Pith. sign in

REVIEW 3 major objections 8 minor 81 references

MIRROR claims that a vision-language model's best-performing view of a geometry problem can act as an internal teacher for its weaker views, and that training with this self-supervision improves both accuracy and cross-view consistency.

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 ↗

T0 review · deepseek-v4-flash

2026-08-01 07:06 UTC pith:RUSD2JXQ

load-bearing objection A smart, well-run paper on turning view disagreement into self-supervision; the core mechanism deserves referee time, but the reverse-KL estimator and dataset construction need to be nailed down before the claims can be fully trusted. the 3 major comments →

arxiv 2607.21552 v1 pith:RUSD2JXQ submitted 2026-07-23 cs.AI cs.LG

MIRROR: Learning from the Other View for Multi-Modal Reasoning

Wen Ye , Yuxiao Qu , Aviral Kumar , Xuezhe Ma This is my paper
classification cs.AI cs.LG
keywords multimodal reasoningvision-language modelsgeometry problem solvingreinforcement learningself-supervisioncross-view transferreverse-KL distillationpaired multimodal data
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.

MIRROR claims that cross-view inconsistency is not a bug but a signal: when a vision-language model solves a geometry problem from a diagram but not from text, or vice versa, the successful view can act as an internal teacher for the failing view. The paper builds a paired geometry dataset with text-dominant, image-dominant, and combined views of the same problems, then trains with outcome-reward reinforcement learning plus an on-policy reverse-KL regularizer that pulls restricted-view students toward the per-problem best-performing view. Across held-out paired problems and external geometry benchmarks, MIRROR improves over single-view and mixed-view RL baselines and raises the share of problems solved under both views from about 42% to 61%. If correct, the insight would let multimodal reasoning improve without external teachers, curated rationales, or extra labeled data—only the model's own disagreements across views.

Core claim

The paper's central claim is that paired views of the same geometry problem—text, diagram, and combined text+diagram—are not interchangeable prompts for a vision-language model: the same model often succeeds from one view and fails from another, and that asymmetry is itself a training signal. MIRROR selects, per problem and per training step, the view with the highest rollout success rate from the current policy, then trains the restricted text-only and image-only student views with an outcome-reward RL term plus an on-policy reverse-KL regularizer that pulls the student's own generated tokens toward the teacher view's distribution. The teacher is an exponential-moving-average copy of the po

What carries the argument

The central object is the per-problem best-view teacher in MIRROR: for each problem, the current policy samples sixteen rollouts from text, image, and combined views; the view with the highest success rate becomes that problem's teacher, and the teacher distribution is an EMA copy of the policy conditioned on that view. The student, restricted to its own view, generates rollouts from its own policy and is trained with a weighted sum of an outcome-reward RL term and a reverse-KL regularizer. The reverse-KL term rescores the student's sampled tokens under the teacher view and pulls the student's next-token distribution toward the teacher's, so transfer happens without the student imitating tra

Load-bearing premise

The load-bearing premise is that the text and image views of each ODA-Data problem carry exactly the same information, so a view's success or failure reflects reasoning differences rather than a missing detail in the prompt.

What would settle it

Take a random sample of ODA-Data image-dominant prompts, show each to a human solver with the diagram hidden, and check whether the text alone still specifies every constraint needed. If a substantial fraction are underspecified, cross-view gains could come from the image leaking information the text lacks, not from reasoning transfer.

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

If this is right

  • If MIRROR is right, cross-view disagreement in a single model is a sufficient training signal; no external teacher, curated rationale, or added labels are needed to improve multimodal reasoning.
  • Simply training on both views with outcome-reward RL is not enough—mixed-modality RL underperforms single-view RL—so explicit transfer direction matters.
  • Per-problem adaptive teacher selection beats any fixed teacher view, implying supervision source should depend on the problem rather than the modality.
  • The method's roughly 2K-example training budget is orders of magnitude smaller than common post-training corpora, so paired views may offer a data-efficient route to better visual reasoning.
  • Training expands the set of solvable problems and raises consistency across views (both-view solvability rises from 42.5% to 60.7%), not just aggregate accuracy.

Where Pith is reading between the lines

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

  • Because the only supervision is the model's own disagreement across paired views, the general principle should transfer beyond geometry to any domain where equivalent views can be constructed—tables vs. charts, code vs. prose, diagrams vs. formal statements—provided the views are truly information-equivalent.
  • The construction of ODA-Data is where the claim is most exposed: if the LLM judge silently drops a condition from the image-dominant prompt, MIRROR's teacher selection could be rewarding information leakage rather than reasoning transfer; a human audit of prompt-image pairs would settle this.
  • One testable extension: applying MIRROR to naturally paired views (e.g., a chart with its underlying data table) and checking whether gains depend on view-symmetric solvability would test whether the mechanism generalizes beyond geometry.
  • The reverse-KL regularizer could be made finer-grained by aligning at intermediate reasoning steps or by weighting teacher selection with uncertainty instead of binary success; the current design uses only rollout rewards and coarse per-problem teacher choice.

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

3 major / 8 minor

Summary. This paper studies modality-dependent reasoning in vision-language models (VLMs) on geometry problems. The authors construct ODA-Data, a paired dataset with text-dominant, image-dominant, and combined image+text views, and propose MIRROR, an RL method that selects the strongest-performing view per problem as a teacher and trains restricted-view students with GRPO plus an auxiliary reverse-KL regularizer. On ODA-Val, GeoInt, and MathVerse, MIRROR improves over the base model and over standard GRPO baselines, and it increases both-view solvability on ODA-Train. The central claim is that cross-view inconsistency, without external teachers or curated rationales, can serve as a self-supervision signal for more accurate and consistent multimodal reasoning.

