance (cf. Cuadron et al., 2025). Human experts, in contrast, often employ sketch-style reasoning—succinct, numbered outlines targeting only the key logical operations—allowing rapid yet accurate problem resolution (Zhang et al., 6 Jan 2026). This suggests that sketch-style reasoning optimizes both cognitive and computational efficiency.

2. Pipeline Architecture and Training Procedure

SketchThinker-R1 formalizes sketch-style reasoning through a three-stage pipeline:

1. Sketch-Mode Cold Start: Standard multimodal reasoning datasets (e.g., LLaVA-CoT-100K, Vision-R1-cold) are converted to sketch-style outlines via an LLM prompt (e.g., GPT-5). This prompt strips fluff and extraneous detail, extracts salient steps, and outputs a numbered sketch trace $T_\text{sketch}$.
2. Supervised Fine-Tuning: The base vision-LLM $\pi_\theta$ is fine-tuned on $(I, Q) \rightarrow T_\text{sketch}$ samples via cross-entropy minimization:

$$\mathcal{L}_{\mathrm{cold}} = -\frac{1}{N}\sum_{i=1}^N \sum_{t=1}^{T_i}\log\pi_\theta(o_{i,t}|o_{i,<t},q_i)$$

Implementation: LoRA rank=8, batch=16, grad_acc=2, lr=1e-5, warmup=0.1, epochs=10.

1. SketchJudge Reward Model: A 7B-parameter LLM (Qwen2.5-7B-Instruct) is fine-tuned to distinguish long CoT traces (label 0) from sketch-style traces (label 1) using paired data (40K sequences from SketchColdStart-20K).
2. Sketch-Thinking Reinforcement Learning: The cold-start model undergoes further RL optimization supervised by SketchJudge, using Group-Reward Proximal Optimization (GRPO, cf. Shao et al., 2024) with the surrogate objective:

$$J_{\text{GRPO}}(\theta) = \mathbb{E}\left[\min(r_t(\theta)\hat A_t,\,\mathrm{clip}(r_t(\theta),1-\epsilon,1+\epsilon)\hat A_t)\right]-\beta D_\text{KL}[\pi_\theta\,||\,\pi_\text{ref}]$$

Reward for each rollout:

$$R = 0.5\,R_\text{accuracy} + 0.4\,R_\text{format} + 0.1\,r_\text{style}$$

Implementation: rollout_batch=128, KL coeff=0.01, AdamW lr=1e-6, weight_decay=1e-2, epochs=15.

3. Reasoning Trace Compression and Efficiency Metrics

SketchThinker-R1 reduces the average number of reasoning tokens by more than 64%, with sample traces compressed from ~200 to ~65 tokens. Efficiency of Thinking (EoT), defined by $\text{EoT} = \text{Acc} / \#\text{Token}$, is consistently 2–6× higher than conventional R1 training. Main results on Qwen2.5-VL-7B backbone across four multimodal benchmarks are summarized below:

Benchmark	Vanilla-R1 Token Len	SketchThinker-R1 Token Len	Acc (Vanilla)	Acc (Sketch)	EoT (Vanilla)	EoT (Sketch)
MMMU	182.2	64.3	61.0%	62.8%	0.335	0.977
MathVision	221.1	65.5	31.0%	31.7%	0.140	0.484
VisuLogic	240.0	56.3	27.6%	27.8%	0.115	0.494
PhyX	225.1	75.3	46.7%	48.6%	0.207	0.645

Table: Comparative token consumption and performance metrics for SketchThinker-R1 and baseline R1 training (Zhang et al., 6 Jan 2026).

A plausible implication is that sketch-style reasoning may also reduce memory footprint and support deployment on edge devices or resource-constrained platforms, though such claims would require empirical measurement.

4. Qualitative Behavior and Interpretability

Analysis of output traces reveals that SketchThinker-R1 produces numbered lists emphasizing only the logical operations necessary for solution derivation. For example, in MathVision, the model’s output:

1. AB=3+2, BC=2+1, AC=3+1
2. Sides 5,3,4 → right
3. Area = ½·3·4=6

captures all essential deductions (distances, triangle classification, area computation) with minimal extraneous commentary. Across domains, hypotheses produced by SketchThinker-R1 are less verbose, yet retain clear logical lineage between input question and answer (Zhang et al., 6 Jan 2026).

5. Comparative Performance and Benchmarking

Benchmarked against prompt-based (Constrained CoT, Chain-of-Draft), SFT-based (C3oT, VeriThinker), and RL-based (L1, ThinkPrune) reductions, SketchThinker-R1 either matches or slightly surpasses accuracy while achieving much greater brevity. Other methods tend to trade brevity for accuracy or yield modest token savings only. The binary reward signal from SketchJudge allows stable RL training, and format enforcement maintains interpretability. The approach is orthogonal to latent-CoT or transformer pruning methods, and could be synergistic if integrated with them (Zhang et al., 6 Jan 2026).

6. Implementation Details and Technical Components

• Model Base: Qwen2.5-VL-7B backbone.
• SFT Hyperparameters: LoRA rank=8, batch=16, grad_acc=2, lr=1e-5, 10 epochs.
• RL Config: rollout_batch=128, AdamW lr=1e-6, 15 epochs.
• Reward Model: Qwen2.5-7B-Instruct fine-tuned on paired traces.

Training is performed using supervised fine-tuning followed by RL. The reward signal combines final-answer correctness, format adherence, and sketch-style discrimination (Zhang et al., 6 Jan 2026).

7. Future Directions and Limitations

Current limitations include reliance on high-quality sketch-style data (cold-started via LLM conversion), binary sketch-style reward signal (threshold effects), and focus on text+image modalities. Extensions proposed involve applying sketch reasoning to audio and video, incorporating graded style scores, and integrating sketch-style reasoning data into large-scale pretraining for upfront computational savings.

SketchThinker-R1 provides a practical framework for intentional style shaping in LMM reasoning, yielding dramatic efficiency gains with no trade-off in accuracy (Zhang et al., 6 Jan 2026).