Pith. sign in

REVIEW 4 major objections 10 minor 2 cited by

GThinker: Towards General Multimodal Reasoning via Cue-Guided Rethinking

T0 review · 4 major / 10 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A 7B model trained to re-check visual cues outranks a frontier reasoning model on M3CoT.

desk verdict Cue-rethinking is a genuinely fresh idea, but the M3CoT headline is not yet credible because the training data may overlap with the test set and the O4-mini comparison is not controlled. read the letter →

arxiv 2506.01078 v1 pith:7VKRI222 submitted 2025-06-01 cs.CV cs.AI

Yufei Zhan , Ziheng Wu , Yousong Zhu , Rongkun Xue , Ruipu Luo , Zhenghao Chen , Can Zhang , Yifan Li
show 5 more authors
Zhentao He Zheming Yang Ming Tang Minghui Qiu Jinqiao Wang
This is my paper · ORCID
classification cs.CVcs.AI
keywords multimodalreasoningvisualcueschain-of-thoughtreinforcementlearninggeneralscenarioM3CoTbenchmarkQwen2.5-VLCue-Rethinking
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

GThinker claims that the main obstacle to general visual reasoning is not knowledge or logic but grounding in visual cues, and that current slow-thinking models reason too abstractly to catch their own perceptual mistakes. The paper introduces a 'Cue-Rethinking' pattern in which the model tags visual cues during free-form reasoning, then deliberately re-examines and revises those cues before answering. Trained with a two-stage pipeline on its own GThinker-11K data, a 7B model reaches 81.5% on M3CoT, above O4-mini, while keeping math performance on par. If the claim holds, a small open-source model can rival closed-source reasoning systems by adding a targeted reflective step rather than by scaling parameters.

What carries the argument

Cue-Rethinking Pattern: a free-form thinking protocol where the model tags visual cues with <vcues_*> markers during initial reasoning, then is prompted to re-inspect and revise those cues before concluding. It is the load-bearing object because it is what the cold-start SFT data, the formatting selection, and the RL reward design all conspire to teach the model to perform.

What would settle it

Search the released GThinker-11K training data for exact or near-duplicate question-image-answer instances that also appear in the M3CoT evaluation set. If any evaluation items are found in the training data, the 81.5% result is inflated and the O4-mini comparison is invalid. A second falsifier: apply the Cue-Rethinking prompt at inference time to the base Qwen2.5-VL-7B without any training; if zero-shot Cue-Rethinking does not produce a large gain over the base model, the headline improvement depends on the trained cold-start stage, not on the pattern alone.

Watch

Extended reading notes

Core claim

A two-stage training recipe built around the Cue-Rethinking pattern endows Qwen2.5-VL-7B with multimodal reasoning that beats O4-mini on M3CoT (81.5% vs. 80.9%). The pattern lets the model reason freely while tagging visual cues with <vcues_*> markers, then triggers a rethinking phase where it checks those cues against the image and corrects flawed or missing cues before finalizing the answer. The paper attributes the gain to a pattern-guided cold start on 7,358 iteratively annotated reasoning paths, followed by incentive reinforcement learning (DAPO) on 4K curated samples.

Load-bearing premise

The paper never explicitly states that the 7,358 cold-start samples and the 4K RL samples were drawn from training splits, so the entire M3CoT result rests on the assumption that no training or near-duplicate samples leaked into the M3CoT evaluation set.

Editorial extensions

If this is right

  • The cold-start stage that teaches Cue-Rethinking contributes about 6.9 points on M3CoT, and pattern-guided selective formatting adds another 5.2 points.
  • GThinker-7B improves 2.5 points on MMStar and 1.6 points on RealWorld QA, so the pattern transfers to general multimodal scenarios beyond M3CoT.
  • Math is not sacrificed: GThinker-7B reaches 72.7% on MathVista and 26.6% on MathVision, in line with or above other RL-enhanced reasoning models.
  • MMMU-Pro improves about 4 points, indicating the pattern also helps multidisciplinary science reasoning.
  • The full pipeline trains in about 9 hours on 32 H100 GPUs, so the recipe is cheap enough to reproduce and extend.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The Cue-Rethinking prompt could be applied at test time to an existing MLLM without any fine-tuning, simply asking the model to tag and re-check its visual cues; if that works, the gains may not be tied to the specific training pipeline.
  • The authors leave open whether rethinking helps because the model literally re-reads pixels or because the prompt encourages more token spending on hard questions; a token-controlled ablation would separate these explanations.
  • If the 4K RL samples overlap with the M3CoT evaluation set, the headline number could be inflated; the paper never states the training/evaluation split explicitly.
  • A natural extension is to apply the same cue-tagging plus rethinking pattern to a larger base model (e.g., a 32B or 72B MLLM), which should push M3CoT scores well beyond closed-source leaders if the mechanism is additive to scale.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 10 minor

Summary. The paper introduces GThinker, a 7B multimodal reasoning model built on Qwen2.5-VL-7B, and proposes a Cue-Rethinking reasoning pattern in which the model tags visual cues during free-form reasoning and then revisits them before answering. Training proceeds in two stages: a pattern-guided cold start on 7,358 annotated reasoning paths and an incentive reinforcement learning stage (DAPO) on 4K curated samples. The authors construct the GThinker-11K dataset via iterative annotation with GPT-4o, O1, and O3, and report strong results on M3CoT (81.5%, 'surpassing' O4-mini), plus improvements on MMStar, RealWorldQA, MMMU-Pro, and MathVista.