Significance. The idea of turning view disagreement into a training signal is appealing and potentially useful. The paper includes several thoughtful controls: a compute-matched comparison, fixed-teacher ablations, a λ-KL sweep, and an EMA-stability analysis showing that the moving-average teacher prevents collapse. If the mechanism is correct, MIRROR would be a practical method for improving multimodal consistency using only paired views. However, the central loss definition in Eq. (2) as written has a zero-mean gradient concern, and the dataset construction's information-equivalence premise is not validated with human audit. The paper does not release code, so the exact implemented loss cannot be independently checked. These issues are load-bearing for the claimed contribution.

major comments (3)
  1. [Section 5, Eq. (2)] Eq. (2) defines L_rKL = E_{y~πθ}[log πθ(y|x_m) - log π_{θ̄}(y|x_{j*})]. The text says gradients are applied only through the student-view log-probabilities and the teacher-view log-probabilities are stop-gradient. Under the standard on-policy sample estimator, this gives a per-sample gradient of ∇θ log πθ(y) (or its negative), whose expectation under y~πθ is zero by the score-function identity. A non-zero expected gradient requires a REINFORCE-style correction, e.g., weighting the log-ratio by stop_gradient(log πθ(y) - log π_teacher(y)) or using a baseline. The manuscript does not state such a correction, and Algorithm 1 line 29 only restates the expectation. As written, MIRROR reduces to GRPO plus zero-mean noise, and the λ-dependent effects in Table 4 and Figure 5 are consistent with noise that can destabilize when scaled. Please provide the exact loss implementation and a gradient der
  2. [Appendix C, ODA-Data construction] The method's premise is that text-dominant and image-dominant views are information-equivalent. The construction pipeline uses Gemini-3-Pro to generate TikZ diagrams, an LLM judge to verify semantic alignment, and an LLM prompt to remove from the text any information 'already presented in the diagram.' No human audit, no judge error-rate, and no examples of failed verification or failed rewrites are reported. If the judge silently drops a constraint that is not actually recoverable from the diagram, or if the TikZ compilation alters the geometry, the 'same problem' premise fails and cross-view gains conflate information leakage with reasoning transfer. Please report a human audit on a random sample, judge agreement statistics, and/or a coverage check ensuring every numerical and relational condition in the original problem remains recoverable from the image-dominant view.
  3. [Section 6, Result 3 and Figure 4] The headline consistency metric (both-view solvability rising from 42.5% to 60.7%) is reported on ODA-Train, the same data used for training and for the online teacher-selection reward. The teacher selection in Section 5 uses the same final-answer reward that defines all reported pass@k metrics, so part of the measured gain is the training objective doing its own job. Please report both-view solvability on ODA-Val (where problem identity is preserved) and, if possible, on external paired-view benchmarks, and analyze whether the consistency gain is attributable to the teacher signal rather than simply to improved marginal accuracy under each view.
minor comments (8)
  1. [Section 5, Eq. (3)] Typo: 'begining' should be 'beginning'.
  2. [Appendix D.3] In the summary, 'asnwer' should be 'answer'.
  3. [Algorithm 1] Caption says 'MIRRORupdate loop' — missing space: 'MIRROR update loop'.
  4. [Section 6, Result 1] 'consistent withray interference' should be 'consistent with ray interference'.
  5. [Table 2] The Vero-8B row has '71.7367.3077.4269.33', which appears to be several numbers run together; please format columns properly.
  6. [Section 6, Result 1] 'Figure 2 and 3' should be 'Figures 2 and 3'.
  7. [Section 5, Eq. (2)] The relationship between the per-token sum in Eq. (2) and the KL divergence D_KL(πθ||πT) is not made explicit for autoregressive policies; please define the notation clearly.
  8. [Algorithm 1, line 23] The algorithm excludes examples where the selected teacher is the current view from the reverse-KL loss, but this exclusion is not mentioned in the main text or reflected in Eq. (2); please reconcile.

Circularity Check

0 steps flagged

No significant circularity: MIRROR's teacher signal is an on-policy self-distillation objective evaluated on held-out and external benchmarks, not a fitted quantity renamed as a prediction.

full rationale

The paper's derivation chain is not circular. MIRROR's objective (Eqs. 1-2) is a GRPO loss plus an on-policy reverse-KL regularizer, with the teacher being an EMA copy of the same model conditioned on the selected best-performing view. The teacher selection uses the same final-answer reward as the RL term, but the advertised predictions - ODA-Val pass@k, GeoInt pass@16, MathVerse judge score - are held-out or external metrics, so no fitted parameter is renamed as a prediction. Training-reward comparisons (Figs. 2-3) are optimization diagnostics, not the load-bearing evidence, and are supplemented by compute-matched comparisons and external transfer. ODA-Data is deliberately filtered for modality-asymmetric examples and the paper explicitly labels ODA-Val as a diagnostic for cross-view asymmetry; that is a dataset-design choice, not a definitional equivalence between input and output. The information-equivalence premise rests on an LLM-based verification/rewriting procedure, not on construction; any failure would be a data-quality threat, not circularity. Self-citations ([32],[33]) name prior phenomena but the empirical claims are supported by the paper's own experiments, so they are not load-bearing. No uniqueness theorem is imported, no ansatz is smuggled via citation, and no known result is merely renamed. The reverse-KL term as displayed is an expectation whose gradient is the standard nonzero policy-gradient form; the skeptic's zero-gradient concern is a correctness/implementation question, not a reduction of Eq. (2) to the input.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

