Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

A fully automated 'collective adversarial' loop can synthesize multimodal training data that matches or beats real data for vision-language reasoning.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A collective adversarial data-synthesis pipeline produces 20K synthetic multimodal training examples whose GRPO-trained 7B model beats several listed open-source MLLMs on reasoning benchmarks.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection A plausible new pipeline for synthetic multimodal QA data, but the central premise—faithful rendering of the image from the text prompt—is asserted, not demonstrated. the 4 major comments →

arxiv 2602.03300 v2 pith:ZCZWLVYM submitted 2026-02-03 cs.LG cs.AIcs.CLcs.CV

R1-SyntheticVL: Is Synthetic Data from Generative Models Ready for Multimodal Large Language Model?

Jingyi Zhang , Tianyi Lin , Huanjin Yao , Xiang Lan , Shunyu Liu , Jiaxing Huang This is my paper
classification cs.LG cs.AIcs.CLcs.CV
keywords synthetic datamultimodal large language modelsdata synthesiscollective intelligenceadversarial learningvisual reasoningreinforcement learningtext-to-image generation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper asks whether synthetic data from generative image models is ready to replace or supplement human-annotated data for training multimodal large language models (MLLMs). It argues yes, if generation is organized as a 'collective adversarial' loop: several MLLMs jointly write questions, answers, and visual prompts; a text-to-image model renders the images; several judge MLLMs try to solve each question; only samples with enough judge agreement are kept, and hard, contested samples are fed back to sharpen the generator. The authors build a 20K dataset this way and train R1-SyntheticVL. They report it beats all listed 7B open models on most benchmarks, including MathVista 75.6 and MMMU-Pro 42.0, and beats a 2K real-data sample head-to-head on MathVista. The point matters because it suggests a scalable, low-annotation route to training data for domains where real multimodal annotations are scarce.

Core claim

CADS is presented as the first general framework to synthesize multimodal training data end-to-end from a generative image model. It runs two cyclic phases. In CAD-Generate, four frontier MLLMs collectively analyze a seed problem's rationale, choose among four transformation strategies (numerical and parameter variation, logic reversion, auxiliary extension, isomorphic scenario transfer), and write a detailed visual prompt so the image's spatial layout, object attributes, data values, and geometric relationships match the question. In CAD-Judge, the same collective tries to solve each synthesized question, and a consensus score counts how many judges reproduce the generated answer. Instances

What carries the argument

The load-bearing object is the consensus score in Eq. 3: C = sum over judges of I(p_k = a'), the number of judge models that reproduce the generated answer exactly. This single scalar routes every sample — C=K keeps it, C=0 discards it, and 1<=C<K flags it as a high-value adversarial instance whose failure patterns are reflected into the next generation context via Adversarial Context Optimization. The second mechanism is the visual prompt generation step, which forces explicit specification of spatial layout, object attributes, data values, and geometric relations so the text-to-image model cannot drift into hallucinated visuals.

Load-bearing premise

The whole loop rests on the unverified premise that the image generator faithfully draws the exact values, positions, and relationships written in the visual prompt; if image-text misalignment is common, both the judge consensus and the training gains lose their foundation.