Significance. If the empirical claims survive controlled evaluation, this is a useful contribution: it extends slow-thinking reasoning beyond math/science into general multimodal scenarios, addresses visual-cue grounding with a simple trainable pattern, and shows that a 7B open model can reach state-of-the-art on a multi-domain benchmark. The two-stage pipeline with hybrid rewards and the detailed data-construction appendix are concrete and reproducible in spirit. However, the headline result depends on the M3CoT evaluation, and the current manuscript does not demonstrate that the training data are disjoint from the M3CoT test set, nor that the comparison to O4-mini and other baselines is controlled. These issues are load-bearing for the central claims.

major comments (4)
  1. [Appendix A.1, Table 5, Section 4.1] The training data are built from M3CoT itself, yet the paper never states that the cold-start 7,358 samples and the 4K RL samples are disjoint from the M3CoT evaluation set. Appendix A.1 says M3CoT is the 'baseline' and Table 5 lists M3CoT as a source for Science (5,266 samples) and Commonsense (1,471 questions, 'generated by M3CoT'); Appendix A.6 describes RL data curated from sources that themselves are M3CoT-derived. Since M3CoT has a public test split, any overlap or near-duplicate between GThinker-11K and the M3CoT evaluation set directly inflates the 81.5% headline. The authors must specify exactly which split of M3CoT was used, report a near-duplicate analysis between the training data and the evaluation set, and, if any overlap exists, re-evaluate on the non-overlapping subset or a separate held-out benchmark.
  2. [Section 4.1, Table 1] The claim of 'surpassing O4-mini' is not a controlled comparison. The O4-mini M3CoT score (80.9%) is taken from the O3/O4-mini system card, while GThinker is evaluated with VLMEvalKit under 'each model's official settings and prompts.' Different harnesses, prompt formats, decoding parameters, and answer extraction procedures can easily account for a 0.6% difference. The authors need to reproduce O4-mini (and ideally other closed-source models like Gemini-2.5 Pro and O3) under the same VLMEvalKit harness with identical prompts and decoding settings, or explicitly label the comparison as cross-pipeline and not a valid benchmark result.
  3. [Section 4.1, Section 4.2] The evaluation protocol for open-source baselines on M3CoT is under-specified. The manuscript reports a single accuracy value per model and states only that 'we follow each model's official settings and prompts and use VLMEvalKit.' It does not provide the exact prompts, temperature, max tokens, number of runs, or whether chain-of-thought is enabled for the reasoning baselines (e.g., LLaVA-CoT, MM-Eureka, R1-OneVision, VLAA-Thinker). These models are known to be sensitive to such settings, and without a detailed evaluation configuration the reported margins (e.g., +13.1 over VLAA-Thinker, +14.2 over R1-OneVision) are not reproducible. The full evaluation configuration should be included in an appendix or supplement.
  4. [Section 4.1, Table 2] For MMStar, RealWorldQA, MMMU-Pro, MathVista, and MathVision, the paper uses 'the results reported in their original papers' for most open-source baselines while evaluating GThinker under its own protocol. This mixes evaluation setups across papers (different few-shot settings, different answer extraction, possibly different prompt templates) and undermines the claim of a '+2.1% average improvement on general scenario multimodal reasoning benchmarks.' The authors should either re-evaluate all open-source baselines under a single unified protocol or provide a detailed comparison table showing which settings were used for each model, so that the claimed average improvement is meaningful.
minor comments (10)
  1. [Abstract / Section 4.2] The abstract says GThinker is 'surpassing the latest O4-mini model' while Section 4.2 says 'performing on par with the latest reasoning model O4-mini'; these are inconsistent and should be reconciled.
  2. [Section 4.3, Table 4] The sentence 'The final row (with all modules removed) corresponds to training with the same QA pairs but without any of our proposed methods' is confusing because the row is labeled 'Qwen2.5-VL-7B-Zero,' which suggests zero-shot evaluation, not training. Please clarify what the last row actually represents.
  3. [Section 4.3, Table 4] The phrase 'using the Cue-Rethinking Pattern for Pattern-Guided Cold Start—without Pattern-Guided Selection (PGS) Formatting—yields a 6.9% average improvement' is ambiguous about the baseline; specifying that the improvement is relative to the '- PG Cold Start' row would make the incremental contributions clearer.
  4. [Appendix A.1] The text says 'we verify the remaining samples’ compliance with closed-source model usage policies using GPT-4o'; it would be helpful to state which policies were checked and whether the annotation pipeline (which also uses O1 and O3) complies with the respective terms of service.
  5. [Appendix A.6] The description of the RL data curation says data are sampled from sources [32, 56, 57] with 'less overlap with the previous cold start data by comparison,' but it does not specify how overlap was measured or whether the clustering algorithm used any M3CoT test information. Please provide the overlap detection method and the exact source datasets with their splits.
  6. [Section 3.2] The pattern-guided selection process is described only verbally: samples with 'flawed visual cues' are selected for full Cue-Rethinking sequences, but the criterion for determining 'flawed' from the base model's reasoning is not operationalized. A brief algorithmic description or pseudocode would improve reproducibility.
  7. [Section 4.1, paragraph 2] The sentence 'For other benchmarks, we use the results reported in their original papers' mixes evaluation protocols; even if the authors choose not to re-run all baselines, they should at least list the reported values with their original papers and settings to avoid implying a uniform protocol.
  8. [Table 1] The GThinker-7B row has a typographical issue in the Science/Soc. column ('68.982.2' appears without a separator); also, the Science/Nat. column for GThinker is 90.7, which should be double-checked for formatting.
  9. [References] Reference [48] and [49] are the same paper (mixed preference optimization); one duplicate should be removed.
  10. [Introduction] The phrase 'GThinker-11k, compromised 7K cold-start data' should read 'comprising 7K cold-start data.'