Everything the central claim rests on: (1) the ODA-Data construction — an unaudited proprietary-LLM judge is the sole guarantor of diagram fidelity and view information-equivalence, and the pool is filtered for the phenomenon under study; (2) the reward — a single final-answer correctness signal with imperfect extraction; (3) the training configuration — λKL and α are chosen values, K=16 rollouts drive teacher selection; (4) standard benchmark ground truths. The method itself introduces no new theoretical entities; its claims reduce to engineering choices plus the data premise. The most consequential ledger entries are the judge-based view-equivalence assumption and the two tuned coefficients (λKL=0.01, α=0.99) that the headline results depend on.

free parameters (5)
  • reverse-KL coefficient λKL = 0.01
    Selected by rank on ODA-Val/GeoInt/MathVerse after a sweep at 100 steps (Table 4); λ=0.1 collapses training, 0.001–0.05 are worse on most metrics. Headline runs use the tuned value and the tuning set overlaps the reported metrics.
  • EMA decay α = 0.99
    Chosen by hand (Section 5 / Table 5); no ablation on α beyond comparing EMA vs current-policy teacher (Figure 5).
  • rollouts per prompt K = 16
    Used both for GRPO and for per-problem teacher selection (Section 5); not ablated. Teacher selection thus has a noisy success-rate estimate with K=16.
  • GRPO hyperparameters (clip 0.2/0.26, ref-KL 0.001, entropy 0.001, lr 1e-6, temp 0.8, 200 steps) = see Table 5
    Standard-ish RL bundle, but tuned for the setup; results are reported for this configuration.
  • dataset difficulty filter = remove pass@5 == 1
    Ad hoc filtering of ODA-Math-460k down to 16K candidates then ~2K curated; the threshold choice shapes the distribution and is not motivated independently.
axioms (5)
  • domain assumption Final-answer correctness is a faithful binary reward for these geometry problems, including reliable answer extraction.
    Invoked in Section 3 ('we assume that correctness can be evaluated from the final answer') and throughout training/eval; corrupted final answers in Appendix D (e.g., '3249✗') show extraction is imperfect.
  • domain assumption The three views of each ODA-Data problem are information-equivalent; Gemini-3-Pro's diagram generation, fidelity judgment, and text-rewrite steps are accurate.
    Appendix C: diagrams kept only if TikZ compiles and the LLM judge says 'Accurate'; prompts rewritten to strip diagram-redundant info. No human audit or error-rate estimate. This is the weakest load-bearing premise.
  • ad hoc to paper Problems the base model solves under exactly one view are the right training pool — base-model asymmetry predicts improvable asymmetry.
    Section 4/Appendix C filter to 'solved under one view but not the other'; Ablation 1 shows asymmetric pairs help more than non-asymmetric ones, but only post hoc on ODA-Val.
  • domain assumption K=16-rollout success-rate estimates identify the best teacher view reliably enough, and EMA of the policy is a stable teacher.
    Section 5/Algorithm 1; ties broken randomly; Figure 5 empirically justifies EMA, but teacher-selection noise is not characterized.
  • standard math Benchmark ground truths (GeoInt answer keys, MathVerse judge scores via gpt-4o) are accepted as correct.
    Section 6 evaluation protocol; taken as external ground truth rather than independently verified here.
invented entities (1)
  • ODA-Data / ODA-Val no independent evidence
    purpose: Paired multi-view geometry datasets built from ODA-Math-460k with Gemini-3-Pro-generated diagrams and rewritten prompts; used as training pool and primary validation.
    Not a postulated physical/mathematical entity; included because it is a new manufactured artifact whose independent verifiability is currently absent — no release URL, no human-audited subset, so a reader cannot check the view-equivalence claims from the paper alone.

pith-pipeline@v1.3.0-alltime-deepseek · 22865 in / 26995 out tokens · 250073 ms · 2026-08-01T07:06:54.647545+00:00 · methodology

0 comments
read the original abstract

Unlike large language models (LLMs) that exhibit strong reasoning capabilities, vision-language models (VLMs) struggle with visual reasoning, even on geometry problems that admit equivalent text, diagram, and combined diagram+text views. We show that these views often elicit different behaviors: a model may solve a problem from text but fail on the corresponding diagram, or succeed visually while failing textually. This inconsistency suggests that different views expose complementary reasoning paths and failure modes that standard multimodal post-training does not fully exploit. To study and exploit this phenomenon, we construct ODA-Data, a high-quality paired multimodal geometry dataset with text-dominant, image-dominant, and combined image+text views of the same problems, together with splits for training and evaluating modality-dependent reasoning behaviors. We then develop Modality-Informed Reciprocal Reasoning Optimization (MIRROR), a reinforcement learning approach for improving multimodal reasoning via self supervision. For each problem, MIRROR evaluates the model under all views, selects the best-performing view as a teacher, and trains other views with a reverse-KL objective towards the teacher. Across reasoning benchmarks that evaluate on geometry problems, MIRROR improves over standard RL and yields more accurate and consistent behavior across modalities

Figures

Figures reproduced from arXiv: 2607.21552 by Aviral Kumar, Wen Ye, Xuezhe Ma, Yuxiao Qu.

