Pith. sign in

REVIEW 4 major objections 6 minor 32 references

CLGRPO: Reasoning Ability Enhancement for Small VLMs

T0 review · 4 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 Four-stage post-training lifts a 1B vision-language model to 8B-level accuracy on emotion recognition.

desk verdict A coherent four-stage recipe for small-VLM reasoning, but the headline number is likely selected on the test set, so the central gain is unverified. read the letter →

arxiv 2506.18048 v2 pith:3WGAYBYJ submitted 2025-06-22 cs.CV

Fanyi Wang , Binzhi Dong , Haotian Hu , Jinjin Xu , Zhiwang Zhang This is my paper
classification cs.CV
keywords smallvision-languagemodelchain-of-thoughtreasoningGroupRelativePolicyOptimizationCLGRPOincrementaltrainingemotionrecognitionEmoSet-118Kpost-training
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 argues that a small vision-language model (at or below 2B parameters) can acquire chain-of-thought reasoning ability through a carefully staged post-training recipe rather than by scaling parameters. On the EmoSet-118K emotion recognition benchmark, a 1B model trained with four stages—SFT on self-supervised chain-of-thought data, format-only reinforcement learning, full reinforcement learning, and a clipped variant called CLGRPO—reaches 81.45% accuracy, beating the 78.68% baseline and matching an 8B model. The central claim is that inexpensive small models can become competitive on domain-specific visual reasoning if both the training signal and the reinforcement-learning search space are shaped to their limited capacity.

What carries the argument

The central mechanism is ClipLow GRPO (CLGRPO), a variant of Group Relative Policy Optimization in which both clipping thresholds $\epsilon_l$ and $\epsilon_h$ are reduced from 0.2 to 0.1, deliberately narrowing the space of policy updates during reinforcement learning. It carries the argument by showing that a small model benefits from a smaller search space, and it is preceded by a format-alignment stage that teaches the model the required output structure before reasoning rewards are applied.

What would settle it

Take the COT training set and replace every REASONING section with a fixed generic sentence while keeping the image, the SUMMARY, and the correct CONCLUSION, then run the same four-stage pipeline and measure test accuracy on EmoSet-118K; if accuracy stays near 81.45%, the reasoning content is not what drives the improvement.

Watch

Extended reading notes

Core claim

The paper proposes an Incremental Training Strategy that enhances the reasoning ability of small vision-language models through four sequential stages. Stage 1 performs supervised fine-tuning on chain-of-thought data generated by larger VLMs and self-supervisedly filtered; Stage 2 aligns the model to the required output format using a small amount of Group Relative Policy Optimization with only a format reward; Stage 3 applies GRPO with both format and accuracy rewards; Stage 4 introduces ClipLow GRPO, which lowers both clipping thresholds from 0.2 to 0.1 to constrain the policy-update search space. On EmoSet-118K, this staged recipe improves the 1B model's accuracy by 2.77 points over the baseline fine-tuned on original data, achieves 81.45% accuracy, and surpasses the accuracy of an 8B model fine-tuned on the same original data while trailing slightly in recall.

Load-bearing premise

The judge model that filters the self-supervised chain-of-thought data correctly decides whether a generated conclusion matches the ground-truth emotion, because the entire training pipeline relies on that filtered COT data being accurate rather than just plausible.

Editorial extensions

If this is right

  • A 1B vision-language model can outperform its own supervised-fine-tuning baseline by 2.77 accuracy points on a domain-specific visual reasoning task.
  • The staged strategy closes the accuracy gap to an 8B model fine-tuned on the same data, with the 1B model reaching 81.45% versus 80.58%.
  • Format-alignment training before full GRPO is necessary for small models: skipping it drops accuracy from 80.31% to 75.00%.
  • Lowering both clip thresholds from 0.2 to 0.1 improves accuracy and recall, while raising the upper threshold to 0.28 degrades both metrics.
  • Direct GRPO on chain-of-thought data without knowledge injection underperforms GRPO on original data, confirming that the staged knowledge injection is load-bearing.