Circularity Check

1 steps flagged · score 6.0 of 10

M3CoT is both a training-data source and the headline evaluation benchmark with no stated split, so the 81.5% 'prediction' is not shown to be out-of-sample.

  1. fitted input called prediction [Appendix A.1, Table 5, and §4.1 Evaluation Settings; Table 1]
    "Building on baseline, we apply a two-step filtering process to ensure data quality: ... resulting in 7,358 high-quality samples. ... Table 5: Science 5266 KiloGram[20], ScienceQA [28], M3CoT [7]; ... Commonsense 1471 Sherlock [15](Questions generated by M3CoT). ... M3CoT: A challenging benchmark that spans science, commonsense, and math domains ... We primarily use this benchmark to comprehensively evaluate models' multimodal reasoning capabilities across diverse scenarios."

    The cold-start SFT data is explicitly sourced from the same M3CoT benchmark that Table 1 uses for the headline 81.5% result and the O4-mini comparison. The paper never states that the 7,358 samples are drawn from an M3CoT training split or that M3CoT evaluation questions were excluded from filtering. If any of the 5,266 science samples or 1,471 commonsense samples (described as 'Questions generated by M3CoT') coincide with or are near-duplicates of M3CoT evaluation items, then the reported M3CoT accuracy is a fitted-input retrieval result rather than an out-of-sample prediction. The evaluation section also provides no contamination check or holdout statement.

full rationale

The strongest circularity is benchmark contamination: Appendix A.1 and Table 5 make M3CoT a training source for 5,266 science and 1,471 commonsense cold-start samples, while §4.1 and Table 1 use M3CoT as the primary evaluation benchmark for the 81.5% headline. No train/test split, exclusion rule, or near-duplicate check is stated anywhere in the data-construction or evaluation sections, so the paper's own text permits the evaluation to be an input to training by construction. This makes the central 'prediction' possibly a fitted-input result. The rest of the paper is not circular: Cue-Rethinking is a new method claim, the DAPO-based RL stage is implemented from external work, and the MMStar, RealWorldQA, MathVista, and MMMU-Pro results provide independent (though secondary) evaluation signal. I found no load-bearing self-citation chain, no uniqueness theorem imported from the authors' prior work, and no ansatz smuggled in via citation. The score of 6 reflects one load-bearing benchmark-contamination circularity affecting the headline claim, while acknowledging that independent benchmarks and method ablations give the paper partial non-circular content.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No numeric free parameters are fitted to a target result; the training hyperparameters are standard. The central claim depends on the quality of the annotated dataset and on benchmark validity, which are recorded as axioms. No invented entities are introduced; the visual cue tags are a formatting convention.

assumptions (3)
  • domain assumption M3CoT is a valid and comprehensive measure of multimodal reasoning across science, commonsense, and math.
    The paper uses M3CoT as its primary evaluation benchmark and as a source of training data (Appendix A.1). If the benchmark does not reliably measure reasoning, the central claim loses its basis.
  • domain assumption The closed-source models used for annotation (GPT-4o, O1, O3, Gemini 2.5 Pro) produce correct reasoning paths and accurate visual cue corrections.
    The 7K cold-start dataset is generated and verified through these models (Appendix A.2, A.5). Flawed annotations would teach the model artifacts instead of genuine visual-cue reasoning.
  • domain assumption The Cue-Rethinking pattern, learned from the constructed data, generalizes to new questions and domains beyond the training distribution.
    This is the paper's core assumption, tested by ablations on M3CoT only. The generalization to other benchmarks is not separately ablated, so the claim rests on this assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GThinker: Towards General Multimodal Reasoning via Cue-Guided Rethinking." pith.science (2026). https://pith.science/paper/7VKRI222

@misc{pith2026250601078,
  author       = {Pith},
  title        = {Pith review of: GThinker: Towards General Multimodal Reasoning via Cue-Guided Rethinking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7VKRI222}},
  note         = {Machine review of arXiv:2506.01078}
}
abstract

Despite notable advancements in multimodal reasoning, leading Multimodal Large Language Models (MLLMs) still underperform on vision-centric multimodal reasoning tasks in general scenarios. This shortfall stems from their predominant reliance on logic- and knowledge-based slow thinking strategies, while effective for domains like math and science, fail to integrate visual information effectively during reasoning. Consequently, these models often fail to adequately ground visual cues, resulting in suboptimal performance in tasks that require multiple plausible visual interpretations and inferences. To address this, we present GThinker (General Thinker), a novel reasoning MLLM excelling in multimodal reasoning across general scenarios, mathematics, and science. GThinker introduces Cue-Rethinking, a flexible reasoning pattern that grounds inferences in visual cues and iteratively reinterprets these cues to resolve inconsistencies. Building on this pattern, we further propose a two-stage training pipeline, including pattern-guided cold start and incentive reinforcement learning, designed to enable multimodal reasoning capabilities across domains. Furthermore, to support the training, we construct GThinker-11K, comprising 7K high-quality, iteratively-annotated reasoning paths and 4K curated reinforcement learning samples, filling the data gap toward general multimodal reasoning. Extensive experiments demonstrate that GThinker achieves 81.5% on the challenging comprehensive multimodal reasoning benchmark M$^3$CoT, surpassing the latest O4-mini model. It also shows an average improvement of 2.1% on general scenario multimodal reasoning benchmarks, while maintaining on-par performance in mathematical reasoning compared to counterpart advanced reasoning models. The code, model, and data will be released soon at https://github.com/jefferyZhan/GThinker.