Figure 1
Figure 1. Figure 1: Modality-Informed Reciprocal Reasoning Optimization (MIRROR). MIRROR exploits view asymmetry by selecting the strongest-performing view of each problem as a teacher and regularizing weaker student views towards the teacher distribution, improving performance without external supervision. Prior work on multimodal reasoning often improves models by using external teachers [27], curated rationales [59], or sp… view at source ↗
Figure 2
Figure 2. Figure 2: Single problem analysis. Across both views, MIRROR achieves the highest training reward, followed by GRPO trained on the corresponding single modality, while naive GRPO over mixed modality views performs worst. This shows that simply exposing RL to multiple views of the same underlying problems is not sufficient for reliable cross-modal transfer; without an explicit transfer direction, the sparse reward si… view at source ↗
Figure 3
Figure 3. Figure 3: Mixed-modality GRPO does not automatically transfer reasoning across views. With full-dataset training, mixed-view GRPO underperforms GRPO trained directly on the corresponding restricted view, while MIRROR obtains the strongest training reward by using paired views directionally. Mixed-modality GRPO is the most direct test of whether paired views are sufficient by them￾selves. Although this baseline train… view at source ↗
Figure 4
Figure 4. Figure 4: Net solvability gain after training. Each bar shows the change in the fraction of solved examples relative to the base model across 3 random seeds. Each random seed samples 309 problems which is the same size as the validation set. Error bars indicate the min and max values across seeds. MIRROR produces the largest net increase. Result 3: MIRROR expands solvability and improves cross-view consistency. Aggr… view at source ↗
Figure 5
Figure 5. Figure 5: The EMA teacher prevents long-horizon collapse. Training MIRROR with the teacher scored by the current policy (no EMA) vs. the EMA teacher (α=0.99), otherwise identical. Left: policy entropy. Middle: reference-policy KL. Right: training reward averaged across image and text modalities. The current-policy teacher destabilizes after ∼90 steps and collapses by step ∼170; the EMA teacher stays stable and reach… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

81 extracted references · 44 linked inside Pith

  1. [1]

    On-policy distillation of language models: Learning from self- generated mistakes, 2024

    Rishabh Agarwal, Nino Vieillard, Yongchao Zhou, Piotr Stanczyk, Sabela Ramos, Matthieu Geist, and Olivier Bachem. On-policy distillation of language models: Learning from self- generated mistakes, 2024. URLhttps://arxiv.org/abs/2306.13649

    Pith/arXiv arXiv 2024

  2. [2]

    Lawrence Zitnick, Dhruv Batra, and Devi Parikh

    Aishwarya Agrawal, Jiasen Lu, Stanislaw Antol, Margaret Mitchell, C. Lawrence Zitnick, Dhruv Batra, and Devi Parikh. Vqa: Visual question answering, 2016. URL https://arxiv. org/abs/1505.00468

    Pith/arXiv arXiv 2016

  3. [3]

    Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. URLhttps://arxiv.org/abs/2308.12966

    Pith/arXiv arXiv 2023

  4. [4]

    Qwen3-vl technical report, 2025

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng Lin, Junyang Lin, Xuejing Liu, Jiawei Liu, Chenglong Liu, Yang Liu, Dayiheng Liu, Shixuan ...

    Pith/arXiv arXiv 2025

  5. [5]

    V old: Reasoning transfer from llms to vision-language models via on-policy distillation, 2025

    Walid Bousselham, Hilde Kuehne, and Cordelia Schmid. V old: Reasoning transfer from llms to vision-language models via on-policy distillation, 2025. URL https://arxiv.org/abs/ 2510.23497

    Pith/arXiv arXiv 2025

  6. [6]

    Opendataarena: A fair and open arena for benchmarking post-training dataset value.arXiv preprint arXiv:2512.14051, 2025

    Mengzhang Cai, Xin Gao, Yu Li, Honglin Lin, Zheng Liu, Zhuoshi Pan, Qizhi Pei, Xiao- ran Shang, Mengyuan Sun, Zinan Tang, et al. Opendataarena: A fair and open arena for benchmarking post-training dataset value.arXiv preprint arXiv:2512.14051, 2025

    arXiv 2025

  7. [7]

    Milestones over outcome: Unlocking geometric reasoning with sub-goal verifiable reward, 2026

    Jianlong Chen, Daocheng Fu, Shengze Xu, Jiawei Chen, Yuan Feng, Yue Yang, Junchi Yan, Hongyuan Zha, and Renqiu Xia. Milestones over outcome: Unlocking geometric reasoning with sub-goal verifiable reward, 2026. URLhttps://arxiv.org/abs/2601.05073

    arXiv 2026

  8. [8]

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian...

    Pith/arXiv arXiv 2021

  9. [9]

    Unitymas-o: A general rl optimization framework for llm-based multi-agent systems.arXiv preprint arXiv:2605.26646, 2026

    Yiqun Chen, Wei Yang, Erhan Zhang, Shijie Wang, Qi Liu, Zechun Niu, Bin Zhang, Haitao Li, Rui Li, Lingyong Yan, et al. Unitymas-o: A general rl optimization framework for llm-based multi-agent systems.arXiv preprint arXiv:2605.26646, 2026

    Pith/arXiv arXiv 2026

  10. [10]

    Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks, 2024

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks, 2024. URL https://arxiv.org/abs/2312.14238

    Pith/arXiv arXiv 2024

  11. [11]

    Plane geometry diagram formalization via vision-language models

    Xiaoteng Cui and Yi Liu. Plane geometry diagram formalization via vision-language models. In2nd AI for Math Workshop@ ICML 2025, 2025. 11

    2025

  12. [12]

    Addressing overthinking in large vision-language models via gated perception-reasoning optimization, 2026

