Pith. sign in

REVIEW 6 major objections 6 minor 37 references

Enhancing Spatial Reasoning in Vision-Language Models via Chain-of-Thought Prompting and Reinforcement Learning

T0 review · 6 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that GRPO, a reinforcement-learning fine-tuning method, gives vision-language models spatial reasoning that survives paraphrase shifts, while supervised fine-tuning memorizes phrasing and fails out-of-distribution.

desk verdict A useful but under-supported empirical study; the OOD claim needs more than one paraphrase and the tables need a careful audit. read the letter →

arxiv 2507.13362 v1 pith:SLDHWO3Z submitted 2025-07-06 cs.CV cs.AIcs.CL

Binbin Ji , Siddharth Agrawal , Qiance Tang , Yvonne Wu This is my paper
classification cs.CVcs.AIcs.CL
keywords spatialreasoningvision-languagemodelschain-of-thoughtpromptingscenegraphsGRPOreinforcementlearningfine-tuningout-of-distributiongeneralizationsupervised
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

This paper claims that reinforcement-learning fine-tuning gives vision-language models spatial reasoning that survives when test questions are reworded, whereas supervised fine-tuning does not. The authors compare Group Relative Policy Optimization (GRPO) against supervised fine-tuning on the same spatial dataset and find GRPO wins under Pass@1 evaluation and degrades far less when "closer to" is flipped to "farther from." They also show that naive chain-of-thought prompting can hurt spatial reasoning, while a structured two-stage prompt that first builds a scene graph and then answers improves accuracy. If the claim holds, RL post-training is the safer route for deployed systems in robotics or navigation where phrasing varies at test time.

What carries the argument

The central mechanism is GRPO (Group Relative Policy Optimization), a reinforcement-learning update that compares responses within a group to compute advantages, applied here with a "No-Thinking" objective that skips chain-of-thought and directly generates the answer. The other load-bearing component is Scene Graph CoT, a two-stage prompt that first extracts objects, attributes, and relations as a JSON scene graph, then uses that graph to generate the final answer. A dynamic-sampling addition borrowed from DAPO addresses reward collapse, where all responses in a group earn identical rewards and the advantage signal vanishes. The out-of-distribution probe is a linguistic polarity flip, rephrasing questions from "closer to" to "farther from" with images unchanged, which separates genuine spatial understanding from memorized phrasing.

What would settle it

Run the same GRPO and SFT models on additional rephrasings, including "near" versus "far," "which object is more distant," and reversed sentence order, on the full CVBench Depth and Distance splits; if SFT matches or beats GRPO on any of these, the claim that SFT overfits to surface phrasing while GRPO generalizes would need revision.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes a generalization hierarchy: for spatial reasoning in vision-language models, GRPO fine-tuning learns transferable spatial concepts, while supervised fine-tuning overfits to surface linguistic patterns. Evidence is the out-of-distribution polarity test: on Depth, SFT drops from 61.67% to 56.33% when questions change from "closer to" to "farther from," worse than the base model, while GRPO reaches 70.50%; on Distance, SFT's in-distribution to out-of-distribution gap is 12.03 percentage points versus 3.17 for GRPO. The paper further argues that two-stage Scene Graph CoT prompting, generating a JSON scene graph first and then answering, consistently beats direct answering by 5 to 15 percent, and that splitting the two steps prevents reward hacking. It also reports that a segmentation-aware variant of Qwen2.5-VL improves counting, with CVBench counting accuracy rising from 21.05% to 64.21%.

Load-bearing premise

The GRPO-over-SFT conclusion assumes that flipping "closer to" into "farther from" is a representative out-of-distribution test and that the measured accuracy gaps on Depth and Distance are not noise from a small evaluation subset.

Editorial extensions

If this is right

  • Deploying GRPO-tuned vision-language models is safer than supervised fine-tuning when user queries paraphrase the training questions, since GRPO keeps most of its in-distribution accuracy under polarity flips.
  • Scene Graph CoT should be preferred over naive CoT for spatial visual question answering: simple CoT can degrade accuracy, while the split two-stage prompt improves it by 5 to 15 percent and avoids shortcutting.
  • Supervised fine-tuning can actively hurt out-of-distribution performance, with Depth OOD accuracy falling below the base model, so SFT gains on benchmark phrasing may overstate real improvement.
  • GRPO-v2 with dynamic sampling achieves the highest Pass@1 total on CVBench, 72.38% versus 66.79% for SFT, indicating that attentive reward computation matters for the generalization benefit.