Reading between the lines

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

  • If the mechanism is generic, the same four-stage recipe should transfer to other small-model vertical domains whenever a larger model can generate trustworthy chain-of-thought data; the paper only tests one dataset.
  • The failure of Dr.GRPO on this visual task suggests that reasoning-enhancement algorithms tuned for large, math-capable models do not automatically transfer to small visual models, and clip schedules may need to be model-scale-specific.
  • The per-class pattern of rising accuracy with falling recall (e.g., the 'amuse' class) hints that the small model learns overconfident decision boundaries; pairing CLGRPO with calibration or class-balanced rewards could lift low-recall classes.
  • A testable extension is to measure whether the benefit of the lower clip thresholds grows as model size shrinks, which would directly support the capacity-based explanation.
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 / 6 minor

Summary. The paper proposes a post-training recipe for enhancing the reasoning ability of small vision-language models (SVLMs), specifically InternVL2.5-1B-Instruct. The recipe consists of a self-supervised chain-of-thought (COT) data construction system, in which larger LVLMs generate COT-format answers and a 7B judge model verifies that the predicted emotion matches the ground truth, followed by a four-stage Incremental Training Strategy: Stage 1 SFT on COT data, Stage 2 GRPO with format-only reward, Stage 3 GRPO with format and accuracy rewards, and Stage 4 CLGRPO, which lowers the PPO clipping range. Experiments are conducted on EmoSet-118K. The central empirical claim is that the full pipeline raises accuracy from 78.68% (baseline SFT on original data) to 81.45% and recall from 78.51% to 79.20%, reaching accuracy comparable to an 8B model fine-tuned on the original data. Ablations are presented for the format-alignment stage and for the clipping-range choice.

Significance. If the reported gains are statistically reliable, the paper would provide a practical and cost-effective recipe for improving domain-specific reasoning in sub-2B VLMs, a topic of clear industrial relevance. The staged approach is conceptually clean, and the result that a short format-alignment phase materially helps a small model is an interesting and non-obvious empirical finding. The paper does not propose a new theoretical mechanism, and CLGRPO itself is a minor modification of GRPO's clipping range, but the contribution could still be useful as an engineering recipe. The main value depends on whether the evaluation protocol is trustworthy; at present the evidence base is a single dataset, a single base model, single runs, and an evaluation protocol that appears to use the test set for model selection. These issues limit the paper's significance until addressed.