What would settle it

Take a random sample of MMSynthetic-20K, have human experts check each image against its question and answer; if a substantial fraction show wrong values, missing objects, or impossible geometry, the judge consensus is passing misaligned data, and removing those samples should change training outcomes. Alternatively, evaluate R1-SyntheticVL on a rendering-independent visual-reasoning benchmark that requires reading novel charts and diagrams; a drop would indicate the model learned generator-specific patterns rather than general reasoning.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Synthetic data alone at 2K scale beat a random 2K slice of a real reasoning dataset on MathVista (73.3 vs 72.2), and mixing them (74.6) beat either alone.
  • Performance scales with synthetic data size from 0.5K to 20K with no saturation, suggesting the pipeline can keep absorbing more generated data.
  • Each CADS component contributes: direct text-to-image generation gives 70.8; adding collective generation raises it to 73.0; collective judgment to 74.6; adversarial context optimization to 75.6.
  • R1-SyntheticVL, trained only on MMSynthetic-20K, reaches the best average across six benchmarks among the listed 7B open models, and the best MMMU-Pro score.
  • The ablation and scaling results together imply that the synthetic data, not the base model, is what drives the reported improvements.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if the pipeline generalizes, domain-specific multimodal data scarcity (medicine, safety-critical diagrams, rare languages) could be addressed by swapping the generator and the judge set, since the method makes no domain-specific assumptions beyond the seed data.
  • Editorial inference: the consensus-score filter is only as trustworthy as the judge models' independence; if the judges share blind spots, consistently wrong answers could pass. A testable extension is to audit failures on human-verified samples and vary the judge panel composition.
  • Editorial inference: the visual prompt explicitly encoding quantities and relations suggests a natural verification hook: parse the rendered image with object detectors or OCR and check whether the stated values actually appear, which the paper does not do.
  • Editorial inference: because the same kind of models both generate and judge the data, there is a potential for self-confirming feedback; an independent, human-labeled holdout set would clarify whether the gains reflect true visual reasoning or generator-specific patterns.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes Collective Adversarial Data Synthesis (CADS), an automated pipeline for synthesizing multimodal QA training data. In the generation phase, a collective of four MLLMs (GPT-4o, Gemini-2.5-Flash, DeepSeek-R1, Claude-4-Sonnet) generates new questions and answers from seed data using four meta-strategies, and a detailed visual prompt is rendered by Nano Banana Pro. In the judgment phase, the same four models solve the question and an exact-match consensus score (Eq. 3) is used to discard unsolvable items and to identify adversarial instances with partial consensus (Eq. 4), which are then reflected into the generation context. The pipeline produces MMSynthetic-20K, and R1-SyntheticVL is trained by GRPO on Qwen2.5-VL-7B. The paper reports the best average score in Table 2 (52.0), a high MathVista score (75.6), and favorable comparisons to real data at 2K scale (Table 4).

Significance. If the claims hold, CADS would be a significant step: a fully automated collective/adversarial synthesis loop that produces visual QA data competitive with real data for MLLM reasoning training. The paper has clear strengths: it evaluates on established public benchmarks, describes the method in sufficient detail to be implemented, includes ablations and a scaling curve, and makes the framework falsifiable through benchmark comparisons. However, the core claim rests on assumptions that are currently unverified: that Nano Banana Pro faithfully renders the constraints in the Step-3 prompts, and that the four-model consensus judges ground their answers in the rendered image rather than in the question text alone. Because code, data, and prompts are not included at submission, these assumptions cannot be checked by readers. The reported gains are also single-run, with several margins small enough to be within run-to-run variability. The result is plausible and worth pursuing, but the manuscript as submitted under-supports its strongest claims.