    Xingjian Diao, Zheyuan Liu, Chunhui Zhang, Weiyi Wu, Keyi Kong, Lin Shi, Kaize Ding, Soroush V osoughi, and Jiang Gui. Addressing overthinking in large vision-language models via gated perception-reasoning optimization, 2026. URL https://arxiv.org/abs/2601. 04442

    2026

  13. [13]

    Insight-v: Exploring long-chain visual reasoning with multimodal large language models, 2025

    Yuhao Dong, Zuyan Liu, Hai-Long Sun, Jingkang Yang, Winston Hu, Yongming Rao, and Ziwei Liu. Insight-v: Exploring long-chain visual reasoning with multimodal large language models, 2025. URLhttps://arxiv.org/abs/2411.14432

    Pith/arXiv arXiv 2025

  14. [14]

    Gemini 3 pro model card

    Google DeepMind. Gemini 3 pro model card. https://storage.googleapis. com/deepmind-media/Model-Cards/Gemini-3-Pro-Model-Card.pdf , December 2025. Model card for Gemini 3 Pro

    2025

  15. [15]

    Bootstrap your own latent-a new approach to self-supervised learning.Advances in neural information processing systems, 33:21271–21284, 2020

    Jean-Bastien Grill, Florian Strub, Florent Altch ´e, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning.Advances in neural information processing systems, 33:21271–21284, 2020

    2020

  16. [16]

    Minillm: On-policy distillation of large language models, 2026

    Yuxian Gu, Li Dong, Furu Wei, and Minlie Huang. Minillm: On-policy distillation of large language models, 2026. URLhttps://arxiv.org/abs/2306.08543

    Pith/arXiv arXiv 2026

  17. [17]

    Momentum contrast for unsupervised visual representation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729–9738, 2020

    2020

  18. [18]

    Rae, Oriol Vinyals, and Laurent Sifre

    Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland, Katie Millican, George van den Driessche, Bogdan Damoc, Aurelia Guy, Simon Osindero, Karen Simonyan, Erich Elsen, Jack W. Rae, Oriol Vinyals, and Laurent Sifre...

    Pith/arXiv arXiv 2022

  19. [19]

    Uni-opd: Unifying on-policy distillation with a dual-perspective recipe, 2026

    Wenjin Hou, Shangpin Peng, Weinong Wang, Zheng Ruan, Yue Zhang, Zhenglin Zhou, Mingqi Gao, Yifei Chen, Kaiqi Wang, Hongming Yang, Chengquan Zhang, Zhuotao Tian, Han Hu, Yi Yang, Fei Wu, and Hehe Fan. Uni-opd: Unifying on-policy distillation with a dual-perspective recipe, 2026. URLhttps://arxiv.org/abs/2605.03677

    Pith/arXiv arXiv 2026

  20. [20]

    Vision-r1: Incentivizing reasoning capability in multimodal large language models, 2026

    Wenxuan Huang, Bohan Jia, Zijie Zhai, Shaosheng Cao, Zheyu Ye, Fei Zhao, Zhe Xu, Xu Tang, Yao Hu, and Shaohui Lin. Vision-r1: Incentivizing reasoning capability in multimodal large language models, 2026. URLhttps://arxiv.org/abs/2503.06749

    Pith/arXiv arXiv 2026

  21. [21]

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

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

    Pith/arXiv arXiv 2024

  22. [22]

    Reinforcement learning via self-distillation.arXiv preprint arXiv:2601.20802, 2026

    Jonas H ¨ubotter, Frederike L ¨ubeck, Lejs Behric, Anton Baumann, Marco Bagatella, Daniel Marta, Ido Hakimi, Idan Shenfeld, Thomas Kleine Buening, Carlos Guestrin, and Andreas Krause. Reinforcement learning via self-distillation.arXiv preprint arXiv:2601.20802, 2026

    Pith/arXiv arXiv 2026

  23. [23]

    Promptception: How sensitive are large multimodal models to prompts?, 2025

    Mohamed Insaf Ismithdeen, Muhammad Uzair Khattak, and Salman Khan. Promptception: How sensitive are large multimodal models to prompts?, 2025. URL https://arxiv.org/ abs/2509.03986

    Pith/arXiv arXiv 2025

  24. [24]

    Rethinking on-policy distillation of large language models: Phenomenology, mechanism, and recipe.arXiv preprint arXiv:2604.13016, 2026

    Yaxuan Li, Yuxin Zuo, Bingxiang He, Jinqian Zhang, Chaojun Xiao, Cheng Qian, Tianyu Yu, Huan-ang Gao, Wenkai Yang, Zhiyuan Liu, et al. Rethinking on-policy distillation of large language models: Phenomenology, mechanism, and recipe.arXiv preprint arXiv:2604.13016, 2026

    Pith/arXiv arXiv 2026

  25. [25]

    A survey of state of the art large vision language models: Alignment, benchmark, evaluations and challenges, 2025

    Zongxia Li, Xiyang Wu, Hongyang Du, Fuxiao Liu, Huy Nghiem, and Guangyao Shi. A survey of state of the art large vision language models: Alignment, benchmark, evaluations and challenges, 2025. URLhttps://arxiv.org/abs/2501.02189. 12

    Pith/arXiv arXiv 2025

  26. [26]

    Lillicrap, Jonathan J

    Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning,

  27. [27]

    Visual instruction tuning, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. URLhttps://arxiv.org/abs/2304.08485

    Pith/arXiv arXiv 2023

  28. [28]

    Act2see: Emergent active visual perception for video reasoning

    Martin Q Ma, Yuxiao Qu, Aditya Agrawal, Willis Guo, Paul Pu Liang, Ruslan Salakhutdinov, and Louis-Philippe Morency. Act2see: Emergent active visual perception for video reasoning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5455–5464, 2026