Figures

Figures reproduced from arXiv: 2506.01078 by the authors.

Figure 1
Figure 1. Multimodal reasoning methods comparison across scenarios. Multimodal Reasoning [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overall pipeline for constructing GThinker. We collect multi-domain data covering general, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Toy example of the Cue-Rethinking Pattern. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Constructed Data Example with Cue Rethinking. The visual cues in red are flawed ones, [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Data pipeline for cold start data. A.1 Data Preparations Though several datasets are constructed to enhance multimodal reasoning capabilities in MLLMs [56–58] spanning diverse domains, they often present challenges such as high knowledge dependency, limited visual cues…
Figure 5
Figure 5. Figure 5: These checks target three critical aspects. First, for format validation, we ensure that for [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: Response sample from GThinker with visual cue-based rethinking enabled. [PITH_FULL_IMAGE:figures/full_fig_p022_6.png]
Figure 7
Figure 7. Figure 7: Response sample from GThinker with reflection for the math problem. [PITH_FULL_IMAGE:figures/full_fig_p023_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. SafeWork-R1: Coevolving Safety and Intelligence under the AI-45$^{\circ}$ Law

    cs.AI 2025-07 conditional novelty 6.0 of 10

    SafeWork-R1 shows that a staged RL pipeline with safety, value, and knowledge verifiers can improve both safety and general reasoning scores over a base multimodal model.

  2. Reinforcement Learning Meets Large Language Models: A Survey of Advancements and Applications Across the LLM Lifecycle

    cs.CL 2025-09 conditional novelty 3.0 of 10

    A survey that maps reinforcement learning methods, datasets, benchmarks, and open-source tools across the full training lifecycle of large language models, focusing on verifiable-reward reasoning.

Reference graph

Works this paper leans on

84 extracted references · 29 canonical work pages · cited by 2 Pith papers

  1. [1]

    Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  2. [2]

    Qwen2.5-vl technical report.arXiv preprint arXiv:2502.13923, 2025

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical report.ar...

    arXiv 2025

  3. [3]

    Perception tokens enhance visual reasoning in multimodal language models.arXiv preprint arXiv:2412.03548, 2024

    Mahtab Bigverdi, Zelun Luo, Cheng-Yu Hsieh, Ethan Shen, Dongping Chen, Linda G Shapiro, and Ranjay Krishna. Perception tokens enhance visual reasoning in multimodal language models.arXiv preprint arXiv:2412.03548, 2024

    arXiv 2024

  4. [4]

    Sft or rl? an early investigation into training r1-like reasoning large vision-language models.arXiv preprint arXiv:2504.11468, 2025

    Hardy Chen, Haoqin Tu, Fali Wang, Hui Liu, Xianfeng Tang, Xinya Du, Yuyin Zhou, and Cihang Xie. Sft or rl? an early investigation into training r1-like reasoning large vision-language models.arXiv preprint arXiv:2504.11468, 2025

    arXiv 2025

  5. [5]

    Are we on the right way for evaluating large vision-language models? InThe Thirty-eighth Annual Conference on Neural Information Processing Systems

    Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? InThe Thirty-eighth Annual Conference on Neural Information Processing Systems

  6. [6]

    R1-v: Reinforcing super generalization ability in vision-language models with less than $3

    Liang Chen, Lei Li, Haozhe Zhao, Yifan Song, and Vinci. R1-v: Reinforcing super generalization ability in vision-language models with less than $3. https://github.com/Deep-Agent/R1-V , 2025. Accessed: 2025-02-02

    2025

  7. [7]

    M3cot: A novel benchmark for multi-domain multi-step multi-modal chain-of-thought.arXiv preprint arXiv:2405.16473, 2024

    Qiguang Chen, Libo Qin, Jin Zhang, Zhi Chen, Xiao Xu, and Wanxiang Che. M3cot: A novel benchmark for multi-domain multi-step multi-modal chain-of-thought.arXiv preprint arXiv:2405.16473, 2024

    arXiv 2024

  8. [8]

    Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185–24198, 2024

    2024

Show all 84 references
  1. [9]

    Gemini 2.5 pro preview model card

    Google DeepMind. Gemini 2.5 pro preview model card. https://storage.googleapis.com/ model-cards/documents/gemini-2.5-pro-preview.pdf, 2025

    2025

  2. [10]

    Virgo: A preliminary exploration on reproducing o1-like mllm.arXiv preprint arXiv:2501.01904, 2025

    Yifan Du, Zikang Liu, Yifan Li, Wayne Xin Zhao, Yuqi Huo, Bingning Wang, Weipeng Chen, Zheng Liu, Zhongyuan Wang, and Ji-Rong Wen. Virgo: A preliminary exploration on reproducing o1-like mllm.arXiv preprint arXiv:2501.01904, 2025

    2025 arXiv

  3. [11]

    Vlmevalkit: An open-source toolkit for evaluating large multi-modality models

    Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, et al. Vlmevalkit: An open-source toolkit for evaluating large multi-modality models. InProceedings of the 32nd ACM International Conference on Multimedia,...