major comments (4)
  1. [§4.2, Table 2/3, Figure 5] The headline numbers are selected from the test set rather than from a pre-specified holdout rule. Section 4.2 states that "the two best-performing weights on the test data are listed" in Tables 2 and 3, and Figure 5 shows accuracy and recall "evaluated in the test data during the training process of Stage 4." This is in tension with Section 4.1, which states that "the best-performing weights on the evaluation data were selected." Reporting the maximum over the test-data trajectory makes 81.45% an extreme order statistic rather than an unbiased estimate of generalization, and it makes the comparison against baselines, which appear to have been selected on evaluation data, unfair. This is load-bearing because the paper's central conclusion rests on this single number. The authors should either pre-specify a validation-based checkpoint rule, report the final checkpoint after a fixed number of iterations, or evaluate a randomly chosen checkpoint on the test set, and they should clearly distinguish validation and test sets.
  2. [§4.1, Tables 2-7] Every result in the paper appears to be a single run with no repeated seeds, error bars, or significance testing. Many of the comparisons that support the method are small: Stage 3 accuracy is 80.31% versus baseline 78.68%, Stage 4-v1 is 81.45% versus Stage 3's 80.31%, and Stage 4-v1 versus Stage 4-v2 differ by only 0.25 accuracy points. Without multiple seeds or at least a bootstrap analysis, the observed improvements could be within run-to-run or checkpoint-to-checkpoint noise. This is particularly relevant given that Section 4.2 explicitly mentions selecting the "two best-performing weights," which suggests nontrivial variance over the training trajectory. The authors should report mean and standard deviation over at least three independent runs, or provide a convincing statistical analysis.
  3. [§3.1, Figure 1] The Self-Supervised COT Data Construction System relies entirely on InternVL2.5-7B-Instruct as the judge that decides whether the generated CONCLUSION is semantically equivalent to the ground truth, but the paper provides no independent validation of this judge's accuracy. The entire training pipeline consumes the filtered COT data, and the paper attributes part of the improvement to this data, yet there is no check that the judge's accept/reject decisions agree with human judgment or with a stronger model, no estimate of judge false-positive/false-negative rates, and no quantitative analysis of what was filtered beyond four example images. A concrete test would be to have the judge labels verified on a random sample, or to compare training on judge-accepted COT data against training on COT data filtered by a different procedure. Without such a check, the claim that the COT data, rather than the original labels or the extra fine-tuning compute, is what drives the improvement is not fully supported.
  4. [§4.1, §4.3, Table 6] The main hyperparameters of the method, including the Stage 4 clipping range epsilon=0.1, the Stage 2 training length of 10 iterations, the Stage 4 training length of 500 iterations, and the specific COT prompt, are all selected empirically on EmoSet-118K, and the final evaluation is reported on that same benchmark. Because the test set appears to have been used for checkpoint selection (see first comment), the generalization claim is not supported by a true held-out evaluation. Even if the authors argue that most tuning was done on the evaluation split, the paper should explicitly state which splits were used for hyperparameter decisions, which split was used for checkpoint selection, and which split was used exactly once for the reported final numbers. Preferably, the method should be evaluated on an additional emotion-recognition or other abstract-semantic dataset to demonstrate that the recipe is not overfit to this benchmark.
minor comments (6)
  1. [§2.2, reference [29]] The Dr.GRPO method is cited as [29], but the listed reference is "Understanding R1-zero-like training: A critical perspective," which is not the original Dr.GRPO paper; the correct citation for Dr.GRPO ("Demystifying Long Chain-of-Thought Reasoning in LLMs") should be used.
  2. [Throughout] There are several typos and formatting inconsistencies: "RESASONING" in Figure 1 should be "REASONING," "benifit" in Section 4.3 should be "benefit," "F ear" appears in Tables 2 and 3 as a column name, and the baseline model is written both as "InternVL2.5-1B" and "InternVL-2.5-1B." These should be corrected.
  3. [§4.2, paragraph 1] The text says "the second row to tuning-based experiments," but Tables 2 and 3 contain multiple tuning-based rows; the intended meaning appears to be "the following rows" or "the tuning-based rows." Please rephrase.
  4. [Table 1] The table uses the abbreviation "Eval" without defining it in the caption or in Section 4.1. Since the distinction between evaluation data and test data is critical for the checkpoint-selection concern, the paper should explicitly define the evaluation set and state that it is disjoint from the test set.
  5. [Tables 4 and 5] The row label "InternVL2.5-1B(stage3)%" appears to be a typo; it should probably be "InternVL2.5-1B(stage3) w/o Stage2" or similar. As written, the percent sign makes the label confusing.
  6. [§4.1] Training details are given, but the paper does not state whether the COT data construction system, its judge prompts, or the training code will be released. If the recipe is meant to be reproducible, releasing the generated COT data and the training configuration would be valuable; otherwise the empirical claims are hard to verify.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are direct empirical measurements, not derivations that reduce to their inputs.

