Pith. sign in

REVIEW 4 major objections 7 minor 59 references

Reinforcing the Generation Order of Multimodal Masked Diffusion Models

T0 review · 4 major / 7 minor · reviewed 2026-07-10 · glm-5.2

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

Pith's one-line read Learning what to draw first: RL-trained order control for multimodal diffusion

desk verdict Learned generation order for multimodal diffusion models: real idea, modest gains, no variance estimates read the letter →

arxiv 2607.08056 v1 pith:GTAO4ZZ2 submitted 2026-07-09 cs.LG cs.AIstat.ML

Yidong Ouyang , Zhe Wang , Sourav Bhabesh , Dmitriy Bespalov This is my paper
classification cs.LGcs.AIstat.ML
keywords maskeddiffusionmodelsgenerationordermultimodalreinforcementlearningGRPOtext-to-imageunderstandingtokenorderingpolicy
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

Masked diffusion models can generate tokens in any order, but which order they choose matters. Prior work showed that for structured language tasks like Sudoku, you can pick the next token to reveal based on how confident the model is at each position — high-confidence tokens go first. This paper shows that the same trick fails for image generation and multimodal understanding: a token's predicted probability says nothing about whether generating it now will help the overall image composition or reasoning chain. Instead, the authors train a separate neural module — a control block — that learns to rank which tokens to unmask at each denoising step. They train it with Group Relative Policy Optimization, a reinforcement learning method that compares groups of sampled generation trajectories and rewards those that produce better images or more correct answers. The key technical move is accumulating policy gradients across all denoising steps of the full trajectory, rather than estimating quality from a single step as prior diffusion-RL work did. On text-to-image generation, this improves spatial relationship handling and multi-object composition. On multimodal understanding, it lets the model produce longer, more structured reasoning chains instead of collapsing to single-token answers.

What carries the argument

A control block built on the Unmask Policy Module (UPM) that takes hidden states from the final transformer layer plus a step index and mask indicator, and outputs ranking scores for each masked token. These scores define a Plackett-Luce policy that samples which K tokens to unmask at each denoising step without replacement. The policy is optimized via GRPO, where G trajectories are sampled per prompt, rewards are computed (PickScore for images, rule-based correctness for understanding), advantages are normalized within the group, and gradients are accumulated across all T denoising steps rather than estimated from a single transition.

What would settle it

If the learned control block's improvements vanish when the reward model is swapped for a different one of comparable quality, or if the gains disappear under a different base model architecture, the contribution would reduce to 'GRPO with a good reward improves outputs' rather than 'generation order is the lever.' The strongest falsification would be showing that a randomly initialized control block trained with the same GRPO pipeline yields comparable gains — meaning the order policy is irrelevant and the improvement comes entirely from the RL fine-tuning of the base model.

Watch

Extended reading notes

Core claim

The central finding is that the generation order problem in multimodal masked diffusion is not solvable by logit-based heuristics — confidence at a position does not predict its importance to the global output — but is solvable by a learned policy optimized via reinforcement learning over full generation trajectories. The control block, trained with GRPO using trajectory-level reward signals, produces orderings that improve both image composition (particularly spatial relations and multi-object scenes) and multimodal reasoning depth. This separates the 'what to generate next' question from the 'what value to assign' question, and shows that the former requires learned, task-specific policies

Load-bearing premise

The method assumes that the reward signal used during training — PickScore for image generation and rule-based correctness for understanding — is a sufficient proxy for the true quality of a generation order policy. PickScore is a learned human-preference model whose alignment with compositional correctness is imperfect: the paper's own results show that training with PickScore but evaluating on GenEval yields weaker gains than training directly with GenEval rewards (0.50 vs

Editorial extensions

If this is right

  • If learned order policies generalize beyond the tested tasks, any masked diffusion model — for audio, video, or 3D generation — could benefit from a similar control block, since the core issue (positional confidence ≠ global importance) is not specific to images.
  • The separation of order policy from token prediction suggests a modular training paradigm: freeze the base diffusion model and train only the lightweight order controller, reducing the cost of adapting a single backbone to many downstream tasks.
  • The trajectory-level gradient accumulation method could apply to other multi-step generative processes (autoregressive, flow matching) where per-step credit assignment is difficult.
  • The failure of logit-based heuristics in multimodal settings implies that confidence calibration research focused on language may not transfer to visual token spaces, redirecting where calibration work is needed.