Reading between the lines

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

  • Editorial: the out-of-distribution result rests on a single antonym pair, so a natural extension is a paraphrase battery covering near versus far, above versus below, left versus right, and passive voice to map the boundary of GRPO's robustness.
  • Editorial: if GRPO primarily prunes suboptimal response paths rather than acquiring new abilities, as the Pass@1 versus Pass@4 tradeoff suggests, then its advantage over SFT should shrink when the base model is already well aligned, which is testable by varying base-model strength.
  • Editorial: the reward-collapse observation points to a practical monitoring rule: when group-wise rewards are identical, GRPO degenerates toward the reference policy, so tracking reward diversity could predict when RL fine-tuning stalls.
  • Editorial: the scene-graph separation implies that answer quality depends on graph quality, so feeding oracle scene graphs versus model-generated graphs would measure how much of the CoT gain comes from perception rather than reasoning.
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

6 major / 6 minor

Summary. The paper investigates spatial reasoning in vision-language models (VLMs) through prompting and reinforcement learning. It reports three main findings: (i) naive chain-of-thought (CoT) prompting fails to help or even hurts performance; (ii) structured two-step SceneGraph CoT prompting improves spatial reasoning accuracy; and (iii) GRPO fine-tuning on the SAT benchmark generalizes better than supervised fine-tuning on CVBench, especially under an out-of-distribution polarity shift from "closer to" to "farther from." The paper also proposes conversational few-shot scene-graph prompting and provides an ablation study over alternative CoT variants such as PoT, CoS, and VoT. Code is released on GitHub.

Significance. If the central claims hold, the paper would provide a useful empirical result for VLM post-training: that GRPO induces semantic generalization beyond surface linguistic patterns, while SFT overfits to training phrasing. The strengths are the breadth of benchmarks and prompting variants, the explicit SFT-versus-GRPO comparison on the same base model, the release of code, and the attempt to decompose the benefit of structured prompting. However, the headline OOD claim rests on a single polarity flip with no error bars or sample sizes, the stated GRPO loss in Eq. (1) does not match the standard algorithm, and there are internal table inconsistencies that undermine confidence in the reported numbers. The claim that naive CoT harms performance is asserted without data. These issues must be resolved before the results can be considered conclusive.

