Thinking with images empowers vision-language models (VLMs) to use an external Zoom-in tool for fine-grained visual search. Recent works employ supervised fine-tuning (SFT) and reinforcement learning (RL) with tool-specific designs to enable such capabilities. However, achieving logical reasoning alongside accurate and efficient tool use remains a significant challenge. To address this, we shift the focus away from increasing RL complexity and instead identify SFT data quality as a key bottleneck. We then propose LogT (short for Logically Think with Images), a fully automated SFT data engine that explicitly enforces logic-chain-guided reasoning. LogT generates an SFT dataset featuring challenging, visually grounded questions and structured tool-use traces with human-like logic. We validate our approach by fine-tuning two base models, Qwen2.5-VL-7B and InternVL3.5-8B, on the dataset and evaluating them across seven benchmark test sets. We show that LogT-SFT based on Qwen2.5-VL-7B outperforms the state-of-the-art trained with both SFT and RL, despite using 3.1× less data and >120× fewer GPU hours. Furthermore, applying a simple RL recipe without tool-specific design on LogT-SFT yields a model that surpasses prior methods with complex RL setups by a substantial 9.59-point margin. In addition to strong performance gains, our model exhibits higher tool-use accuracy and efficiency, as well as more structured and logically consistent reasoning traces, directly validating the effectiveness of our approach.
Figures & tables
Figure 1
Figure 3: A detailed overview of LogT SFT data synthesis engine. Part 1 (top; Section 3.1 ) synthesizes visual search QA pairs using the full image ( I0 ) along with rich visual contexts ( IC and IF , top left) related to randomly selected target objects. After generating candidate QAs, we apply a simple heuristic filter and use a VLM-as-a-Judge to ensure quality. Part 2 (bottom; Section 3.2 ) generates reasoning traces for the selected QA. For easy questions, we directly generate traces without tool calls (omitted here). For other questions, we first create logic chains ( LM and LE ) based on two human-like strategies. A logic chain selector then chooses the more appropriate one, which is subsequently used to guide the synthesis of multi-turn reasoning traces with tool use.
Figure 3
Figure 6: Tool call efficiency. We report the average number of tool calls per question across the seven test sets. Despite achieving better performance compared to prior SOTA, LogT -SFT model uses 2.7 × fewer tool calls, and LogT -RL uses 2.3 × fewer.
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: System prompt for the QA generator. The remaining rules and in-context examples are omitted for brevity.
Figure 8: System prompt for the reasoning trace generator. Each logic-chain entry of the form ⟨short noun phrase⟩(k) specifies a target object or region and its source image Ik . Here, I0 is the original image, and Ik for k≥1 is the crop returned by the k -th Zoom-in call. The reasoning trace generator locates the specified target in Ik , predicts its bounding box, and invokes Zoom-in on that image using the predicted box. For example, “Backpack (2)” corresponds to “Backpack ( I2 )” in Figure 3 and instructs the generator to zoom in on the backpack within the person crop returned by the second tool call. The remaining rules and in-context examples are omitted for brevity.
Figure 9: An example of an exhaustive search logic chain and its corresponding tool-use reasoning trace. To locate the pink sign held by a person in the crowd, the logic chain begins by instructing the trace generator (see Section 3.2 ) to identify the crowd, then systematically search it from left to right. Since the trace generator successfully detects the target sign and the person in the crop returned by the third Zoom-in call, it stops scanning the rest of the crowd and instead zooms in on the identified region to extract the final answer.
Figure 8
Figure 12: Qualitative case study of SFT data synthesized by Pixel Reasoner , Chain-of-Focus , Mini-o3 and LogT .
Figure 13: Instructions for Human Evaluation.
Figure 14: Per-annotator breakdown of human evaluation. All three annotators preferred responses generated by LogT -RL over those from Mini-o3 -RL.
Split
V *
VisualProbe
Easy
Medium
Hard
Easier
85.34
59.57
35.07
35.85
Harder
87.43
58.87
34.70
40.57
Appendix
Table 4: Ablation on question difficulty. Training the VLM on the harder half of our SFT dataset leads to significantly improved performance on challenging tasks like those in VisualProbe-Hard.
VisualProbe
HR-Bench
Training Recipe
Logic Chain
V *
Easy
Medium
Hard
4K
8K
MME- RW-Lite
Avg.
✗
87.43
59.57
38.06
39.62
72.25
69.50
44.09
58.65
SFT
✓
87.43
60.99
41.04
39.62
72.00
70.12
45.96
59.59
✗
90.05
66.67
46.27
46.23
74.75
75.12
51.80
64.41
SFT+RL
✓
92.15
70.92
52.24
51.89
76.12
74.62
53.00
67.28
Appendix
Table 5: Ablation on logic chain-driven reasoning trace synthesis. Incorporating logical chains into the data generation process of LogT improves overall SFT performance and yields substantial additional gains after subsequent RL.
Model Usage
Avg.
Original
59.55
Qwen3-VL-8B, Qwen2.5-VL-32B & GPT-OSS-20B
59.43
Appendix
Table 6: Ablation on models used for data synthesis. We replace the models used by LogT in Section 3 with much smaller variants and re-synthesize 4k SFT samples. After SFT, the average performance across V * , VisualProbe, and HR-Bench remains nearly unchanged, indicating that our method is robust to the choice of synthesis models.