    2026

  29. [29]

    Chartqa: A benchmark for question answering about charts with visual and logical reasoning, 2022

    Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning, 2022. URL https://arxiv.org/abs/2203.10244

    Pith/arXiv arXiv 2022

  30. [30]

    Human-level control through deep reinforcement learning.nature, 518(7540):529–533, 2015

    V olodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning.nature, 518(7540):529–533, 2015

    2015

  31. [31]

    Multimath: Bridging visual and mathematical reasoning for large language models, 2024

    Shuai Peng, Di Fu, Liangcai Gao, Xiuqin Zhong, Hongguang Fu, and Zhi Tang. Multimath: Bridging visual and mathematical reasoning for large language models, 2024. URL https: //arxiv.org/abs/2409.00147

    Pith/arXiv arXiv 2024

  32. [32]

    Recursive introspection: Teaching language model agents how to self-improve.Advances in Neural Information Processing Systems, 37:55249–55285, 2024

    Yuxiao Qu, Tianjun Zhang, Naman Garg, and Aviral Kumar. Recursive introspection: Teaching language model agents how to self-improve.Advances in Neural Information Processing Systems, 37:55249–55285, 2024

    2024

  33. [33]

    Pope: Learning to reason on hard problems via privileged on-policy exploration, 2026

    Yuxiao Qu, Amrith Setlur, Virginia Smith, Ruslan Salakhutdinov, and Aviral Kumar. Pope: Learning to reason on hard problems via privileged on-policy exploration, 2026. URLhttps: //arxiv.org/abs/2601.18779

    arXiv 2026

  34. [34]

    Vero: An open rl recipe for general visual reasoning, 2026

    Gabriel Sarch, Linrong Cai, Qunzhong Wang, Haoyang Wu, Danqi Chen, and Zhuang Liu. Vero: An open rl recipe for general visual reasoning, 2026. URL https://arxiv.org/abs/ 2604.04917

    Pith/arXiv arXiv 2026

  35. [35]

    Beyond chinchilla-optimal: Accounting for inference in language model scaling laws, 2025

    Nikhil Sardana, Jacob Portes, Sasha Doubov, and Jonathan Frankle. Beyond chinchilla-optimal: Accounting for inference in language model scaling laws, 2025. URL https://arxiv.org/ abs/2401.00448

    Pith/arXiv arXiv 2025

  36. [36]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y . K. Li, Y . Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024. URL https://arxiv.org/abs/ 2402.03300

    Pith/arXiv arXiv 2024

  37. [37]

    Self-distillation enables continual learning, 2026

    Idan Shenfeld, Mehul Damani, Jonas H¨ubotter, and Pulkit Agrawal. Self-distillation enables continual learning, 2026. URLhttps://arxiv.org/abs/2601.19897

    Pith/arXiv arXiv 2026

  38. [38]

    Hybridflow: A flexible and efficient rlhf framework

    Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. Hybridflow: A flexible and efficient rlhf framework. In Proceedings of the Twentieth European Conference on Computer Systems, pages 1279–1297, 2025

    2025

  39. [39]

    Math-llava: Bootstrapping mathematical reasoning for multimodal large language models, 2024

    Wenhao Shi, Zhiqiang Hu, Yi Bin, Junhua Liu, Yang Yang, See-Kiong Ng, Lidong Bing, and Roy Ka-Wei Lee. Math-llava: Bootstrapping mathematical reasoning for multimodal large language models, 2024. URLhttps://arxiv.org/abs/2406.17294

    Pith/arXiv arXiv 2024

  40. [40]

    Leangeo: Formalizing competitional geometry problems in lean, 2025

    Chendong Song, Zihan Wang, Frederick Pu, Haiming Wang, Xiaohan Lin, Junqi Liu, Jia Li, and Zhengying Liu. Leangeo: Formalizing competitional geometry problems in lean, 2025. URLhttps://arxiv.org/abs/2508.14644

    Pith/arXiv arXiv 2025

  41. [41]

    Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged con- sistency targets improve semi-supervised deep learning results.Advances in neural information processing systems, 30, 2017. 13

    2017

  42. [42]

    Mimo-v2-flash technical report, 2026

    Core Team, Bangjun Xiao, Bingquan Xia, Bo Yang, Bofei Gao, Bowen Shen, Chen Zhang, Chenhong He, Chiheng Lou, Fuli Luo, Gang Wang, Gang Xie, Hailin Zhang, Hanglong Lv, Hanyu Li, Heyu Chen, Hongshen Xu, Houbin Zhang, Huaqiu Liu, Jiangshan Duo, Jianyu Wei, Jiebao Xiao, Jinhao Dong, Jun Shi, Junhao Hu, Kainan Bao, Kang Zhou, Lei Li, Liang Zhao, Linghao Zhang,...