Reading between the lines

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

  • The reward model's quality bounds the control block's learned policy: the paper shows PickScore-trained evaluation on GenEval yields 0.50 vs 0.51 for GenEval-trained, suggesting the ceiling is the reward signal's fidelity to the true objective. Better reward models would likely unlock larger gains than architectural improvements to the control block itself.
  • The observation that the baseline model collapses to single-token answers despite being allowed 256 tokens suggests the generation order policy implicitly controls response length and reasoning depth — making order optimization a form of learned chain-of-thought scaffolding.
  • If the control block is lightweight relative to the base model, the marginal inference cost of learned ordering may be small enough that it becomes a default add-on rather than a research luxury — but the paper does not report the control block's parameter count or latency overhead, leaving this open.
  • The 4–5% relative improvements are modest in absolute terms, but they appear concentrated in the hardest subtasks (spatial relations, multi-object composition, open-ended reasoning), which are exactly where current multimodal models struggle most — suggesting the method addresses a genuine bottleneck rather than uniformly smoothing performance.
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 / 7 minor

Summary. This paper investigates the optimization of token generation order in multimodal masked diffusion models (MDMs). The authors first establish that existing logit-based heuristics (Top-K, Top-K margin), which are effective for structured language tasks like Sudoku, do not transfer to multimodal settings (text-to-image generation and multimodal understanding). To address this, they propose a learnable control block — built on the Unmask Policy Module (UPM) from Huang et al. (2025) — that assigns ranking scores to masked tokens at each denoising step, with selection governed by a Plackett-Luce policy. The control block is trained via Group Relative Policy Optimization (GRPO), accumulating policy gradients across all denoising steps rather than using a single-step density ratio as in prior diffu-GRPO work. Experiments are conducted on the GenEval benchmark for text-to-image generation and VLMEvalKit for multimodal understanding, using MMaDA-8B-MixCoT as the base model. The authors report 4.08% and 4.85% relative improvements on these benchmarks, respectively.

Significance. The paper addresses a timely and well-motivated question: whether generation order matters in multimodal MDMs and, if so, how to optimize it. The negative result — that logit-based heuristics fail in multimodal settings — is a useful empirical finding. The proposed solution (a GRPO-trained control block with full-trajectory gradient accumulation) is technically reasonable and builds cleanly on prior work (MMaDA, diffu-GRPO, UPM) with appropriate citations. The experimental setup includes proper deduplication of GenEval prompts to prevent data leakage and evaluation on standard benchmarks. However, the significance of the results is tempered by the small magnitude of improvements (0.02 absolute on GenEval overall), the absence of variance estimates, and apparent data presentation issues in Table 2 that make the multimodal understanding results difficult to interpret.