Stage
Error Rate (%)
QA Generation & Filtering
12
Logic Chain Generation
2
Reasoning Trace Generation
7
Appendix
Table 7: Error rates across different stages. We randomly sampled 100 samples from each stage and manually evaluated their error rates.
Model
V * -Rotated
Qwen2.5-VL-7B
46.07
LogT -RL (with Rotation Tool)
68.06
Δ vs. Base Model
+21.99
Appendix
Table 8: Performance on V * -Rotated. We construct V * -Rotated by rotating images in V * by angles randomly sampled from 30∘ to 330∘ . LogT -RL uses a rotation tool at inference time that was not seen during training, improving over the base model by 21.99 points.
Figure 15: Qualitative case study comparing model responses to a V * question between baseline models and LogT -RL. We show the generated reasoning traces, including Zoom-in calls, for both baseline models and LogT -RL. While Qwen2.5-VL-7B-RL, Pixel Reasoner -RL, and Chain-of-Focus -RL produce the correct final answer, their responses contain hallucinations or self-contradictions. The remaining baselines make erroneous Zoom-in calls and ultimately produce incorrect answers. In contrast, LogT -RL follows a coherent and consistent reasoning chain: it first identifies the front cabin of the truck, then locates the tissue box, and finally answers the question correctly without hallucinations.
Tool use in vision-language models (VLMs) studies how models call visual, computational, and retrieval tools to improve grounded perception and reasoning.
Thinking with images allows a multimodal model to compensate for limited perception by invoking visual tools through code. Yet the prevailing SFT-then-RL recipe creates a different supervision misalignment at each stage. SFT is expected to teach how to use tools, but trajectories from stronger teachers may succeed through perceptual capabilities that a smaller student cannot reliably reproduce or exploit, causing the student to imitate tool-call patterns without learning how to make them useful. RL is expected to teach when to use tools, but outcome-only rewards make fallible tool execution a liability and suppress tool use, whereas a blanket bonus for every correct tool-using trajectory encourages valid but ineffective operations. To address these two misalignments, we introduce ToolVision. During SFT, a multi-agent pipeline explores candidate trajectories, and a committee including student-scale models scores stepwise evidence gain to rank and prune the search branches. Only successfully executed trajectories with correct final answers are retained for SFT. Before RL, ToolVision compares the learner's performance with and without tools, then rewards successful tool use only on questions where tools provide a clear benefit. Both signals are constructed automatically from public task data without additional human annotations of tool use or necessity. ToolVision-8B improves over its base on all seven main benchmarks, surpasses Thyme-7B, CodeVision-8B, and CodeDance-7B on all three high-resolution benchmarks, and outperforms Qwen3-VL-32B-Thinking on V* and HRBench 8K. We will publicly release the datasets and source code.
Delin Mao, Chenghao Sun, Jingwei Song +2
Shanghai Jiao Tong University · University of Chinese Academy of Sciences · Fudan University
Vision language models (VLMs) are expected to perform effective multimodal reasoning and make logically coherent decisions, which is critical to tasks such as diagram understanding and spatial problem solving. However, current VLM reasoning lacks large-scale and well-structured training datasets. To bridge this gap, we propose VisualSphinx, a first-of-its-kind large-scale synthetic visual logical reasoning training dataset. To tackle the challenge of image synthesis with grounded answers, we propose a rule-to-image synthesis pipeline that extracts and expands puzzle rules from seed questions and generates executable rendering programs for puzzle assembly. Using standard GRPO with binary correctness rewards as a validation recipe, experiments show that VLMs trained on VisualSphinx benefit from the logical coherence and readability of our dataset and exhibit improved performance on logical reasoning tasks. The enhanced reasoning capabilities developed from VisualSphinx also benefit other reasoning tasks such as algebraic reasoning, arithmetic reasoning and geometry reasoning.
Yichen Feng, Zhangchen Xu, Fengqing Jiang +5
University of Washington · Western Washington University
Large Vision-Language Models (LVLMs) have shown strong promise for multimodal reasoning, yet often struggle with tasks requiring concepts beyond what is directly observable in the input image. Existing methods generate intermediate images or latent visual tokens to guide reasoning, but these representations can introduce errors and increasingly interfere with textual reasoning as reasoning progresses. We propose Visual-to-Text Chain-of-Thought Distillation (V2T), a framework that enables LVLMs to internalize visual reasoning without generating intermediate visual representations at inference time. V2T first trains a teacher LVLM using interleaved visual and textual chains of thought, and then uses knowledge distillation to train a student LVLM using the teacher's logits and cross-entropy supervision from ground-truth textual reasoning. When reasoning images can be mapped to the original image, V2T can additionally distill the teacher's attention to corresponding regions, while ground-truth bounding boxes can further guide a subsequent reinforcement learning stage. Experiments across multiple multimodal reasoning benchmarks show that V2T consistently outperforms the teacher and existing baselines, improving average accuracy by 14.3% on a held-out set and 2.7% on the broader visual evaluation suite. Moreover, lightweight SFT and substantially reduced RL make V2T up to 42x faster to train than state-of-the-art baselines.
Wenhan Yang, Nilay Naharas, Ali Payani +1
University of California, Los Angeles · Cisco Research