major comments (6)
  1. [Section III-B, Eq. (1)] The GRPO loss in Eq. (1) is mis-specified relative to the GRPO algorithm cited in [1]: it lacks the clipped surrogate objective, the group-relative advantage normalization over sampled responses, and the placement/sign of the KL penalty is nonstandard. As written, the equation is not the objective used by GRPO implementations and cannot serve as the basis for reproducing the experiments. Please replace it with the correct GRPO objective or explicitly clarify the exact objective used in the implementation.
  2. [Section IV-B, Table IV] The OOD generalization claim rests on a single linguistic polarity flip ('closer to' -> 'farther from') with no reported sample size, no error bars, no repeated seeds, and no statistical test. The observed gaps (e.g., Depth OOD: SFT 56.33% vs GRPO 70.50%) could be due to chance or to prompt-specific artifacts. To support the abstract's claim of superior robustness under OOD conditions, the authors should report variance over seeds, test additional paraphrases (e.g., 'near'/'far', reversed sentence order), and include per-item paired comparisons.
  3. [Section III-C] The scene-graph prompting protocol leaks the answer in step 1: the text states that 'the model frequently also includes the correct answer implicitly, while step 2 mainly serves to structure the output for evaluation.' This means the reported gains of SceneGraph CoT (Section IV-C and Table III) may be attributable to answer leakage rather than to the value of structured scene-graph reasoning. The authors should either mask answer content in the extracted scene graphs or compare against a control that removes such leakage.
  4. [Section IV-C and Table V] There is an internal inconsistency between Table III and Table V: Table III assigns Qwen2.5-VL + SceneGraph CoT a total of 68.08% and Llama-4-Scout + SceneGraph CoT 75.66%, while Table V lists the reverse for the 'VC Bench (SG CoT)' column. This suggests a row/column swap or bookkeeping error that could also affect Table IV and the GRPO-vs-SFT comparison; please correct and re-verify all cross-referenced tables.
  5. [Introduction and Abstract] The claim that naive CoT prompting 'not only fails to help, but can even harm' the model's original performance is asserted without presenting any experimental data, table, or figure. As this is one of the paper's headline findings, it needs to be substantiated with a quantitative comparison (for example, a row in Table III or a separate table).
  6. [Section III-B and Table II] The training configuration is underspecified: the paper does not report learning rate, batch size, number of GRPO steps, LoRA alpha, reward function details, or the SFT training protocol (dataset size, epochs, loss type). Without these, the GRPO-vs-SFT comparison cannot be independently reproduced, which is essential for a paper whose central contribution is a training method comparison.
minor comments (6)
  1. [Table III] The column headers say 'Pass@4' while the note says all results are Pass@1 unless specified in parentheses, which are Pass@4 values. Please harmonize the headers with the note to avoid confusion.
  2. [Section IV-D] The paragraph at the end of Section IV-D ends with 'It is also interesting to note that' without any continuation; please complete or delete the sentence.
  3. [Table V] The column labeled 'VC Bench' should be 'CV-Bench' for consistency with the rest of the paper.
  4. [Table I] The model name 'Regular QwenVL2.5-3B' differs from 'Qwen2.5-VL-Instruct-3B' used in Section IV-A; please unify the naming.
  5. [Methodology] The metrics Pass@1 and Pass@4 are never formally defined; please add a definition in the methodology section.
  6. [Eq. (2)] The Close-Call Percentage formula has a denominator that can be zero when the model makes no errors; please state the convention for that case (e.g., set to 0 or report as undefined).

Circularity Check

0 steps flagged · score 0.0 of 10

No structural circularity: the GRPO-vs-SFT OOD comparison is evaluated on external CVBench data, and no load-bearing self-citation or fitted-input-to-prediction reduction is present. One admitted answer-leakage limitation weakens the SceneGraph CoT ablation but is not a circular derivation.

full rationale

The derivation chain is self-contained against external benchmarks. The central GRPO-vs-SFT claim is tested by fine-tuning on SAT and evaluating on CVBench (Table III) plus a polarity-flipped OOD variant (Table IV, Section IV-B); the OOD numbers are not fitted parameters and are not constructed from the claim, so the comparison is an independent empirical result. The paper cites external prior work for GRPO (DeepSeekMath), DAPO, and baseline methods, but no load-bearing self-citation or imported uniqueness theorem is used. The one self-admitted issue is in Section III-C: 'Although step 1 is only required to generate the scene graph, the model frequently also includes the correct answer implicitly, while step 2 mainly serves to structure the output for evaluation.' This answer leakage means the SceneGraph CoT accuracy partly measures whether the model can emit the answer inside the scene graph, a validity threat to that specific ablation, but it is not a circular derivation: the final answer is not fed into the scene graph by construction, the scene graph is generated by the same model, and the central GRPO/SFT OOD comparison does not depend on this leakage. Table V's apparent row/column mismatch with Table III (Qwen2.5-VL and Llama-4-Scout 'VC Bench (SG CoT)' values) is a bookkeeping/correctness concern, not circularity. Score 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

This is an empirical ML paper, so there is no mathematical derivation with fitted constants. The central comparison rests on hand-chosen training hyperparameters and on domain assumptions about benchmark validity and the correctness of the GRPO implementation. No new entities (particles, forces, etc.) are introduced.