major comments (4)
  1. Tables 1 and 2 report only single-run point estimates with no standard deviations, confidence intervals, or multiple-seed runs. The GenEval overall improvement is 0.49 to 0.51 (absolute gain of 0.02). Given that the generation process involves stochastic unmasking via Plackett-Luce sampling (Section 3.1) and GRPO training samples G=4 stochastic trajectories per prompt (Section 3.2), there are multiple sources of variance. Without variance estimates, it is impossible to determine whether these small gains reflect a real effect of the learned generation order policy or run-to-run noise. This is load-bearing for the central claim of 'substantial improvements.'
  2. Table 2 appears to contain data presentation errors that undermine the multimodal understanding results. For MMMU, the baseline MMaDA-COT reports 80.2 while the proposed method reports 10.3 — this looks like a decimal point or formatting error (likely 80.2 vs 10.3 should be 80.2 vs 10.3 is implausible; perhaps 80.2 vs 10.3 should be 80.2 vs 10.3 — regardless, a drop from 80.2 to 10.3 would be catastrophic, not an improvement). Similarly, MMB shows 70.2 for the baseline and 10.3 for the proposed method, and SEED shows 0.52 for the baseline and 0.24 for the proposed method. These values are inconsistent with the claimed 4.85% relative improvement and must be corrected before the results can be properly assessed.
  3. The '4.85% relative improvement in multimodal understanding' claimed in the abstract is not clearly supported by Table 2. Several sub-benchmarks show the proposed method performing worse than the baseline (GQA: 48.5 to 47.6; SEED: 0.52 to 0.24). The paper does not specify how the aggregate improvement is computed — whether it is an average across benchmarks, a weighted average, or something else. The authors should clarify the aggregation method and report per-benchmark improvements transparently, including which benchmarks show degradation.
  4. Section 3.2, Eq. (1): The paper states that gradients are accumulated across all T denoising steps, in contrast to the single-step density ratio estimation in Zhao et al. (2025) and Yang et al. (2025). However, the paper does not provide an ablation isolating the contribution of full-trajectory gradient accumulation versus single-step estimation. Since this is presented as a key methodological difference from prior work, an ablation or at least a discussion of why full-trajectory accumulation is necessary would strengthen the contribution. Additionally, the relationship between the per-step policy ratio in Eq. (1) and the sequence-level likelihood approximation used in diffu-GRPO (Section 2.2) should be clarified — is the product of per-step ratios equivalent to the mean-field sequence likelihood used in diffu-GRPO?
minor comments (7)
  1. Only one base model (MMaDA-8B-MixCoT) is tested. While this is understandable given computational constraints, the generality of the approach across different multimodal MDM architectures is unverified. A brief discussion of whether the control block is expected to transfer to other architectures would strengthen the generality claim.
  2. Table 1: The two variants of the proposed method (PickScore reward vs GenEval reward) show very similar overall scores (0.50 vs 0.51). The paper notes that training with PickScore and evaluating on GenEval demonstrates generalization, but the difference is within what could be noise given no variance estimates.
  3. Section 4.1: The prompt sampling ratio (Position:Counting:Attribute Binding:Colors:Two Objects:Single Object = 7:5:3:1:1:0) is stated without justification. How sensitive are the results to this ratio? Was it tuned on a validation set?
  4. The reward weight R = 2*R_correctness + 0.4*R_format (Section 4.2) is presented without justification for the specific coefficients. A brief ablation or sensitivity analysis on this weighting would be helpful.
  5. Figure references (Figures 2, 3, 4) are mentioned in the text but the figures themselves were not fully legible in the review materials. The authors should ensure trajectory visualizations and qualitative comparisons are clearly readable and properly labeled.
  6. The notation switches between x_T for clean data (Section 2.1) and x_0 for the initial masked state, which may confuse readers familiar with the standard convention where x_0 is clean data. The footnote explains this choice but it adds cognitive overhead.
  7. The KL penalty coefficient beta is set to 0 (Appendix A). This means the KL regularization term in the GRPO objective is effectively disabled. The authors should discuss whether this affects training stability or policy collapse, and whether experiments with non-zero beta were attempted.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the learned control block is trained via GRPO against external reward signals and evaluated on independent benchmarks; no step reduces to its inputs by construction.

full rationale

The paper's central claim is that a learnable control block, trained via GRPO, produces better generation orderings than logit-based heuristics (Top-K, Top-K margin) for multimodal masked diffusion models. Walking the derivation chain: (1) The control block (UPM-based Plackett-Luce policy, Section 3.1) is a genuinely new learnable module whose parameters are not defined in terms of the evaluation metrics. (2) The GRPO objective (Eq. 1, Section 3.2) optimizes this policy against external reward functions — PickScore for image generation and rule-based correctness/format rewards for multimodal understanding. These rewards are computed by external evaluators, not by the control block itself, so the optimization target is not self-referential. (3) The base model (MMaDA-8B), the diffu-GRPO framework (Zhao et al. 2025), and the UPM module (Huang et al. 2025) are cited as independent prior work; none of these citations are load-bearing for the core claim in a way that would make the result tautological. The paper does train with GenEval-derived rewards and then evaluates on GenEval (Table 1, 'MMaDA-Ours (GenEval)' row), which could raise a train-test contamination concern, but the paper also reports a 'MMaDA-Ours (PickScore)' row trained on a different reward and still evaluated on GenEval, showing the improvement is not purely an artifact of reward fitting. The evaluation benchmarks (GenEval, VLMEvalKit) are standard external suites. No equation or definition in the paper creates a self-definitional loop where the output is equivalent to the input by construction. The concerns about statistical significance and reward model fidelity are correctness/empirical-validity issues, not circularity issues. The derivation is self-contained against external benchmarks.