full rationale

The paper is an empirical post-training study rather than a formal derivation. The four-stage pipeline (SFT, format-reward GRPO, accuracy-plus-format GRPO, and CLGRPO with epsilon 0.1) is validated by measured accuracy and recall on EmoSet-118K. The Self-Supervised COT Data Construction System uses external LVLMs and a judge model to generate and filter data; this is a data-generation procedure, not a fitted parameter renamed as a prediction. CLGRPO's epsilon reduction is presented as an empirical design choice (Section 3.2: "Based on empirical experiments, we adjust epsilon_l and epsilon_h from 0.2 in the stage 3 to 0.1"), and its effect is tested in ablation Tables 6 and 7. No equation or definition makes a reported result equal to an input by construction. The one concerning passage is Section 4.2: "The two best-performing weights on the test data are listed in Table. 2, 3," together with Figure 5's statement that metrics were "evaluated in the test data during the training process of Stage 4." That is a test-set selection and evaluation-protocol risk, and Section 4.1's wording about "evaluation data" is inconsistent, but it is not circularity: selecting a checkpoint by test accuracy may bias the measured value yet does not make the result equivalent to an input. The authors' self-citations appear only as background related work, not as load-bearing support for the central claim. Section 5 openly acknowledges the single-task, single-model scope, reinforcing that the claims are empirical measurements. Therefore no circular step is present; the appropriate score is 0.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The method is empirical, with several hyperparameters tuned on the target benchmark. The main load-bearing assumption is that the COT data generated by larger LVLMs and verified by a judge model is accurate and beneficial for the small model.

free parameters (7)
  • epsilon_l = 0.1
    Lower clip bound for CLGRPO in Stage 4, chosen empirically from 0.2; central to the proposed method.
  • epsilon_h = 0.1
    Upper clip bound for CLGRPO in Stage 4, chosen empirically; ablation against 0.28 shows lower is better on this benchmark.
  • beta (KL penalty) = 0.04
    KL penalty coefficient in GRPO objective, fixed for all RL experiments.
  • RL learning rate = 1e-6
    Learning rate for all RL stages, chosen by the authors.
  • Stage 2 iterations = 10
    Number of GRPO iterations for format alignment; small amount, determined by experiments.
  • Stage 4 iterations = 500
    Number of CLGRPO iterations; chosen empirically.
  • Group size = 8
    Number of sampled outputs per question in GRPO.
assumptions (4)
  • domain assumption Chain-of-thought data improves small VLM reasoning
    The entire method assumes that training on LLM-generated COT data will improve a small model's reasoning, which is a premise not proven here.
  • domain assumption Judge model correctly verifies COT conclusions
    The self-supervised data construction uses InternVL2.5-7B to judge whether the generated conclusion matches the ground truth; its reliability is assumed.
  • domain assumption EmoSet-118K labels are accurate ground truth
    The dataset's emotion labels are taken as correct for reward computation and evaluation.
  • standard math GRPO objective (Eq. 1) is the standard RL update
    The paper uses the standard GRPO loss with clipping and KL penalty from prior work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CLGRPO: Reasoning Ability Enhancement for Small VLMs." pith.science (2026). https://pith.science/paper/3WGAYBYJ

@misc{pith2026250618048,
  author       = {Pith},
  title        = {Pith review of: CLGRPO: Reasoning Ability Enhancement for Small VLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3WGAYBYJ}},
  note         = {Machine review of arXiv:2506.18048}
}
read the original abstract

Small Vision Language Models (SVLMs) generally refer to models with parameter sizes less than or equal to 2B. Their low cost and power consumption characteristics confer high commercial value. However, their reasoning abilities are limited by the number of parameters. To address this issue, this paper proposes a post-training optimization paradigm called the Incremental Training Strategy to enhance the reasoning ability of SVLMs. Firstly, we constructed a Self-Supervised Chain-of-Thought (COT) Data Construction System, which leverages multiple LVLMs with 7B parameters or more to transform original data into COT data in a self-supervised manner. Our proposed Incremental Training Strategy consists of four stages. Stage 1 injects domain knowledge by performing Supervised Fine-Tuning (SFT) to the pretrained model on the COT data. Stage 2 aligns the COT data format by conducting a small amount of Group Relative Policy Optimization (GRPO) training constrained only by format rewards on the COT data. Stage 3 enhances reasoning ability by applying GRPO training on the COT data with constraints on both format and accuracy rewards. The resulting model shows significant improvement compared to the baseline. Stage 4 addresses the limited capacity of the SVLMs and the weak ability to capture complex patterns by proposing ClipLow GRPO (CLGRPO) to constrain the capture space of the training process. We conducted extensive comparative and ablation experiments on the abstract semantic recognition dataset EMOSet-118K. Experimental results demonstrate that our method significantly improves the reasoning ability of 1B SVLM. Compared to the baseline model fine-tuned on the original data, accuracy increased by 2.77 and recall by 0.69, achieving performance comparable to that of 8B models.