free parameters (3)
  • GRPO-v2 KL penalty β = 0.01
    Selected by the authors; the relative performance of GRPO vs SFT could change with different KL weights.
  • GRPO-v2 LoRA rank = 16
    Selected by the authors; one of the differences between GRPO-v1 and GRPO-v2.
  • GRPO-v2 LoRA target modules = q,k,v,o,gate,up,down projections
    Selected by the authors; expanding beyond q_proj,v_proj was needed to see gains.
assumptions (4)
  • domain assumption CVBench, SAT, VSR, and CLEVR-CoGenT are valid measures of spatial reasoning ability.
    The paper's conclusions about spatial reasoning and generalization are drawn from accuracy on these benchmarks; if the benchmarks are biased or leak information, the conclusions do not follow. The prompt-only baseline (52.25% vs 48.82% with image) hints at possible information leakage in the CLEVR prompt design.
  • domain assumption The GRPO implementation used for fine-tuning corresponds to the algorithm described in the paper and to the reference implementation.
    The loss in Eq. (1) is written incorrectly as a combined importance-weighted objective with KL inside the expectation, so the reader must assume the actual code implements standard GRPO correctly. The GRPO-v2 dynamic sampling is credited to DAPO [27].
  • domain assumption The out-of-distribution test set, formed by replacing 'closer to' with 'farther from', preserves the intended spatial semantics and does not change the answer distribution.
    The OOD comparison in Table IV is the main evidence for the claim that SFT overfits to phrasing. If the rephrasing changes difficulty or answer distribution differently for the two models, the observed gaps (Depth OOD 56.33 vs 70.50) would not support the conclusion.
  • domain assumption The models' weights and the SAT/CVBench data are used as-is; no training/evaluation data overlap or leakage exists.
    The paper fine-tunes on SAT and evaluates on CVBench to claim OOD generalization. Any overlap or metadata leakage would invalidate that claim.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Spatial Reasoning in Vision-Language Models via Chain-of-Thought Prompting and Reinforcement Learning." pith.science (2026). https://pith.science/paper/SLDHWO3Z

@misc{pith2026250713362,
  author       = {Pith},
  title        = {Pith review of: Enhancing Spatial Reasoning in Vision-Language Models via Chain-of-Thought Prompting and Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SLDHWO3Z}},
  note         = {Machine review of arXiv:2507.13362}
}
read the original abstract

This study investigates the spatial reasoning capabilities of vision-language models (VLMs) through Chain-of-Thought (CoT) prompting and reinforcement learning. We begin by evaluating the impact of different prompting strategies and find that simple CoT formats, where the model generates a reasoning step before the answer, not only fail to help, but can even harm the model's original performance. In contrast, structured multi-stage prompting based on scene graphs (SceneGraph CoT) significantly improves spatial reasoning accuracy. Furthermore, to improve spatial reasoning ability, we fine-tune models using Group Relative Policy Optimization (GRPO) on the SAT dataset and evaluate their performance on CVBench. Compared to supervised fine-tuning (SFT), GRPO achieves higher accuracy on Pass@1 evaluations and demonstrates superior robustness under out-of-distribution (OOD) conditions. In particular, we find that SFT overfits to surface-level linguistic patterns and may degrade performance when test-time phrasing changes (e.g., from "closer to" to "farther from"). GRPO, on the other hand, generalizes more reliably and maintains stable performance under such shifts. Our findings provide insights into how reinforcement learning and structured prompting improve the spatial reasoning capabilities and generalization behavior of modern VLMs. All code is open source at: https://github.com/Yvonne511/spatial-vlm-investigator

Figures

Figures reproduced from arXiv: 2507.13362 by the authors.

Figure 1
Figure 1. Accuracy Reward of GRPO-v2 model formance under out-of-distribution (OOD) conditions, but ac￾tually reduces accuracy compared to the base model (56.33% vs. 59.00%). This suggests that SFT encourages the model to memorize surface-level linguistic patterns rather than fostering transferable spatial understanding. When test queries shift semantically from ”closer to” to ”farther from,” the SFT model struggles to adapt,… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 7 canonical work pages

  1. [1]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models,

    Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y . K. Li, Y . Wu, and D. Guo, “Deepseekmath: Pushing the limits of mathematical reasoning in open language models,” 2024. [Online]. Available: https://arxiv.org/abs/2402.03300

    arXiv 2024

  2. [2]

    Visual spatial reasoning,

    F. Liu, G. Emerson, and N. Collier, “Visual spatial reasoning,” 2023. [Online]. Available: https://arxiv.org/abs/2205.00363

    arXiv 2023

  3. [3]

    Sat: Dynamic spatial aptitude training for multimodal language models,

    A. Ray, J. Duan, E. Brown, R. Tan, D. Bashkirova, R. Hendrix, K. Ehsani, A. Kembhavi, B. A. Plummer, R. Krishna, K.-H. Zeng, and K. Saenko, “Sat: Dynamic spatial aptitude training for multimodal language models,” 2025. [Online]. Available: https: //arxiv.org/abs/2412.07755

    arXiv 2025

  4. [4]