major comments (4)
  1. [§3.2.1, Step 3; §3.2.2, Eq. (3)] The paper's central premise is that Nano Banana Pro renders the spatial layout, object attributes, data values, and geometric relationships in the Step-3 prompt 'strictly aligned' (§3.2.1). No quantitative render-success metric or human audit is reported; Fig. 4 shows only a few selected examples. The filter in Eq. (3) is a consensus among the same collective that generated a', and it can return a' even if the image is wrong or entirely unused: a judge may derive a' from q' alone, e.g., from textual arithmetic or common-sense reasoning. Consequently, the benchmark gains in Table 2 cannot yet be attributed to visual fidelity. Please add (i) a human or automated audit of a random sample of MMSynthetic-20K reporting image-question-answer alignment rates, and (ii) a text-only control in which judges solve q' without the image; the consensus pass rate in the text-only condition quantifies how
  2. [Tables 2–4 and Fig. 3; §4.2–4.4] All headline results are single runs with no seeds, error bars, or significance tests. Several margins are likely within run-to-run noise: MathVista 75.6 vs 75.1, CharXiv-Desc 75.5 vs 75.2, and average 52.0 vs 50.3 over an uncontrolled mix of checkpoints. At minimum, report mean ± std over 2–3 independent GRPO runs for the main comparison and for the Table 4 real-vs-synthetic comparison. In addition, Table 3 is presented as a cumulative chain, not a full factorial ablation; with a single run it does not identify the individual contribution of each CADS component.
  3. [§4.4, Table 4] The comparison with real data is performed at 2K instances only, using a random sample of MM-Eureka as the real-data baseline. There is no 20K real-data GRPO baseline. The paper's conclusion that synthetic data is 'a reliable alternative to real-world data' requires a same-size comparison at the actual dataset scale used for MMSynthetic-20K, especially because Fig. 3 shows that synthetic-data gains are still increasing at 20K; real-data scaling may behave differently. Without this, the advantage over open-source real-data models in Table 2 is not an apples-to-apples comparison and does not support the readiness claim.
  4. [§3.2.2, Eq. (4); §4.2; §4.3] The treatment of the adversarial set Dadv is unclear and load-bearing. Samples with 1≤C<K are used to optimize the generation context, but the paper does not say whether such instances are also included in the final MMSynthetic-20K, in what proportion, or whether they are re-judged after context optimization. If Dadv samples are included directly, the method intentionally keeps instances on which four strong judges disagree; these may be ambiguous, mis-rendered, or otherwise defective rather than 'high-value'. Please report the size and composition of Dadv, the inclusion rule for the 20K, and a sensitivity analysis in which Dadv is excluded. Also specify the matching criterion in Eq. (3), e.g., normalized answer equivalence, or a tolerant scorer.
minor comments (5)
  1. [Title, Abstract, §5] The title and abstract use 'Multimodal Large Language Model' in the singular; the intended plural is 'Models'. The conclusion contains the typo 'Specifcailly'.
  2. [Table 1] Table 1 needs a caption and clearer row semantics. The first column label 'Qwen2.5-VL-7B' presumably denotes the base model without any synthetic training data, but this is not stated. The evaluation protocol for this comparison should also be described.
  3. [Table 3] The row labels in Table 3 are ambiguous. It appears to be a cumulative ablation, but the header 'Nano Banana Pro Collective Adversarial Data Synthesis' does not make it clear which components are active in which row. Add explicit checkmarks with component names.
  4. [Fig. 3 and §4.4] In Fig. 3, the x-axis label '0.5k' is inconsistent with '20k' capitalization; there are no error bars, and the underlying numbers are not given in a table. Also, 'perfromance' in §4.4 is a typo.
  5. [Reproducibility] The paper says code, model, and data 'will be available' but provides no artifacts or supplementary material at submission. For a data-generation paper, including the generation prompts, judge prompts, and a sample of the dataset would greatly help review and reproducibility.

Circularity Check

1 steps flagged

Partial circularity: Eq. 3's quality filter is self-consensus among the same models that generated a′, but external benchmarks keep the central claim independent.

specific steps
  1. self definitional [Sec. 3.2.2 (CAD-Judge), Eq. 3; Sec. 4.1 implementation details]
    "we adopt a group of four models, including GPT-4o, Gemini-2.5-Flash, DeepSeek-R1 and Claude-4-Sonnet, to collectively perform data generation and data judgment. ... Each judge model πk attempts to solve the synthesized question, generating a prediction pk = πk(v′, q′). ... consensus score C, which represents the number of models that successfully match the generated ground truth a′: C = Σ I(pk = a′)."