Assumptions & free parameters 6 free parameters · 4 assumptions · 1 invented entities

The paper introduces one new architectural entity (the control block configuration) and relies on several empirically chosen hyperparameters. The free parameters are standard RL/diffusion training hyperparameters with no principled derivation. The key axioms are inherited from prior work (mean-field approximation) or established empirically within the paper (logit insufficiency). No invented physical or mathematical entities are introduced.

free parameters (6)
  • Prompt sampling ratio (Position:Counting:Attribute:Colors:TwoObj:SingleObj) = 7:5:3:1:1:0
    Set based on base model's initial task accuracy; weights are hand-tuned, not derived from any principle.
  • Reward weight R = 2*R_correctness + 0.4*R_format = 2.0, 0.4
    Weights for combining correctness and format rewards in multimodal understanding training; chosen empirically without stated justification.
  • Learning rate = 5e-6
    Standard hyperparameter, not a derived quantity.
  • Clipping coefficient epsilon = 0.2
    Standard PPO/GRPO hyperparameter.
  • KL penalty coefficient beta = 0
    Set to zero, disabling KL regularization entirely; this is a modeling choice with no stated justification for why KL regularization is unnecessary.
  • Group size G and repeat rounds R = G=4, R=6 (T2I) or R=3 (VQA)
    Sampling configuration chosen empirically.
assumptions (4)
  • domain assumption Model logits alone are insufficient for determining optimal generation sequences in multimodal tasks (Section 1, preliminary experiments).
    This is established empirically by showing Top-K margin fails on multimodal tasks, but the preliminary experiments showing this failure are not detailed in the paper — only stated as motivation.
  • domain assumption The mean-field approximation of sequence likelihood under masked diffusion policies is valid (Section 2.2, citing Zhao et al. 2025).
    The diffu-GRPO objective relies on this approximation; the paper inherits it from prior work without independent verification.
  • domain assumption PickScore is a sufficient reward proxy for human-preference-aligned image generation quality.
    Used as reward model for text-to-image training; the paper itself shows weaker results with PickScore vs task-specific rewards, suggesting this axiom is only partially valid.
  • domain assumption Accumulating policy gradients across all denoising steps provides better training signal than single-step estimation.
    Stated in Section 3.2 as motivation for the full-trajectory approach; no ablation comparing single-step vs full-trajectory is provided.
invented entities (1)
  • Control block (UPM-based ranking module for multimodal MDMs) independent evidence
    purpose: Assigns ranking scores to masked tokens to determine generation order at each denoising step.
    The control block is evaluated on external benchmarks (GenEval, VLMEvalKit) with metrics not used during training (for the PickScore variant), providing falsifiable evidence of its effectiveness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reinforcing the Generation Order of Multimodal Masked Diffusion Models." pith.science (2026). https://pith.science/paper/GTAO4ZZ2

@misc{pith2026260708056,
  author       = {Pith},
  title        = {Pith review of: Reinforcing the Generation Order of Multimodal Masked Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GTAO4ZZ2}},
  note         = {Machine review of arXiv:2607.08056}
}
read the original abstract

Diffusion Language Models (DLMs) have recently achieved substantial progress in natural language generation tasks. Recent research demonstrates that adaptive token generation ordering can significantly improve performance in mathematical reasoning and code synthesis applications. In this work, we investigate the optimization of generation order for both text-to-image synthesis and multimodal understanding. We first establish that, unlike structured problems in language generation such as Sudoku puzzles, model logits alone are insufficient for determining optimal generation sequences in text-to-image generation and multimodal understanding. To address this challenge, we introduce a learnable control module trained via Group Relative Policy Optimization (GRPO) to determine the generation order. Our results demonstrate that learning this control block substantially improves both text-to-image alignment and multimodal understanding in DLMs. In particular, it enhances the model's ability to capture fine-grained spatial relationships in generated images while also strengthening performance on multimodal reasoning and comprehension tasks. We evaluate our framework on GenEval, an object-focused benchmark for text-to-image alignment, where it achieves 4.08% relative improvements. In addition, experiments on VLMEvalKit confirm 4.85% relative improvements in multimodal understanding, highlighting the broad effectiveness of our approach.

Figures

Figures reproduced from arXiv: 2607.08056 by the authors.