    Cambrian-1: A fully open, vision-centric exploration of multimodal llms,

    S. Tong, E. Brown, P. Wu, S. Woo, M. Middepogu, S. C. Akula, J. Yang, S. Yang, A. Iyer, X. Pan, Z. Wang, R. Fergus, Y . LeCun, and S. Xie, “Cambrian-1: A fully open, vision-centric exploration of multimodal llms,” 2024. [Online]. Available: https://arxiv.org/abs/2406.16860

    arXiv 2024

  5. [5]

    Clevr: A diagnostic dataset for compositional language and elementary visual reasoning,

    J. Johnson, B. Hariharan, L. Van Der Maaten, L. Fei-Fei, C. Lawrence Zitnick, and R. Girshick, “Clevr: A diagnostic dataset for compositional language and elementary visual reasoning,” in Proceed- ings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2901–2910

    work page 2017

  6. [6]

    Visual-rft: Visual reinforcement fine-tuning,

    Z. Liu, Z. Sun, Y . Zang, X. Dong, Y . Cao, H. Duan, D. Lin, and J. Wang, “Visual-rft: Visual reinforcement fine-tuning,” 2025. [Online]. Available: https://arxiv.org/abs/2503.01785

    arXiv 2025

  7. [7]

    Sft memorizes, rl generalizes: A comparative study of foundation model post-training,

    T. Chu, Y . Zhai, J. Yang, S. Tong, S. Xie, D. Schuurmans, Q. V . Le, S. Levine, and Y . Ma, “Sft memorizes, rl generalizes: A comparative study of foundation model post-training,” 2025. [Online]. Available: https://arxiv.org/abs/2501.17161

    arXiv 2025

  8. [8]

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

    L. Chen, L. Li, H. Zhao, Y . 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

    work page 2025

Show all 37 references
  1. [9]

    Rlvr in vision language models: Findings, questions and directions,

    L. Chen, L. Li, H. Zhao, Y . Song, Vinci, L. Kong, Q. Liu, and B. Chang, “Rlvr in vision language models: Findings, questions and directions,” Notion Post , Feb 2025. [Online]. Available: https: //deepagent.notion.site/rlvr-in-vlms

    2025

  2. [10]

    R1-zero’s

    H. Zhou, X. Li, R. Wang, M. Cheng, T. Zhou, and C.-J. Hsieh, “R1-zero’s ”aha moment” in visual reasoning on a 2b non-sft model,”

  3. [11]

    Vlm-r1: A stable and generalizable r1-style large vision-language model,

    H. Shen, P. Liu, J. Li, C. Fang, Y . Ma, J. Liao, Q. Shen, Z. Zhang, K. Zhao, Q. Zhang, R. Xu, and T. Zhao, “Vlm-r1: A stable and generalizable r1-style large vision-language model,” 2025. [Online]. Available: https://arxiv.org/abs/2504.07615

    2025 arXiv

  4. [12]

    Think or not think: A study of explicit thinking in rule-based visual reinforcement fine-tuning,

    M. Li, J. Zhong, S. Zhao, Y . Lai, H. Zhang, W. B. Zhu, and K. Zhang, “Think or not think: A study of explicit thinking in rule-based visual reinforcement fine-tuning,” 2025. [Online]. Available: https://arxiv.org/abs/2503.16188

    2025

  5. [13]

    Video-r1: Reinforcing video reasoning in mllms,

    K. Feng, K. Gong, B. Li, Z. Guo, Y . Wang, T. Peng, B. Wang, and X. Yue, “Video-r1: Reinforcing video reasoning in mllms,” 2025. [Online]. Available: https://arxiv.org/abs/2503.21776

    2025 arXiv

  6. [14]

    Spatial-r1: Enhancing mllms in video spatial reasoning,

    K. Ouyang, “Spatial-r1: Enhancing mllms in video spatial reasoning,”

  7. [15]

    Improved visual-spatial reasoning via r1-zero-like training,

    Z. Liao, Q. Xie, Y . Zhang, Z. Kong, H. Lu, Z. Yang, and Z. Deng, “Improved visual-spatial reasoning via r1-zero-like training,” 2025. [Online]. Available: https://arxiv.org/abs/2504.00883