    The answer a′ used as ground truth in Eq. 3 is produced by the same four-model collective that later serves as judges (Sec. 4.1). The 'quality' filter therefore measures self-consensus: a sample passes when the judge models reproduce an answer the same collective generated. The paper's claim that CAD-Judge 'identifies and filters out the samples with multimodal misalignment and factual inconsistencies, guaranteeing the reliability' goes beyond what Eq. 3 can certify; the filter is definitional with respect to the collective's own outputs, not an independent check of image-text alignment or factual correctness. This is partial circularity, not full: downstream MathVista/MMMU-Pro evaluations are external and do not reduce to Eq. 3.

full rationale

The central benchmark results are grounded in external public benchmarks (MathVista, MathVerse, MathVision, MMMU, MMMU-Pro, CharXiv) and include a controlled comparison against MM-Eureka real data (Table 4) and ablations (Table 3). Those evaluations do not reduce to the paper's internal quality score. The one recognizably circular component is the internal data-quality signal: Eq. 3's consensus score is computed by the same four models that generated a′, so the claim that CAD-Judge 'guarantees reliability' is self-referential. This makes the 'high-quality' label partly definitional, but it does not by construction force the external benchmark results. Self-citations (R1-VL, Mulberry, surveys) are used as baselines or related work, not as load-bearing justification. No uniqueness theorem or imported ansatz is used. The unverified rendering-fidelity premise (Nano Banana Pro strictly aligning with the Step-3 prompt) is a correctness risk, not a circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central claim depends on an empirical pipeline rather than a mathematical derivation. The main pulled-in assumptions are the reliability of the text-to-image generator and the validity of self-consensus filtering, plus standard RL transfer. The hand-chosen hyperparameters (consensus thresholds, model panel, generation iterations, dataset size, GRPO settings) are not derived or swept.

free parameters (5)
  • Judge consensus thresholds = C=0 discard; C=K keep; 1<=C<K treat as adversarial
    Eq. 3-4; these thresholds define data composition and are chosen without sensitivity analysis.
  • Generation/judgment ensemble size K = 4 models (GPT-4o, Gemini-2.5-Flash, DeepSeek-R1, Claude-4-Sonnet)
    Sec. 4.1; the choice and number of models affects diversity and filtering stringency, with no ablation on K or model set.
  • Maximum generation iterations = 10
    Sec. 4.1; controls dataset size and content; no sensitivity analysis is reported.
  • Dataset size = 20,000
    MMSynthetic-20K; chosen for compute; the scaling curve only reaches 20K.
  • GRPO hyperparameters = rollout 8, global batch 128, lr 1e-6, rollout temperature 1.0
    Sec. 4.1; standard EasyR1-style defaults with no reported tuning or ablation.
axioms (4)
  • domain assumption Nano Banana Pro renders images that faithfully reflect detailed textual, spatial, and geometric constraints.
    Step 3 in Sec. 3.2.1 relies on this for visual-text alignment; no quantitative render-success rate or human audit is provided.
  • domain assumption Consensus of judge models matching the generated answer is a valid measure of data quality and solvability.
    Eq. 3; judges are part of the same model collective that produced the data, and no external or human ground-truth validation is reported.
  • domain assumption Seed data D_seed is of sufficient quality, coverage, and diversity to bootstrap useful synthesis.
    Sec. 3.2.1 starts from seed samples but their source, curation, and size are not specified.
  • domain assumption Training on synthetic data improves performance on external benchmarks via standard RL generalization.
    Sec. 4.2 assumes positive transfer without detailed analysis of distribution shift between synthetic training data and benchmark distributions.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of R1-SyntheticVL: Is Synthetic Data from Generative Models Ready for Multimodal Large Language Model?." pith.science (2026). https://pith.science/paper/ZCZWLVYM

@misc{pith2026260203300,
  author       = {Pith},
  title        = {Pith review of: R1-SyntheticVL: Is Synthetic Data from Generative Models Ready for Multimodal Large Language Model?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZCZWLVYM}},
  note         = {Machine review of arXiv:2602.03300}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