Figure 1
Figure 1. Generation pipeline of the diffusion language model with a control block for determining the [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Trajectory visualization on text-to-image generation. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Trajectory visualization on multimodal understanding. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of generated samples. Our method demonstrates superior understand [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 59 canonical work pages

  1. [1]

    ArXiv , year=

    Deep Unsupervised Learning using Nonequilibrium Thermodynamics , author=. ArXiv , year=

    work page

  2. [2]

    ArXiv , year=

    Structured Denoising Diffusion Models in Discrete State-Spaces , author=. ArXiv , year=

    work page

  3. [3]

    Neural Information Processing Systems , year=

    Argmax Flows and Multinomial Diffusion: Learning Categorical Distributions , author=. Neural Information Processing Systems , year=

    work page

  4. [4]

    ArXiv , year=

    A Continuous Time Framework for Discrete Denoising Models , author=. ArXiv , year=

    work page

  5. [5]

    ArXiv , year=

    Score-based Continuous-time Discrete Diffusion Models , author=. ArXiv , year=

    work page

  6. [6]

    ArXiv , year=

    Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models , author=. ArXiv , year=

    work page

  7. [7]

    International Conference on Machine Learning , year=

    Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution , author=. International Conference on Machine Learning , year=

    work page

  8. [8]

    ArXiv , year=

    Simple and Effective Masked Diffusion Language Models , author=. ArXiv , year=

    work page

Show all 59 references
  1. [9]

    ArXiv , year=

    Masked Diffusion Models are Secretly Time-Agnostic Masked Models and Exploit Inaccurate Categorical Sampling , author=. ArXiv , year=

  2. [10]

    ArXiv , year=

    Your Absorbing Discrete Diffusion Secretly Models the Conditional Distributions of Clean Data , author=. ArXiv , year=

  3. [11]

    ArXiv , year=

    Generalized Interpolating Discrete Diffusion , author=. ArXiv , year=

  4. [12]

    ArXiv , year=

    Simplified and Generalized Masked Diffusion for Discrete Data , author=. ArXiv , year=

  5. [13]

    ArXiv , year=

    Large Language Diffusion Models , author=. ArXiv , year=

  6. [14]

    ArXiv , year=

    Discrete Flow Matching , author=. ArXiv , year=

  7. [15]

    ArXiv , year=

    Generative Flows on Discrete State-Spaces: Enabling Multimodal Flows with Applications to Protein Co-Design , author=. ArXiv , year=

  8. [16]

    ArXiv , year=

    d1: Scaling Reasoning in Diffusion Large Language Models via Reinforcement Learning , author=. ArXiv , year=

  9. [17]

    ArXiv , year=

    MMaDA: Multimodal Large Diffusion Language Models , author=. ArXiv , year=

  10. [18]

    ArXiv , year=

    FUDOKI: Discrete Flow-based Unified Understanding and Generation via Kinetic-Optimal Velocities , author=. ArXiv , year=

  11. [19]

    ArXiv , year=

    Reinforcing the Diffusion Chain of Lateral Thought with Diffusion Language Models , author=. ArXiv , year=

  12. [20]

    ArXiv , year=

    Train for the Worst, Plan for the Best: Understanding Token Ordering in Masked Diffusions , author=. ArXiv , year=

  13. [21]

    ArXiv , year=

    A Reparameterized Discrete Diffusion Model for Text Generation , author=. ArXiv , year=

  14. [22]

    ArXiv , year=

    Training language models to follow instructions with human feedback , author=. ArXiv , year=

  15. [23]

    ArXiv , year=

    Proximal Policy Optimization Algorithms , author=. ArXiv , year=

  16. [24]

    ArXiv , year=

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , author=. ArXiv , year=

  17. [25]

    ArXiv , year=

    GenEval: An Object-Focused Framework for Evaluating Text-to-Image Alignment , author=. ArXiv , year=

  18. [26]

    ArXiv , year=

    Flow-GRPO: Training Flow Matching Models via Online RL , author=. ArXiv , year=

  19. [27]

    ArXiv , year=

    Pick-a-Pic: An Open Dataset of User Preferences for Text-to-Image Generation , author=. ArXiv , year=

  20. [28]

    ArXiv , year=

    MixGRPO: Unlocking Flow-based GRPO Efficiency with Mixed ODE-SDE , author=. ArXiv , year=

  21. [29]

    ArXiv , year=

    Autoregressive Model Beats Diffusion: Llama for Scalable Image Generation , author=. ArXiv , year=

  22. [30]

    2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    High-Resolution Image Synthesis with Latent Diffusion Models , author=. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    2022