    2025 arXiv

  8. [16]

    Available: https://arxiv.org/abs/2504.01805

    [Online]. Available: https://arxiv.org/abs/2504.01805

    arXiv

  9. [17]

    Compositional chain- of-thought prompting for large multimodal models,

    C. Mitra, B. Huang, T. Darrell, and R. Herzig, “Compositional chain- of-thought prompting for large multimodal models,” 2024. [Online]. Available: https://arxiv.org/abs/2311.17076

    2024 arXiv

  10. [18]

    Thinking in space: How multimodal large language models see, remember, and recall spaces,

    J. Yang, S. Yang, A. W. Gupta, R. Han, L. Fei-Fei, and S. Xie, “Thinking in space: How multimodal large language models see, remember, and recall spaces,” 2024. [Online]. Available: https://arxiv.org/abs/2412.14171

    2024 arXiv

  11. [19]

    Mind’s eye of llms: Visualization-of-thought elicits spatial reasoning in large language models,

    W. Wu, S. Mao, Y . Zhang, Y . Xia, L. Dong, L. Cui, and F. Wei, “Mind’s eye of llms: Visualization-of-thought elicits spatial reasoning in large language models,” arXiv preprint arXiv:2404.03622 , 2024. [Online]. Available: https://arxiv.org/abs/2404.03622

    2024 arXiv

  12. [20]

    Spatialcot: Advancing spatial reasoning through coordinate alignment and chain-of-thought for embodied task planning,

    Y . Liu, D. Chi, S. Wu, Z. Zhang, Y . Hu, L. Zhang, Y . Zhang, S. Wu, T. Cao, G. Huang, H. Huang, G. Tian, W. Qiu, X. Quan, J. Hao, and Y . Zhuang, “Spatialcot: Advancing spatial reasoning through coordinate alignment and chain-of-thought for embodied task planning,” 2025. [On...

    2025 arXiv

  13. [21]

    Clevr cogent valb,

    MMInstruction, “Clevr cogent valb,” https://huggingface.co/datasets/ MMInstruction/Clevr CoGenT ValB, accessed: 2025-05-14

    2025

  14. [22]

    Chain-of-symbol prompting elicits planning in large language models,

    H. Hu, H. Lu, H. Zhang, Y .-Z. Song, W. Lam, and Y . Zhang, “Chain-of-symbol prompting elicits planning in large language models,” arXiv preprint arXiv:2305.10276 , 2023. [Online]. Available: https://arxiv.org/abs/2305.10276

    2023 arXiv

  15. [23]

    Depth anything v2,

    L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, and H. Zhao, “Depth anything v2,” arXiv preprint arXiv:2406.09414 , 2024

    2024 arXiv

  16. [24]

    Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning,

    Z. Li, X. Wang, E. Stengel-Eskin, A. Kortylewski, W. Ma, B. V . Durme, and A. Yuille, “Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning,” 2023. [Online]. Available: https://arxiv.org/abs/2212.00259

    2023 arXiv

  17. [25]

    Clipsam: Clip and sam collaboration for zero-shot anomaly segmentation,

    S. Li, J. Cao, P. Ye, Y . Ding, C. Tu, and T. Chen, “Clipsam: Clip and sam collaboration for zero-shot anomaly segmentation,” 2024. [Online]. Available: https://arxiv.org/abs/2401.12665

    2024 arXiv

  18. [26]

    Molmo and pixmo: Open weights and open data for state-of-the-art vision-language models,

    M. Deitke, C. Clark, S. Lee, R. Tripathi, Y . Yang, J. S. Park, M. Salehi, N. Muennighoff, K. Lo, L. Soldaini, J. Lu, T. Anderson, E. Bransom, K. Ehsani, H. Ngo, Y . Chen, A. Patel, M. Yatskar, C. Callison-Burch, A. Head, R. Hendrix, F. Bastani, E. VanderBilt, N. Lambert, Y . ...