    2024

  4. [12]

    Cantor: Inspiring multimodal chain-of-thought of mllm

    Timin Gao, Peixian Chen, Mengdan Zhang, Chaoyou Fu, Yunhang Shen, Yan Zhang, Shengchuan Zhang, Xiawu Zheng, Xing Sun, Liujuan Cao, et al. Cantor: Inspiring multimodal chain-of-thought of mllm. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 9096–9105, 2024

    2024

  5. [13]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  6. [14]

    Measuring mathematical problem solving with the math dataset

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. InThirty-fifth Conference on Neural Information Processing Systems, 2021

    2021

  7. [15]

    The abduction of sherlock holmes: A dataset for visual abductive reasoning

    Jack Hessel, Jena D Hwang, Jae Sung Park, Rowan Zellers, Chandra Bhagavatula, Anna Rohrbach, Kate Saenko, and Yejin Choi. The abduction of sherlock holmes: A dataset for visual abductive reasoning. In European Conference on Computer Vision, pages 558–575. Springer, 2022

    2022

  8. [16]

    Vision-r1: Incentivizing reasoning capability in multimodal large language models.arXiv preprint arXiv:2503.06749, 2025

    Wenxuan Huang, Bohan Jia, Zijie Zhai, Shaosheng Cao, Zheyu Ye, Fei Zhao, Zhe Xu, Yao Hu, and Shaohui Lin. Vision-r1: Incentivizing reasoning capability in multimodal large language models.arXiv preprint arXiv:2503.06749, 2025. 11

    2025 arXiv

  9. [17]

    Gpt-4o system card.arXiv preprint arXiv:2410.21276, 2024

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card.arXiv preprint arXiv:2410.21276, 2024

    2024 arXiv

  10. [18]

    Math-Verify: Math Verification Library, 2023

    Hynek. Math-Verify: Math Verification Library, 2023. If you use this software, please cite it using the metadata from this file

    2023

  11. [19]

    Openai o1 system card.arXiv preprint arXiv:2412.16720, 2024

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Hel- yar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card.arXiv preprint arXiv:2412.16720, 2024

    2024 arXiv

  12. [20]

    Abstract visual reasoning with tangram shapes

    Anya Ji, Noriyuki Kojima, Noah Rush, Alane Suhr, Wai Keen V ong, Robert Hawkins, and Yoav Artzi. Abstract visual reasoning with tangram shapes. InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 582–601, 2022

    2022

  13. [21]

    Dcot: Dual chain-of-thought prompting for large multimodal models

    Zixi Jia, Jiqiang Liu, Hexiao Li, Qinghua Liu, and Hongbin Gao. Dcot: Dual chain-of-thought prompting for large multimodal models. InThe 16th Asian Conference on Machine Learning (Conference Track), 2024

    2024

  14. [22]

    Llava-onevision: Easy visual task transfer.arXiv preprint arXiv:2408.03326, 2024

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer.arXiv preprint arXiv:2408.03326, 2024

    2024 arXiv

  15. [23]

    Silkie: Preference distillation for large visual language models.arXiv preprint arXiv:2312.10665, 2023

    Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, and Lingpeng Kong. Silkie: Preference distillation for large visual language models.arXiv preprint arXiv:2312.10665, 2023

    2023 arXiv

  16. [24]

    V ocot: Unleashing visually grounded multi-step reasoning in large multi-modal models.arXiv preprint arXiv:2405.16919, 2024

    Zejun Li, Ruipu Luo, Jiwen Zhang, Minghui Qiu, Xuanjing Huang, and Zhongyu Wei. V ocot: Unleashing visually grounded multi-step reasoning in large multi-modal models.arXiv preprint arXiv:2405.16919, 2024

    2024 arXiv

  17. [25]

    Diving into self-evolving training for multimodal reasoning.arXiv preprint arXiv:2412.17451, 2024

    Wei Liu, Junlong Li, Xiwen Zhang, Fan Zhou, Yu Cheng, and Junxian He. Diving into self-evolving training for multimodal reasoning.arXiv preprint arXiv:2412.17451, 2024

    2024 arXiv

  18. [26]

    Chain-of-spot: Interactive reasoning improves large vision-language models.arXiv preprint arXiv:2403.12966, 2024

    Zuyan Liu, Yuhao Dong, Yongming Rao, Jie Zhou, and Jiwen Lu. Chain-of-spot: Interactive reasoning improves large vision-language models.arXiv preprint arXiv:2403.12966, 2024

    2024 arXiv

  19. [27]

    Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts

    Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. InThe Twelfth International Conference on Learning ...

  20. [28]

    Learn to explain: Multimodal reasoning via thought chains for science question answering.Advances in Neural Information Processing Systems, 35:2507–2521, 2022

    Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering.Advances in Neural Information Processing Systems, 35:2507–2521, 2022

    2022

  21. [29]

    Dynamic prompt learning via policy gradient for semi-structured mathematical reasoning

    Pan Lu, Liang Qiu, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, Tanmay Rajpurohit, Peter Clark, and Ashwin Kalyan. Dynamic prompt learning via policy gradient for semi-structured mathematical reasoning. InThe Eleventh International Conference on Learning Representations, 2022

    2022

  22. [30]

    Ovis: Structural embedding alignment for multimodal large language model.arXiv:2405.20797, 2024

    Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model.arXiv:2405.20797, 2024

    2024 arXiv

  23. [31]

    Textcot: Zoom in for enhanced multimodal text-rich image understanding.arXiv preprint arXiv:2404.09797, 2024

    Bozhi Luan, Hao Feng, Hong Chen, Yonghui Wang, Wengang Zhou, and Houqiang Li. Textcot: Zoom in for enhanced multimodal text-rich image understanding.arXiv preprint arXiv:2404.09797, 2024

    2024 arXiv

  24. [32]

    Mm-eureka: Exploring the frontiers of multimodal reasoning with rule-based reinforcement learning.arXiv preprint arXiv:2503.07365, 2025