    Pith/arXiv arXiv 2026

  43. [43]

    Mad-opd: Breaking the ceiling in on-policy distillation via multi-agent debate, 2026

    Jianze Wang, Ying Liu, Jinlong Chen, Xuchun Hu, Qilong Zhang, Yu Cao, Jun Wang, Hua Yang, Yong Xie, and Qianglong Chen. Mad-opd: Breaking the ceiling in on-policy distillation via multi-agent debate, 2026. URLhttps://arxiv.org/abs/2605.01347

    Pith/arXiv arXiv 2026

  44. [44]

    Measuring multimodal mathematical reasoning with math-vision dataset, 2024

    Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset, 2024. URL https://arxiv. org/abs/2402.14804

    Pith/arXiv arXiv 2024

  45. [45]

    Mathcoder-vl: Bridging vision and code for enhanced multimodal mathematical reasoning, 2025

    Ke Wang, Junting Pan, Linda Wei, Aojun Zhou, Weikang Shi, Zimu Lu, Han Xiao, Yunqiao Yang, Houxing Ren, Mingjie Zhan, and Hongsheng Li. Mathcoder-vl: Bridging vision and code for enhanced multimodal mathematical reasoning, 2025. URL https://arxiv.org/ abs/2505.10557

    Pith/arXiv arXiv 2025

  46. [46]

    Perception-aware policy optimization for multimodal reasoning.arXiv preprint arXiv:2507.06448, 2025

    Zhenhailong Wang, Xuehang Guo, Sofia Stoica, Haiyang Xu, Hongru Wang, Hyeonjeong Ha, Xiusi Chen, Yangyi Chen, Ming Yan, Fei Huang, et al. Perception-aware policy optimization for multimodal reasoning.arXiv preprint arXiv:2507.06448, 2025

    Pith/arXiv arXiv 2025

  47. [48]

    Geoint-r1: Formalizing multimodal geometric reasoning with dynamic auxiliary constructions, 2025

    Jingxuan Wei, Caijun Jia, Qi Chen, Honghao He, Linzhuang Sun, Conghui He, Lijun Wu, Bihui Yu, and Cheng Tan. Geoint-r1: Formalizing multimodal geometric reasoning with dynamic auxiliary constructions, 2025. URLhttps://arxiv.org/abs/2508.03173

    Pith/arXiv arXiv 2025

  48. [49]

    Do vision- language models truly perform vision reasoning? a rigorous study of the modality gap, 2026

    Yige Xu, Yongjie Wang, Zizhuo Wu, Kaisong Song, Jun Lin, and Zhiqi Shen. Do vision- language models truly perform vision reasoning? a rigorous study of the modality gap, 2026. URLhttps://arxiv.org/abs/2604.16256

    Pith/arXiv arXiv 2026

  49. [50]

    Learning to deliberate: Meta-policy collaboration for agentic llms with multi-agent reinforcement learning.arXiv preprint arXiv:2509.03817, 2025

    Wei Yang and Jesse Thomason. Learning to deliberate: Meta-policy collaboration for agentic llms with multi-agent reinforcement learning.arXiv preprint arXiv:2509.03817, 2025

    arXiv 2025

  50. [51]

    Maestro: Learning to collaborate via conditional listwise policy optimization for multi-agent llms.arXiv preprint arXiv:2511.06134, 2025

    Wei Yang, Jiacheng Pang, Shixuan Li, Paul Bogdan, Stephen Tu, and Jesse Thomason. Maestro: Learning to collaborate via conditional listwise policy optimization for multi-agent llms.arXiv preprint arXiv:2511.06134, 2025

    arXiv 2025

  51. [52]

    Auditing multi-agent llm reasoning trees outperforms majority vote and llm-as-judge.arXiv preprint arXiv:2602.09341, 2026

    Wei Yang, Shixuan Li, Heng Ping, Peiyu Zhang, Paul Bogdan, and Jesse Thomason. Auditing multi-agent llm reasoning trees outperforms majority vote and llm-as-judge.arXiv preprint arXiv:2602.09341, 2026

    arXiv 2026

  52. [53]

    Learning beyond teacher: Generalized on-policy distillation with reward extrapolation, 2026

    Wenkai Yang, Weijie Liu, Ruobing Xie, Kai Yang, Saiyong Yang, and Yankai Lin. Learning beyond teacher: Generalized on-policy distillation with reward extrapolation, 2026. URL https://arxiv.org/abs/2602.12125. 14

    Pith/arXiv arXiv 2026

  53. [54]

    R1-onevision: Advancing generalized multimodal reasoning through cross-modal formalization, 2025

    Yi Yang, Xiaoxuan He, Hongkun Pan, Xiyan Jiang, Yan Deng, Xingtao Yang, Haoyu Lu, Dacheng Yin, Fengyun Rao, Minfeng Zhu, Bo Zhang, and Wei Chen. R1-onevision: Advancing generalized multimodal reasoning through cross-modal formalization, 2025. URL https: //arxiv.org/abs/2503.10615

    Pith/arXiv arXiv 2025

  54. [55]

    Lens: Multi-level evaluation of multimodal reasoning with large language models, 2025

    Ruilin Yao, Bo Zhang, Jirui Huang, Xinwei Long, Yifang Zhang, Tianyu Zou, Yufei Wu, Shichao Su, Yifan Xu, Wenxi Zeng, Zhaoyu Yang, Guoyou Li, Shilan Zhang, Zichan Li, Yaxiong Chen, Shengwu Xiong, Peng Xu, Jiajun Zhang, Bowen Zhou, David Clifton, and Luc Van Gool. Lens: Multi-level evaluation of multimodal reasoning with large language models, 2025. URLhtt...