  23. [31]

    ArXiv , year=

    Hierarchical Text-Conditional Image Generation with CLIP Latents , author=. ArXiv , year=

  24. [32]

    ArXiv , year=

    SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis , author=. ArXiv , year=

  25. [33]

    ArXiv , year=

    DreamLLM: Synergistic Multimodal Comprehension and Creation , author=. ArXiv , year=

  26. [34]

    ArXiv , year=

    SEED-X: Multimodal Models with Unified Multi-granularity Comprehension and Generation , author=. ArXiv , year=

  27. [35]

    ArXiv , year=

    LLaVA-Plus: Learning to Use Tools for Creating Multimodal Agents , author=. ArXiv , year=

  28. [36]

    ArXiv , year=

    World Model on Million-Length Video And Language With Blockwise RingAttention , author=. ArXiv , year=

  29. [37]

    ArXiv , year=

    Generative Pretraining in Multimodality , author=. ArXiv , year=

  30. [38]

    ArXiv , year=

    Show-o: One Single Transformer to Unify Multimodal Understanding and Generation , author=. ArXiv , year=

  31. [39]

    ArXiv , year=

    Janus: Decoupling Visual Encoding for Unified Multimodal Understanding and Generation , author=. ArXiv , year=

  32. [40]

    ArXiv , year=

    Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context , author=. ArXiv , year=

  33. [41]

    ArXiv , year=

    VARGPT: Unified Understanding and Generation in a Visual Autoregressive Multimodal Large Language Model , author=. ArXiv , year=

  34. [42]

    2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering , author=. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    2019

  35. [43]

    Proceedings of CVPR , year=

    MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI , author=. Proceedings of CVPR , year=

  36. [44]

    European conference on computer vision , pages=

    Mmbench: Is your multi-modal model an all-around player? , author=. European conference on computer vision , pages=. 2024 , organization=

    2024

  37. [45]

    International Conference on Learning Representations (ICLR) , year=

    MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts , author=. International Conference on Learning Representations (ICLR) , year=

  38. [46]

    European Conference on Computer Vision , year=

    Microsoft COCO: Common Objects in Context , author=. European Conference on Computer Vision , year=

  39. [47]

    ArXiv , year=

    Choosing to Rank , author=. ArXiv , year=

  40. [48]

    Proceedings of the 21st ACM international conference on Information and knowledge management , year=

    A new probabilistic model for top-k ranking problem , author=. Proceedings of the 21st ACM international conference on Information and knowledge management , year=

  41. [49]

    ArXiv , year=

    DanceGRPO: Unleashing GRPO on Visual Generation , author=. ArXiv , year=

  42. [50]

    ArXiv , year=

    TempFlow-GRPO: When Timing Matters for GRPO in Flow Models , author=. ArXiv , year=

  43. [51]

    ArXiv , year=

    SFT or RL? An Early Investigation into Training R1-Like Reasoning Large Vision-Language Models , author=. ArXiv , year=

  44. [52]

    ArXiv , year=

    Lumina-DiMOO: An Omni Diffusion Large Language Model for Multi-Modal Generation and Understanding , author=. ArXiv , year=

  45. [53]

    ArXiv , year=

    Understanding R1-Zero-Like Training: A Critical Perspective , author=. ArXiv , year=

  46. [54]

    ArXiv , year=

    Muddit: Liberating Generation Beyond Text-to-Image with a Unified Discrete Diffusion Model , author=. ArXiv , year=

  47. [55]

    ArXiv , year=

    LaViDa: A Large Diffusion Language Model for Multimodal Understanding , author=. ArXiv , year=

  48. [56]

    ArXiv , year=

    Meissonic: Revitalizing Masked Generative Transformers for Efficient High-Resolution Text-to-Image Synthesis , author=. ArXiv , year=

  49. [57]

    ArXiv , year=

    d2: Improved Techniques for Training Reasoning Diffusion Language Models , author=. ArXiv , year=

  50. [58]

    ArXiv , year=

    DiRL: An Efficient Post-Training Framework for Diffusion Language Models , author=. ArXiv , year=

  51. [59]

    ArXiv , year=

    Revolutionizing Reinforcement Learning Framework for Diffusion Large Language Models , author=. ArXiv , year=

Pith tools

Reviewed July 10, 2026 · model on record in the stance chip above.