    Fanqing Meng, Lingxiao Du, Zongkai Liu, Zhixiang Zhou, Quanfeng Lu, Daocheng Fu, Tiancheng Han, Botian Shi, Wenhai Wang, Junjun He, et al. Mm-eureka: Exploring the frontiers of multimodal reasoning with rule-based reinforcement learning.arXiv preprint arXiv:2503.07365, 2025

    2025 arXiv

  25. [33]

    Compositional chain-of-thought prompting for large multimodal models

    Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420–14431, 2024

    2024

  26. [34]

    O3 and o4-mini system card

    OpenAI. O3 and o4-mini system card. https://cdn.openai.com/pdf/ 2221c875-02dc-4789-800b-e7758f3722c1/o3-and-o4-mini-system-card.pdf , 2025. Ac- cessed: 2025-05-07

    2025

  27. [35]

    Skywork r1v: Pioneering multimodal reasoning with chain-of-thought.arXiv preprint arXiv:2504.05599, 2025

    Yi Peng, Xiaokun Wang, Yichen Wei, Jiangbo Pei, Weijie Qiu, Ai Jian, Yunzhuo Hao, Jiachun Pan, Tianyidan Xie, Li Ge, et al. Skywork r1v: Pioneering multimodal reasoning with chain-of-thought.arXiv preprint arXiv:2504.05599, 2025. 12

    2025 arXiv

  28. [36]

    O1 replication journey: A strategic progress report–part 1.arXiv preprint arXiv:2410.18982, 2024

    Yiwei Qin, Xuefeng Li, Haoyang Zou, Yixiu Liu, Shijie Xia, Zhen Huang, Yixin Ye, Weizhe Yuan, Hector Liu, Yuanzhi Li, et al. O1 replication journey: A strategic progress report–part 1.arXiv preprint arXiv:2410.18982, 2024

    2024 arXiv

  29. [37]

    Direct preference optimization: Your language model is secretly a reward model.Advances in Neural Information Processing Systems, 36:53728–53741, 2023

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model.Advances in Neural Information Processing Systems, 36:53728–53741, 2023

    2023

  30. [38]

    Rethinking reflection in pre-training.arXiv preprint arXiv:2504.04022, 2025

    Darsh J Shah, Peter Rushton, Somanshu Singla, Mohit Parmar, Kurt Smith, Yash Vanjani, Ashish Vaswani, Adarsh Chaluvaraju, Andrew Hojel, Andrew Ma, et al. Rethinking reflection in pre-training.arXiv preprint arXiv:2504.04022, 2025

    2025 arXiv

  31. [39]

    Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought reasoning.Advances in Neural Information Processing Systems, 3...

    2024

  32. [40]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

    2024 arXiv

  33. [42]

    Aligning large multimodal models with factually augmented rlhf.arXiv preprint arXiv:2309.14525, 2023

    Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang- Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf.arXiv preprint arXiv:2309.14525, 2023

    2023 arXiv

  34. [43]

    Kimi-vl technical report.arXiv preprint arXiv:2504.07491, 2025

    Kimi Team, Angang Du, Bohong Yin, Bowei Xing, Bowen Qu, Bowen Wang, Cheng Chen, Chenlin Zhang, Chenzhuang Du, Chu Wei, et al. Kimi-vl technical report.arXiv preprint arXiv:2504.07491, 2025

    2025 arXiv

  35. [44]

    Qwq-32b: Embracing the power of reinforcement learning, 2025

    Qwen Team. Qwq-32b: Embracing the power of reinforcement learning, 2025

    2025

  36. [45]

    Llamav-o1: Rethinking step-by-step visual reasoning in llms.arXiv preprint arXiv:2501.06186, 2025

    Omkar Thawakar, Dinura Dissanayake, Ketan More, Ritesh Thawkar, Ahmed Heakl, Noor Ahsan, Yuhao Li, Mohammed Zumri, Jean Lahoud, Rao Muhammad Anwer, et al. Llamav-o1: Rethinking step-by-step visual reasoning in llms.arXiv preprint arXiv:2501.06186, 2025

    2025 arXiv

  37. [46]

    Automatic data curation for self-supervised learning: A clustering-based approach.arXiv preprint arXiv:2405.15613, 2024

    Huy V V o, Vasil Khalidov, Timothée Darcet, Théo Moutakanni, Nikita Smetanin, Marc Szafraniec, Hugo Touvron, Camille Couprie, Maxime Oquab, Armand Joulin, et al. Automatic data curation for self-supervised learning: A clustering-based approach.arXiv preprint arXiv:2405.15613, 2024

    2024 arXiv

  38. [47]

    Measuring multimodal mathematical reasoning with math-vision dataset.Advances in Neural Information Processing Systems, 37:95095–95169, 2024

    Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and Hong- sheng Li. Measuring multimodal mathematical reasoning with math-vision dataset.Advances in Neural Information Processing Systems, 37:95095–95169, 2024

    2024

  39. [49]

    Enhancing the reasoning ability of multimodal large language models via mixed preference optimization.arXiv preprint arXiv:2411.10442, 2024

    Weiyun Wang, Zhe Chen, Wenhai Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Jinguo Zhu, Xizhou Zhu, Lewei Lu, Yu Qiao, et al. Enhancing the reasoning ability of multimodal large language models via mixed preference optimization.arXiv preprint arXiv:2411.10442, 2024

    2024 arXiv

  40. [50]

    Visualprm: An effective process reward model for multimodal reasoning.arXiv preprint arXiv:2503.10291, 2025