    Pith/arXiv arXiv 2025

  55. [56]

    Lmms-eval: Reality check on the evaluation of large multimodal models, 2024

    Kaichen Zhang, Bo Li, Peiyuan Zhang, Fanyi Pu, Joshua Adrian Cahyono, Kairui Hu, Shuai Liu, Yuanhan Zhang, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Lmms-eval: Reality check on the evaluation of large multimodal models, 2024. URL https://arxiv.org/abs/2407.12772

    Pith/arXiv arXiv 2024

  56. [57]

    Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems?, 2024

    Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Peng Gao, and Hongsheng Li. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems?, 2024. URL https://arxiv.org/abs/ 2403.14624

    Pith/arXiv arXiv 2024

  57. [58]

    Open eyes, then reason: Fine-grained visual mathematical understanding in mllms, 2025

    Shan Zhang, Aotian Chen, Yanpeng Sun, Jindong Gu, Yi-Yu Zheng, Piotr Koniusz, Kai Zou, Anton van den Hengel, and Yuan Xue. Open eyes, then reason: Fine-grained visual mathematical understanding in mllms, 2025. URLhttps://arxiv.org/abs/2501.06430

    Pith/arXiv arXiv 2025

  58. [59]

    Mul- timodal chain-of-thought reasoning in language models, 2024

    Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Mul- timodal chain-of-thought reasoning in language models, 2024. URL https://arxiv.org/ abs/2302.00923

    Pith/arXiv arXiv 2024

  59. [60]

    Self-distilled reasoner: On-policy self-distillation for large language models, 2026

    Siyan Zhao, Zhihui Xie, Mengchen Liu, Jing Huang, Guan Pang, Feiyu Chen, and Aditya Grover. Self-distilled reasoner: On-policy self-distillation for large language models, 2026. URLhttps://arxiv.org/abs/2601.18734

    Pith/arXiv arXiv 2026

  60. [61]

    Xing, Hao Zhang, Joseph E

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023. URL https://arxiv.org/ abs/2306.05685. 15 Appendices A FLOPs Estimation We estimate the computational overhead of M...

    Pith/arXiv arXiv 2023

  61. [63]

    A geometry problem described in text

  62. [64]

    Check carefully that:

    A diagram image intended to represent that problem Your task is to judge whether the diagram is an accurate and faithful representation of the problem text. Check carefully that:

  63. [65]

    All geometric entities mentioned in the text (points, lines, angles, shapes, objects) appear in the diagram

  64. [66]

    Spatial relationships match the text (relative positions, intersections, parallelism, perpen- dicularity, containment, motion if described)

  65. [67]

    • The marked angle corresponds to the intended angle, not its complementary or supplementary angle

    Angle markings are correct: • Angles are marked at the correct vertex. • The marked angle corresponds to the intended angle, not its complementary or supplementary angle. • The rays defining the angle match the textual description

  66. [68]

    Any stated angle measures or symbols (e.g., 90◦, 100◦, α, β) are applied to the correct geometric region

  67. [69]

    The diagram does not introduce misleading constructions, extra constraints, or alternative interpretations that contradict the text

  68. [70]

    If any required element is missing, incorrectly marked, ambiguously represented, or misleading, judge the diagram as inaccurate

    Do NOT judge based on visual scale accuracy: • Differences in segment lengths, proportions, or drawing scale (e.g., a segment of length 15 not appearing longer than one of length 13) should NOT be treated as errors unless relative length ordering is explicitly required by the text. If any required element is missing, incorrectly marked, ambiguously repres...

  69. [71]

    Removetextual descriptions of geometric configurations, diagrams, shapes, angles, side lengths, point positions, or any other information that a reader could extract from the image

  70. [72]

    Only remove properties that are explicitly marked in the image.For example, if the text says ‘isosceles triangle’ and the diagram shows a triangle that look isosceles but does NOT have explicit markings (e.g., tick marks on equal sides, or a label saying ‘isosceles’), you must KEEP ‘isosceles triangle’ in the text. The same applies to right angles (keep u...

  71. [73]

    Always keepthe final question being asked (e.g., ‘Find the length of...’, ‘What is the value of...’, ‘Determine...’)

  72. [74]

    Always keepany answer format instructions (e.g., ‘put your final answer in ’, ‘Final Answer: ’)

  73. [75]

    Keepany mathematical constraints or conditions that are purely algebraic and NOT depicted in the image

  74. [76]

    If the question is purely algebraic with no visual information to remove, return it UNCHANGED

  75. [77]

    Do NOT add any new information or rephrase the question – only remove redundant parts

  76. [78]

    Return ONLY the rewritten question text, nothing else

  77. [79]

    ## Original question:<question> ## Rewritten question (visual-redundant info removed): D Examples D.1 Example 1 Summary The correct answer is 3 √ 2 + 2 √

    Do not add placeholder tokens such as ¡image¿. ## Original question:<question> ## Rewritten question (visual-redundant info removed): D Examples D.1 Example 1 Summary The correct answer is 3 √ 2 + 2 √

  78. [80]

    The image-dominant response realizes that P should sit on the intersection of AC and BD to minimize |P A|+|P B|+|P C|+|P D|because this is intuitive from the diagram. The text-dominant response never realizes this and simply guess and exhausts different locations of P and only reaches a rough estimate of|P A|+|P B|+|P C|+|P D| ≈8.74 Image-dominant Prompt ...

  79. [81]

    A square AEF G is constructed with AE as one side, and points A and G are on the same side of line BE

    Then the length of BEis? Text-dominant Response We are given a square ABCD , and point E is on the extension of side BC. A square AEF G is constructed with AE as one side, and points A and G are on the same side of line BE. The extension of diagonal BD intersects line AF at point H, with HD= √ 2 and F H= 5 √

  80. [82]

    share a common side

    We are to find the length ofBE. [...] Final Answer:8✓ D.5 Example 5 Summary The text-dominant response treats ’Square A and B share a common side’ as side length equal and could never solve the problem. Text-dominant Prompt Square A, with side length 20 meters, has an area equal to the areas of square B and square C combined. Square A and square B share a...

Showing first 80 references.