Figures

Figures reproduced from arXiv: 2506.18048 by the authors.

Figure 1
Figure 1. Overview of Self-Supervised COT Data Construction System and our Incremental Training Strategy. Green is SUMMARY part, Blue is CAPTION part, Purple is RESASONING part, Red is CONCLUSION part. Models (SVLMs) with parameter sizes less than or equal to 2B [24] [25]. Many perspectives suggest that the COT ability of SVLMs is constrained by their number of parameters and knowledge density, making it difficult to fully ac… view at source ↗
Figure 2
Figure 2. Failure cases of Self-Supervised COT Data Construction System. The label bellow image above is ”Ground Truth- [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. COT Data produced by our Self-Supervised COT Data Construction System. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Confusion matrix of InternVL-2.5-8B-Instruct SFT on ORI Data and our CLGRPO result train on COT Data. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Accuracy and recall metrics results during Stage4 training process compared with results of Stage3. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 11 canonical work pages

  1. [1]

    Qwen2.5-vl technical report,

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang, H. Zhong, Y . Zhu, M. Yang, Z. Li, J. Wan, P. Wang, W. Ding, Z. Fu, Y . Xu, J. Ye, X. Zhang, T. Xie, Z. Cheng, H. Zhang, Z. Yang, H. Xu, and J. Lin, “Qwen2.5-vl technical report,”arXiv preprint arXiv:2502.13923, 2025

    arXiv 2025

  2. [2]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution,

    P. Wang, S. Bai, S. Tan, S. Wang, Z. Fan, J. Bai, K. Chen, X. Liu, J. Wang, W. Ge, Y . Fan, K. Dang, M. Du, X. Ren, R. Men, D. Liu, C. Zhou, J. Zhou, and J. Lin, “Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution,”arXiv preprint arXiv:2409.12191, 2024

    arXiv 2024

  3. [3]

    Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond,

    J. Bai, S. Bai, S. Yang, S. Wang, S. Tan, P. Wang, J. Lin, C. Zhou, and J. Zhou, “Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond,”arXiv preprint arXiv:2308.12966, 2023

    arXiv 2023

  4. [4]

    Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models,

    J. Zhu, W. Wang, Z. Chen, Z. Liu, S. Ye, L. Gu, H. Tian, Y . Duan, W. Su, J. Shao, Z. Gao, E. Cui, X. Wang, Y . Cao, Y . Liu, X. Wei, H. Zhang, H. Wang, W. Xu, H. Li, J. Wang, N. Deng, S. Li, Y . He, T. Jiang, J. Luo, Y . Wang, C. He, B. Shi, X. Zhang, W. Shao, J. He, Y . Xiong, W. Qu, P. Sun, P. Jiao, H. Lv, L. Wu, K. Zhang, H. Deng, J. Ge, K. Chen, L. W...

    arXiv 2025

  5. [5]

    Mini-internvl: a flexible-transfer pocket multi-modal model with 5% parameters and 90% performance,

    Z. Gao, Z. Chen, E. Cui, Y . Ren, W. Wang, J. Zhu, H. Tian, S. Ye, J. He, X. Zhuet al., “Mini-internvl: a flexible-transfer pocket multi-modal model with 5% parameters and 90% performance,”Visual Intelligence, vol. 2, no. 1, pp. 1–17, 2024

    work page 2024

  6. [6]