    Weiyun Wang, Zhangwei Gao, Lianjie Chen, Zhe Chen, Jinguo Zhu, Xiangyu Zhao, Yangzhou Liu, Yue Cao, Shenglong Ye, Xizhou Zhu, et al. Visualprm: An effective process reward model for multimodal reasoning.arXiv preprint arXiv:2503.10291, 2025

    2025 arXiv

  41. [51]

    Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    2022

  42. [52]

    V?: Guided visual search as a core mechanism in multimodal llms

    Penghao Wu and Saining Xie. V?: Guided visual search as a core mechanism in multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13084– 13094, 2024

    2024

  43. [53]

    Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding.arXiv preprint arXiv:2412.10302, 2024

    Zhiyu Wu, Xiaokang Chen, Zizheng Pan, Xingchao Liu, Wen Liu, Damai Dai, Huazuo Gao, Yiyang Ma, Chengyue Wu, Bingxuan Wang, et al. Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding.arXiv preprint arXiv:2412.10302, 2024. 13

    2024 arXiv

  44. [54]

    Valley2: Exploring multimodal models with scalable vision-language design.arXiv preprint arXiv:2501.05901, 2025

    Ziheng Wu, Zhenghao Chen, Ruipu Luo, Can Zhang, Yuan Gao, Zhentao He, Xian Wang, Haoran Lin, and Minghui Qiu. Valley2: Exploring multimodal models with scalable vision-language design.arXiv preprint arXiv:2501.05901, 2025

    2025 arXiv

  45. [55]

    Grok-1.5 vision preview.https://x.ai/blog/grok-1.5v, 2024

    xAI. Grok-1.5 vision preview.https://x.ai/blog/grok-1.5v, 2024

    2024

  46. [56]

    Llava-o1: Let vision language models reason step-by-step.arXiv preprint arXiv:2411.10440, 2024

    Guowei Xu, Peng Jin, Li Hao, Yibing Song, Lichao Sun, and Li Yuan. Llava-o1: Let vision language models reason step-by-step.arXiv preprint arXiv:2411.10440, 2024

    2024 arXiv

  47. [57]

    R1-onevision: Advancing generalized multimodal reasoning through cross-modal formalization.arXiv preprint arXiv:2503.10615, 2025

    Yi Yang, Xiaoxuan He, Hongkun Pan, Xiyan Jiang, Yan Deng, Xingtao Yang, Haoyu Lu, Dacheng Yin, Fengyun Rao, Minfeng Zhu, et al. R1-onevision: Advancing generalized multimodal reasoning through cross-modal formalization.arXiv preprint arXiv:2503.10615, 2025

    2025 arXiv

  48. [58]

    Mulberry: Empowering mllm with o1-like reasoning and reflection via collective monte carlo tree search.arXiv preprint arXiv:2412.18319, 2024

    Huanjin Yao, Jiaxing Huang, Wenhao Wu, Jingyi Zhang, Yibo Wang, Shunyu Liu, Yingjie Wang, Yuxin Song, Haocheng Feng, Li Shen, et al. Mulberry: Empowering mllm with o1-like reasoning and reflection via collective monte carlo tree search.arXiv preprint arXiv:2412.18319, 2024

    2024 arXiv

  49. [59]

    Tree of thoughts: Deliberate problem solving with large language models.Advances in neural information processing systems, 36:11809–11822, 2023

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models.Advances in neural information processing systems, 36:11809–11822, 2023

    2023

  50. [60]

    Dapo: An open-source llm reinforcement learning system at scale.arXiv preprint arXiv:2503.14476, 2025

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, et al. Dapo: An open-source llm reinforcement learning system at scale.arXiv preprint arXiv:2503.14476, 2025

    2025 arXiv

  51. [61]

    Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthiness.arXiv preprint arXiv:2405.17220, 2024

    Tianyu Yu, Haoye Zhang, Yuan Yao, Yunkai Dang, Da Chen, Xiaoman Lu, Ganqu Cui, Taiwen He, Zhiyuan Liu, Tat-Seng Chua, et al. Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthiness.arXiv preprint arXiv:2405.17220, 2024

    2024

  52. [62]

    Mmmu-pro: A more robust multi-discipline multimodal understanding benchmark.arXiv preprint arXiv:2409.02813, 2024

    Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, et al. Mmmu-pro: A more robust multi-discipline multimodal understanding benchmark.arXiv preprint arXiv:2409.02813, 2024

    2024 arXiv

  53. [63]

    Griffon: Spelling out all object locations at any granularity with large language models

    Yufei Zhan, Yousong Zhu, Zhiyang Chen, Fan Yang, Ming Tang, and Jinqiao Wang. Griffon: Spelling out all object locations at any granularity with large language models. InEuropean Conference on Computer Vision, pages 405–422. Springer, 2024

    2024

  54. [64]

    Ferret-v2: An improved baseline for referring and grounding with large language models

    Haotian Zhang, Haoxuan You, Philipp Dufter, Bowen Zhang, Chen Chen, Hong-You Chen, Tsu-Jui Fu, William Yang Wang, Shih-Fu Chang, Zhe Gan, et al. Ferret-v2: An improved baseline for referring and grounding with large language models. InFirst Conference on Language Modeling

  55. [65]

    Improve vision language model chain-of-thought reasoning.arXiv preprint arXiv:2410.16198, 2024

    Ruohong Zhang, Bowen Zhang, Yanghao Li, Haotian Zhang, Zhiqing Sun, Zhe Gan, Yinfei Yang, Ruoming Pang, and Yiming Yang. Improve vision language model chain-of-thought reasoning.arXiv preprint arXiv:2410.16198, 2024