In this work, we aim to develop effective data synthesis techniques that autonomously synthesize multimodal training data for enhancing MLLMs in solving complex real-world tasks. To this end, we propose Collective Adversarial Data Synthesis (CADS), a novel and general approach to synthesize high-quality, diverse and challenging multimodal data for MLLMs. The core idea of CADS is to leverage collective intelligence to ensure high-quality and diverse generation, while exploring adversarial learning to synthesize challenging samples for effectively driving model improvement. Specifically, CADS operates with two cyclic phases, i.e., Collective Adversarial Data Generation (CAD-Generate) and Collective Adversarial Data Judgment (CAD-Judge). CAD-Generate leverages collective knowledge to jointly generate new and diverse multimodal data, while CAD-Judge collaboratively assesses the quality of synthesized data. In addition, CADS introduces an Adversarial Context Optimization mechanism to optimize the generation context to encourage challenging and high-value data generation. With CADS, we construct MMSynthetic-20K and train our model R1-SyntheticVL, which demonstrates superior performance on various benchmarks.

Figures

Figures reproduced from arXiv: 2602.03300 by Huanjin Yao, Jiaxing Huang, Jingyi Zhang, Shunyu Liu, Tianyi Lin, Xiang Lan.

Figure 1
Figure 1. Figure 1: Comparison between “directly using Nano Banana Pro” and “our proposed Collective Adversarial Data Synthesis (CADS)”. (a) Directly applying Nano Banana Pro often suffers from low data quality, limited data diversity and trivial difficulty in data synthesis of complex tasks, while (b) our proposed CADS effectively synthesizes high-quality, diverse and challenging multimodal data for MLLMs. and (3) Data Diffi… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our proposed Collective Adversarial Data Synthesis (CADS). CADS operates with two cyclic phases, including (a) Collective Adversarial Data Generation (CAD-Generate), which leverages collective knowledge to generate diverse data and (b) Collective Adversarial Data Judgment (CAD-Judge), which collaboratively assesses the quality of the synthesized data. During these two phases, CADS introduces an… view at source ↗
Figure 3
Figure 3. Figure 3: Scaling analysis of synthetic data. Applying’ Nano Banana Pro yields marginal performance improvement as shown in the second row. By introducing our CAD-Generate and CAD-Judge, we can observe a sub￾stantial enhancement in performance, which demonstrates that leveraging collective knowledge to ensure diverse gen￾eration and filter out low-quality instances largely improves the quality of the synthesized dat… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative illustrations of synthesized samples from MMSynthetic-20K. same configurations to train the models and evaluate their performance on the MathVista benchmark As shown in the first two rows in [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

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

  1. Failure-Informed Image Self-Augmentation for Multimodal Large Language Model Self-Improvement

    cs.AI 2026-08 conditional novelty 5.0

    FISA improves MLLM visual question answering by generating and filtering augmented images based on the model's own failure cases.

Reference graph

Works this paper leans on

31 extracted references · 26 linked inside Pith · cited by 1 Pith paper

  1. [1]

    L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774,

    Pith/arXiv arXiv

  2. [4]

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., et al. Qwen2. 5-vl technical report.arXiv preprint arXiv:2502.13923,

    Pith/arXiv arXiv

  3. [7]

    Is synthetic data from generative models ready for image recognition?arXiv preprint arXiv:2210.07574,

    He, R., Sun, S., Yu, X., Xue, C., Zhang, W., Torr, P., Bai, S., and Qi, X. Is synthetic data from generative models ready for image recognition?arXiv preprint arXiv:2210.07574,

    Pith/arXiv arXiv

  4. [8]

    Visual instruction tuning towards general-purpose multimodal large language model: A survey.International Journal of Computer Vision, 133 (11):8151–8189, 2025a

    Huang, J., Zhang, J., Jiang, K., Qiu, H., Zhang, X., Shao, L., Lu, S., and Tao, D. Visual instruction tuning towards general-purpose multimodal large language model: A survey.International Journal of Computer Vision, 133 (11):8151–8189, 2025a. Huang, Q., Zhang, H., Wei, R., Wang, Y ., Tang, R., Song, M., and Song, J. Syn-grpo: Self-evolving data syn- thes...