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

    Z. Chen, J. Wu, W. Wang, W. Su, G. Chen, S. Xing, M. Zhong, Q. Zhang, X. Zhu, L. Luet 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, 2024, pp. 24 185–24 198

    2024

  7. [7]

    Deepseek-v3 technical report,

    DeepSeek-AI, “Deepseek-v3 technical report,” 2024. [Online]. Available: https://arxiv.org/abs/2412.19437

    arXiv 2024

  8. [8]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,

    ——, “Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,” 2025. [Online]. Available: https://arxiv.org/abs/2501.12948

    arXiv 2025

Show all 32 references
  1. [9]

    Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding,

    Z. Wu, X. Chen, Z. Pan, X. Liu, W. Liu, D. Dai, H. Gao, Y . Ma, C. Wu, B. Wang, Z. Xie, Y . Wu, K. Hu, J. Wang, Y . Sun, Y . Li, Y . Piao, K. Guan, A. Liu, X. Xie, Y . You, K. Dong, X. Yu, H. Zhang, L. Zhao, Y . Wang, and C. Ruan, “Deepseek-vl2: Mixture-of-experts vision-langu...

    2024 arXiv

  2. [10]

    Deepseek-vl: Towards real-world vision-language understanding,

    H. Lu, W. Liu, B. Zhang, B. Wang, K. Dong, B. Liu, J. Sun, T. Ren, Z. Li, H. Yang, Y . Sun, C. Deng, H. Xu, Z. Xie, and C. Ruan, “Deepseek-vl: Towards real-world vision-language understanding,” 2024

    2024

  3. [11]

    Fastvlm: Efficient vision encoding for vision language models,

    P. K. A. Vasu, F. Faghri, C.-L. Li, C. Koc, N. True, A. Antony, G. Santhanam, J. Gabrielet al., “Fastvlm: Efficient vision encoding for vision language models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2025

    2025

  4. [12]

    Emoset: A large-scale visual emotion dataset with rich attributes,

    J. Yang, Q. Huang, T. Ding, D. Lischinski, D. Cohen-Or, and H. Huang, “Emoset: A large-scale visual emotion dataset with rich attributes,” inICCV, 2023

    2023

  5. [13]

    How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites,

    Z. Chen, W. Wang, H. Tian, S. Ye, Z. Gao, E. Cui, W. Tong, K. Hu, J. Luo, Z. Maet al., “How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites,”Science China Information Sciences, vol. 67, no. 12, p. 220101, 2024

    2024

  6. [14]

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

    Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, M. Zhang, Y . 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

    2024 arXiv

  7. [15]

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

    2025 arXiv

  8. [16]

    Mobilevlm: A fast, reproducible and strong vision language assistant for mobile devices,

    X. Chu, L. Qiao, X. Lin, S. Xu, Y . Yang, Y . Hu, F. Wei, X. Zhang, B. Zhang, X. Weiet al., “Mobilevlm: A fast, reproducible and strong vision language assistant for mobile devices,”arXiv preprint arXiv:2312.16886, 2023

    2023 arXiv

  9. [17]

    Enhancing the reasoning ability of multimodal large language models via mixed preference optimization,

    W. Wang, Z. Chen, W. Wang, Y . Cao, Y . Liu, Z. Gao, J. Zhu, X. Zhu, L. Lu, Y . Qiao, and J. Dai, “Enhancing the reasoning ability of multimodal large language models via mixed preference optimization,”arXiv preprint arXiv:2411.10442, 2024

    2024 arXiv

  10. [18]

    Flash-vl 2b: Optimizing vision-language model perfor- mance for ultra-low latency and high throughput,

    B. Zhang, S. Li, R. Tian, Y . Yang, J. Tang, J. Zhou, and L. Ma, “Flash-vl 2b: Optimizing vision-language model perfor- mance for ultra-low latency and high throughput,”arXiv preprint arXiv:2505.09498, 2025

    2025 arXiv

  11. [20]