    2024 arXiv

  56. [66]

    Spa-vl: A comprehensive safety preference alignment dataset for vision language model.arXiv preprint arXiv:2406.12030, 2024

    Yongting Zhang, Lu Chen, Guodong Zheng, Yifeng Gao, Rui Zheng, Jinlan Fu, Zhenfei Yin, Senjie Jin, Yu Qiao, Xuanjing Huang, et al. Spa-vl: A comprehensive safety preference alignment dataset for vision language model.arXiv preprint arXiv:2406.12030, 2024

    2024 arXiv

  57. [67]

    Mm-rlhf: The next step forward in multimodal llm alignment.arXiv preprint arXiv:2502.10391, 2025

    Yi-Fan Zhang, Tao Yu, Haochen Tian, Chaoyou Fu, Peiyan Li, Jianshu Zeng, Wulin Xie, Yang Shi, Huanyu Zhang, Junkang Wu, et al. Mm-rlhf: The next step forward in multimodal llm alignment.arXiv preprint arXiv:2502.10391, 2025

    2025 arXiv

  58. [68]

    Multimodal chain-of-thought reasoning in language models.arXiv preprint arXiv:2302.00923, 2023

    Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models.arXiv preprint arXiv:2302.00923, 2023

    2023 arXiv

  59. [69]

    Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models.Advances in Neural Information Processing Systems, 36:5168–5191, 2023

    Ge Zheng, Bin Yang, Jiajin Tang, Hong-Yu Zhou, and Sibei Yang. Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models.Advances in Neural Information Processing Systems, 36:5168–5191, 2023

    2023

  60. [70]

    Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models.arXiv preprint arXiv:2504.10479, 2025

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Yuchen Duan, Hao Tian, Weijie Su, Jie Shao, et al. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models.arXiv preprint arXiv:2504.10479, 2025. 14 Appendix Content...

    2025 arXiv

  61. [71]

    The multi-choice question with the corresponding image

  62. [72]

    The true answer label (e.g. “B”)

  63. [73]

    A short, human-annotated rationale for that true answer

  64. [74]

    model’s chain-of-thought (CoT)

    The model’s PREVIOUS reasoning response, formatted exactly as <think> . . . model’s chain-of-thought (CoT). . . </think> <answer> . . . model’s final letter or text answer. . . </answer> • Inside the <think>. . . </think> block, visual cues that the model claims to use are wra...

  65. [75]

    - Then, wrap the model’s entire thought process in <think></think>

    **Response Segmentation**: - Remove the answer conclusion part in the model. - Then, wrap the model’s entire thought process in <think></think>

  66. [76]

    <vcues_1>, <vcues_2>,

    **Visual Cues Annotation**: - Within the <think> section, identify specific visual cue phrases (not entire paragraphs) and annotate each one with a tag in the format <vcues_*></vcues_*>, starting numbering from 1 (i.e. <vcues_1>, <vcues_2>, . . . )

  67. [77]

    All the data to be processed now concern reasoning errors based on visual cues rather than errors in visual cue perception

    **Visual Cues Reasoning Error Diagnosis and Correction**: 3.0. All the data to be processed now concern reasoning errors based on visual cues rather than errors in visual cue perception. These reasoning errors may include issues such as insufficient knowledge, over-analysis, e...

  68. [78]

    based on the rationale

    **Output Constraints**: - Preserve the original reasoning structure as possible. - **Do not include similar phrases like "based on the rationale", "The reasoning should focus", "aligns with the rationale", "the model", beacuse the processed content is used for the model traini...

  69. [79]

    <vcues_1>The image shows a steam train on the tracks</vcues_1>

  70. [80]

    <vcues_2>There is a small green building on the right side of the image, which appears to be a signal box or a small station</vcues_2>

  71. [81]

    <vcues_3>The train is moving on the tracks, and there is no indication that it has stopped at a station</vcues_3>

  72. [82]

    Let's verify each visual cue and its reasoning before finalizing the answer

    <vcues_4>There are no gates visible in the image, and the train is not shown breaking down</vcues_4>. Let's verify each visual cue and its reasoning before finalizing the answer. Some visual cues seem to need adjustment. <vcues_2>The small green building on the right side of t...

  73. [83]

    Use the properties of parallel lines: - Since AD is parallel to BC, the angle formed by the transversal and line AD will be related to the angle formed by the transversal and line BC

  74. [84]

    - The angle 78° is an interior angle of the triangle, and angle 1 is 42°

    Determine the relationship between the angles: - <vcues_4>Angle 2 is an exterior angle to the triangle formed by the transversal and lines AD and BC</vcues_4>. - The angle 78° is an interior angle of the triangle, and angle 1 is 42°. - The exterior angle (angle 2) is the sum o...

  75. [85]

    However, upon reevaluating the problem, it appears there might be a misunderstanding in the interpretation of the angles

    Calculate the angle: - <vcues_5>Angle 2 = 78° - 42° = 36°</vcues_5>. However, upon reevaluating the problem, it appears there might be a misunderstanding in the interpretation of the angles. The correct approach is to use the fact that angle 2 is the exterior angle to the tria...

  76. [86]

    - <vcues_2>Angle 1 is 42°</vcues_2>

    Identify the given information: - <vcues_1>AD is parallel to BC</vcues_1>. - <vcues_2>Angle 1 is 42°</vcues_2>. - <vcues_3>Angle 78° is an angle formed by the transversal and line AD</vcues_3>. Figure 7: Response sample from GThinker with reflection for the math problem. 9

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.