    arXiv

  5. [9]

    Uniecg: Understanding and generating ecg in one unified model.arXiv preprint arXiv:2509.18588,

    Jin, J., Wang, H., Lan, X., Li, J., Cheng, G., Li, H., and Hong, S. Uniecg: Understanding and generating ecg in one unified model.arXiv preprint arXiv:2509.18588,

    Pith/arXiv arXiv

  6. [10]

    Kimi, T., Du, A., Gao, B., Xing, B., Jiang, C., Chen, C., Li, C., Xiao, C., Du, C., Liao, C., et al. Kimi k1. 5: Scaling reinforcement learning with llms.arXiv preprint arXiv:2501.12599,

    Pith/arXiv arXiv

  7. [11]

    Gem: Empowering mllm for grounded ecg under- standing with time series and images.arXiv preprint arXiv:2503.06073,

    Lan, X., Wu, F., He, K., Zhao, Q., Hong, S., and Feng, M. Gem: Empowering mllm for grounded ecg under- standing with time series and images.arXiv preprint arXiv:2503.06073,

    arXiv

  8. [12]

    Mmr1: Enhancing multimodal reasoning with variance- aware sampling and open resources.arXiv preprint arXiv:2509.21268,

    Leng, S., Wang, J., Li, J., Zhang, H., Hu, Z., Zhang, B., Jiang, Y ., Zhang, H., Li, X., Bing, L., et al. Mmr1: Enhancing multimodal reasoning with variance- aware sampling and open resources.arXiv preprint arXiv:2509.21268,

    arXiv

  9. [13]

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

    9 R1-SyntheticVL: Is Synthetic Data from Generative Models Ready for Multimodal Large Language Model? Li, B., Zhang, Y ., Guo, D., Zhang, R., Li, F., Zhang, H., Zhang, K., Zhang, P., Li, Y ., Liu, Z., et al. Llava- onevision: Easy visual task transfer.arXiv preprint arXiv:2408.03326, 2024a. Li, H., Dong, Q., Tang, Z., Wang, C., Zhang, X., Huang, H., Huang...

    Pith/arXiv arXiv 2025

  10. [14]

    Internvl-x: Advancing and accelerating internvl series with efficient visual token compression.arXiv preprint arXiv:2503.21307,

    Lu, D., Sun, Y ., Zhang, Z., Huang, L., Zeng, J., Shu, M., and Cao, H. Internvl-x: Advancing and accelerating internvl series with efficient visual token compression.arXiv preprint arXiv:2503.21307,

    Pith/arXiv arXiv

  11. [15]

    Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts.arXiv preprint arXiv:2310.02255,

    Lu, P., Bansal, H., Xia, T., Liu, J., Li, C., Hajishirzi, H., Cheng, H., Chang, K.-W., Galley, M., and Gao, J. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts.arXiv preprint arXiv:2310.02255,

    Pith/arXiv arXiv

  12. [16]

    Mathgenie: Generating synthetic data with question back-translation for enhancing mathemati- cal reasoning of llms.arXiv preprint arXiv:2402.16352,

    Lu, Z., Zhou, A., Ren, H., Wang, K., Shi, W., Pan, J., Zhan, M., and Li, H. Mathgenie: Generating synthetic data with question back-translation for enhancing mathemati- cal reasoning of llms.arXiv preprint arXiv:2402.16352,

    Pith/arXiv arXiv

  13. [17]

    Wizard- math: Empowering mathematical reasoning for large lan- guage models via reinforced evol-instruct.arXiv preprint arXiv:2308.09583,

    Luo, H., Sun, Q., Xu, C., Zhao, P., Lou, J., Tao, C., Geng, X., Lin, Q., Chen, S., and Zhang, D. Wizard- math: Empowering mathematical reasoning for large lan- guage models via reinforced evol-instruct.arXiv preprint arXiv:2308.09583,

    Pith/arXiv arXiv

  14. [18]