    Direct preference optimization: Your language model is secretly a reward model,

    R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon, and C. Finn, “Direct preference optimization: Your language model is secretly a reward model,” inThirty-seventh Conference on Neural Information Processing Systems,

  12. [21]

    Bluelm: An open multilingual 7b language model,

    B. Team, “Bluelm: An open multilingual 7b language model,” https://github.com/vivo-ai-lab/BlueLM, 2023

    2023

  13. [22]

    Qwen3 technical report,

    A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lvet al., “Qwen3 technical report,” arXiv preprint arXiv:2505.09388, 2025

    2025 arXiv

  14. [23]

    Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling,

    Z. Chen, W. Wang, Y . Cao, Y . Liu, Z. Gao, E. Cui, J. Zhu, S. Ye, H. Tian, Z. Liuet al., “Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling,”arXiv preprint arXiv:2412.05271, 2024

    2024 arXiv

  15. [24]

    Visual cot: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought reasoning,

    H. Shao, S. Qian, H. Xiao, G. Song, Z. Zong, L. Wang, Y . Liu, and H. 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, vol. 37, pp. 8612–8642, 2024

    2024

  16. [25]

    Cot-vla: Visual chain-of- thought reasoning for vision-language-action models,

    Q. Zhao, Y . Lu, M. J. Kim, Z. Fu, Z. Zhang, Y . Wu, Z. Li, Q. Ma, S. Han, C. Finnet al., “Cot-vla: Visual chain-of- thought reasoning for vision-language-action models,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 1702–1713

    2025

  17. [26]

    Proximal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,”arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  18. [27]

    Direct preference optimization: Your language model is secretly a reward model,

    R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon, and C. Finn, “Direct preference optimization: Your language model is secretly a reward model,”Advances in Neural Information Processing Systems, vol. 36, pp. 53 728– 53 741, 2023

    2023

  19. [28]

    Reveal the mystery of dpo: The connection between dpo and rl algorithms,

    X. Su, Y . Wang, J. Zhu, M. Yi, F. Xu, Z. Ma, and Y . Liu, “Reveal the mystery of dpo: The connection between dpo and rl algorithms,”arXiv preprint arXiv:2502.03095, 2025

    2025 arXiv

  20. [29]

    Understanding r1-zero-like training: A critical perspective,

    Z. Liu, C. Chen, W. Li, P. Qi, T. Pang, C. Du, W. S. Lee, and M. Lin, “Understanding r1-zero-like training: A critical perspective,”arXiv preprint arXiv:2503.20783, 2025

    2025 arXiv

  21. [30]

    u-llava: Unifying multi-modal tasks via large language model,

    J. Xu, L. Xu, Y . Yang, X. Li, F. Wang, Y . Xie, Y .-J. Huang, and Y . Li, “u-llava: Unifying multi-modal tasks via large language model,”arXiv preprint arXiv:2311.05348, 2023

    2023 arXiv

  22. [31]

    Reproducibility companion paper: u-llava: Unifying multi-modal tasks via large language model,

    J. Xu, X. Wang, L. Xu, Y . Yang, X. Li, F. Wang, Y . Xie, Y .-J. Huang, Y . Li, and Y . Hu, “Reproducibility companion paper: u-llava: Unifying multi-modal tasks via large language model,” inProceedings of the 2025 International Conference on Multimedia Retrieval, 2025, pp. 1964–1967

    2025

  23. [32]

    Overcoming heterogeneous data in federated medical vision-language pre-training: A triple-embedding model selector approach,

    A. Wang, Z. Zhang, D. Wang, F. Wang, H. Hu, J. Guo, Y . Zhou, C. Pang, and S. Wen, “Overcoming heterogeneous data in federated medical vision-language pre-training: A triple-embedding model selector approach,” inProceedings of the AAAI Conference on Artificial Intelligence, vo...

    2025

  24. [2023]

    Available: https://arxiv.org/abs/2305.18290

    [Online]. Available: https://arxiv.org/abs/2305.18290

    arXiv

Pith tools

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