    2024 arXiv

  19. [27]

    Dapo: An open-source llm reinforcement learning system at scale,

    Q. Yu, Z. Zhang, R. Zhu, Y . Yuan, X. Zuo, Y . Yue, T. Fan, G. Liu, L. Liu, X. Liu, H. Lin, Z. Lin, B. Ma, G. Sheng, Y . Tong, C. Zhang, M. Zhang, W. Zhang, H. Zhu, J. Zhu, J. Chen, J. Chen, C. Wang, H. Yu, W. Dai, Y . Song, X. Wei, H. Zhou, J. Liu, W.-Y . Ma, Y .-Q. Zhang, L....

    2025 arXiv

  20. [28]

    Paligemma 2: A family of versatile vlms for transfer,

    A. Steiner, A. S. Pinto, M. Tschannen, D. Keysers, X. Wang, Y . Bitton, A. Gritsenko, M. Minderer, A. Sherbondy, S. Long, S. Qin, R. Ingle, E. Bugliarello, S. Kazemzadeh, T. Mesnard, I. Alabdulmohsin, L. Beyer, and X. Zhai, “Paligemma 2: A family of versatile vlms for transfer...

    2024 arXiv

  21. [29]

    Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model?

    Y . Yue, Z. Chen, R. Lu, A. Zhao, Z. Wang, Y . Yue, S. Song, and G. Huang, “Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model?” 2025. [Online]. Available: https://arxiv.org/abs/2504.13837

    2025 arXiv

  22. [30]

    Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks,

    W. Chen, X. Ma, X. Wang, and W. W. Cohen, “Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks,” arXiv preprint arXiv:2211.12588 , 2022

    2022 arXiv

  23. [31]

    Training large language models to reason in a continuous latent space,

    S. Hao, S. Sukhbaatar, D. Su, X. Li, Z. Hu, J. Weston, and Y . Tian, “Training large language models to reason in a continuous latent space,” 2024. [Online]. Available: https://arxiv.org/abs/2412.06769

    2024 arXiv

  24. [32]

    Embodiedvsr: Dynamic scene graph-guided chain-of-thought reasoning for visual spatial tasks,

    Y . Zhang, Q. Zhang, X. Ju, Z. Liu, J. Mao, J. Sun, J. Wu, S. Gao, S. Cai, Z. Qin, L. Liang, J. Wang, Y . Duan, J. Cao, R. Xu, and J. Tang, “Embodiedvsr: Dynamic scene graph-guided chain-of-thought reasoning for visual spatial tasks,” 2025. [Online]. Available: https://arxiv.o...

    2025 arXiv

  25. [33]

    Spatialbot: Precise spatial understanding with vision language models,

    W. Cai, I. Ponomarenko, J. Yuan, X. Li, W. Yang, H. Dong, and B. Zhao, “Spatialbot: Precise spatial understanding with vision language models,” 2025. [Online]. Available: https://arxiv.org/abs/2406.13642

    2025 arXiv

  26. [34]

    Videochat-r1: Enhancing spatio-temporal perception via reinforcement fine-tuning,

    X. Li, Z. Yan, D. Meng, L. Dong, X. Zeng, Y . He, Y . Wang, Y . Qiao, Y . Wang, and L. Wang, “Videochat-r1: Enhancing spatio-temporal perception via reinforcement fine-tuning,” 2025. [Online]. Available: https://arxiv.org/abs/2504.06958

    2025 arXiv

  27. [35]

    Spatialvlm: Endowing vision- language models with spatial reasoning capabilities,

    B. Chen, Z. Xu, S. Kirmani, B. Ichter, D. Driess, P. Florence, D. Sadigh, L. Guibas, and F. Xia, “Spatialvlm: Endowing vision- language models with spatial reasoning capabilities,” 2024. [Online]. Available: https://arxiv.org/abs/2401.12168

    2024 arXiv

  28. [36]

    Spatialpin: Enhancing spatial reasoning capabilities of vision-language models through prompting and interacting 3d priors,

    C. Ma, K. Lu, T.-Y . Cheng, N. Trigoni, and A. Markham, “Spatialpin: Enhancing spatial reasoning capabilities of vision-language models through prompting and interacting 3d priors,” 2024. [Online]. Available: https://arxiv.org/abs/2403.13438

    2024 arXiv

  29. [2025]

    Available: https://arxiv.org/abs/2503.05132

    [Online]. Available: https://arxiv.org/abs/2503.05132

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.