    Mm-eureka: Exploring visual aha moment with rule-based large-scale reinforcement learning.arXiv preprint arXiv:2503.07365,

    Meng, F., Du, L., Liu, Z., Zhou, Z., Lu, Q., Fu, D., Shi, B., Wang, W., He, J., Zhang, K., et al. Mm-eureka: Exploring visual aha moment with rule-based large-scale reinforcement learning.arXiv preprint arXiv:2503.07365,

    Pith/arXiv arXiv

  15. [19]

    Glide: Towards photorealistic image generation and editing with text-guided diffusion models.arXiv preprint arXiv:2112.10741,

    Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., and Chen, M. Glide: Towards photorealistic image generation and editing with text-guided diffusion models.arXiv preprint arXiv:2112.10741,

    Pith/arXiv arXiv

  16. [20]

    Skywork r1v: Pioneering multimodal reasoning with chain-of-thought

    Peng, Y ., Wang, X., Wei, Y ., Pei, J., Qiu, W., Jian, A., Hao, Y ., Pan, J., Xie, T., Ge, L., et al. Skywork r1v: Pioneering multimodal reasoning with chain-of-thought. arXiv preprint arXiv:2504.05599,

    Pith/arXiv arXiv

  17. [21]

    H., Fung, Y

    Qin, Z., Dong, Q., Zhang, X., Dong, L., Huang, X., Yang, Z., Khademi, M., Zhang, D., Awadalla, H. H., Fung, Y . R., et al. Scaling laws of synthetic data for language models. arXiv preprint arXiv:2503.19551,

    arXiv

  18. [23]

    Deepseekmath: Push- ing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300,

    Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y ., Wu, Y ., et al. Deepseekmath: Push- ing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300,

    Pith/arXiv arXiv

  19. [24]

    Mathcanvas: Intrin- sic visual chain-of-thought for multimodal mathematical reasoning.arXiv preprint arXiv:2510.14958,

    Shi, W., Yu, A., Fang, R., Ren, H., Wang, K., Zhou, A., Tian, C., Fu, X., Hu, Y ., Lu, Z., et al. Mathcanvas: Intrin- sic visual chain-of-thought for multimodal mathematical reasoning.arXiv preprint arXiv:2510.14958,

    arXiv

  20. [25]

    M., Hauth, A., Millican, K., et al

    Team, G., Anil, R., Borgeaud, S., Alayrac, J.-B., Yu, J., Sori- cut, R., Schalkwyk, J., Dai, A. M., Hauth, A., Millican, K., et al. Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805,

    Pith/arXiv arXiv

  21. [26]

    Treesynth: Synthesizing diverse data from scratch via tree-guided subspace partitioning.arXiv preprint arXiv:2503.17195, 2025a

    Wang, S., Chen, P., Zhou, J., Li, Q., Dong, J., Gao, J., Xue, B., Jiang, J., Kong, L., and Wu, C. Treesynth: Synthesizing diverse data from scratch via tree-guided subspace partitioning.arXiv preprint arXiv:2503.17195, 2025a. Wang, X., Yang, Z., Feng, C., Lu, H., Li, L., Lin, C.-C., Lin, K., Huang, F., and Wang, L. Sota with less: Mcts-guided sample selec...

    Pith/arXiv arXiv

  22. [27]

    Valley2: Exploring multimodal models with scalable vision-language design

    Wu, Z., Chen, Z., Luo, R., Zhang, C., Gao, Y ., He, Z., Wang, X., Lin, H., and Qiu, M. Valley2: Exploring multimodal models with scalable vision-language design. arXiv preprint arXiv:2501.05901,

    Pith/arXiv arXiv

  23. [28]

    Llava-cot: Let vision language models reason step-by- step, 2024.URL https://arxiv

    Xu, G., Jin, P., Hao, L., Song, Y ., Sun, L., and Yuan, L. Llava-cot: Let vision language models reason step-by- step, 2024.URL https://arxiv. org/abs/2411.10440. Yang, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Li, C., Liu, D., Huang, F., Wei, H., et al. Qwen2. 5 technical report.arXiv preprint arXiv:2412.15115,

    Pith/arXiv arXiv 2024

  24. [29]

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

    Yang, Y ., He, X., Pan, H., Jiang, X., Deng, Y ., Yang, X., Lu, H., Yin, D., Rao, F., Zhu, M., et al. R1-onevision: Ad- vancing generalized multimodal reasoning through cross- modal formalization.arXiv preprint arXiv:2503.10615, 2025a. Yang, Y ., Zhang, Z., Hou, Y ., Li, Z., Liu, G., Payani, A., Ting, Y .-S., and Zheng, L. Effective training data syn- the...

    Pith/arXiv arXiv

  25. [30]

    Zhan, Y ., Zhu, Y ., Zheng, S., Zhao, H., Yang, F., Tang, M., and Wang, J

    URLhttps://arxiv.org/abs/2409.02813. Zhan, Y ., Zhu, Y ., Zheng, S., Zhao, H., Yang, F., Tang, M., and Wang, J. Vision-r1: Evolving human-free alignment in large vision-language models via vision-guided rein- forcement learning.arXiv preprint arXiv:2503.18013,

    Pith/arXiv arXiv

  26. [31]

    R1-vl: Learning to reason with multimodal large language models via step-wise group relative policy optimization.arXiv preprint arXiv:2503.12937, 2025a

    Zhang, J., Huang, J., Yao, H., Liu, S., Zhang, X., Lu, S., and Tao, D. R1-vl: Learning to reason with multimodal large language models via step-wise group relative policy optimization.arXiv preprint arXiv:2503.12937, 2025a. Zhang, L., Cui, Q., Zhao, B., and Yang, C. Oasis: One image is all you need for multimodal instruction data synthesis.arXiv preprint ...

    Pith/arXiv arXiv

  27. [2020]

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

    Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al. Deepseek-r1: In- centivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    Pith/arXiv arXiv

  28. [2022]

    Flames: Improving llm math reasoning via a fine-grained anal- ysis of the data synthesis pipeline.arXiv preprint arXiv:2508.16514,

    Seegmiller, P., Mehta, K., Saha, S., Tao, C., Oraby, S., Gupta, A., Chung, T., Bansal, M., and Peng, N. Flames: Improving llm math reasoning via a fine-grained anal- ysis of the data synthesis pipeline.arXiv preprint arXiv:2508.16514,

    Pith/arXiv arXiv

  29. [2023]

    Llava- onevision-1.5: Fully open framework for democratized multimodal training.arXiv preprint arXiv:2509.23661,

    An, X., Xie, Y ., Yang, K., Zhang, W., Zhao, X., Cheng, Z., Wang, Y ., Xu, S., Chen, C., Zhu, D., et al. Llava- onevision-1.5: Fully open framework for democratized multimodal training.arXiv preprint arXiv:2509.23661,

    Pith/arXiv arXiv

  30. [2024]

    Gemini 2.5 flash image (nano banana), 2025a

    DeepMind, G. Gemini 2.5 flash image (nano banana), 2025a. URL https://aistudio.google.com/ models/gemini-2-5-flash-image. DeepMind, G. Introducing nano banana pro, 2025b. URL https://blog.google/technology/ai/ nano-banana-pro/. Deng, Y ., Bansal, H., Yin, F., Peng, N., Wang, W., and Chang, K.-W. Openvlthinker: An early exploration to complex vision-langua...

    Pith/arXiv arXiv

  31. [2025]

    anthropic.com/news/claude-4

    URL https://www. anthropic.com/news/claude-4. Bai, J., Bai, S., Chu, Y ., Cui, Z., Dang, K., Deng, X., Fan, Y ., Ge, W., Han, Y ., Huang, F., et al. Qwen technical report.arXiv preprint arXiv:2309.16609,

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.