Pith. sign in

REVIEW 5 major objections 4 minor 16 cited by

A masked-diffusion vision-language model can be turned into a robot policy that outperforms auto-regressive baselines on simulation and real hardware.

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

2026-08-04 22:51 UTC pith:KA2VX6DZ

load-bearing objection A genuinely new direction—first VLA on a pretrained masked-diffusion VLM—but the empirical SOTA claim is only as good as the uneven baseline comparisons, which need serious shoring up before I'd trust the headline margins. the 5 major comments →

arxiv 2509.06932 v2 pith:KA2VX6DZ submitted 2025-09-08 cs.RO cs.CV

LLaDA-VLA: Vision Language Diffusion Action Models

Yuqing Wen , Hebei Li , Kefan Gu , Yucheng Zhao , Tiancai Wang , Xiaoyan Sun This is my paper
classification cs.RO cs.CV
keywords vision-language-action modelmasked diffusionrobotic manipulationdiffusion vision-language modelspecial-token classificationhierarchical decodingCALVIN benchmarkSimplerEnv
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.

LLaDA-VLA sets out to show that masked-diffusion vision-language models, not just auto-regressive ones, can serve as the backbone for robotic action generation. To bridge the gap between general-purpose d-VLMs and precise low-level control, the paper restricts prediction to a small set of special action tokens and decodes action chunks hierarchically by confidence. If the comparisons hold, these two changes turn a pretrained d-VLM into a policy that exceeds existing VLA baselines on long-horizon simulation tasks and on a real robot arm. The work matters because it opens a third paradigm—masked diffusion—for building vision-language-action models, alongside auto-regressive and flow-matching approaches.

Core claim

The paper's central claim is that a pretrained masked-diffusion VLM, fine-tuned with a localized special-token classification objective and a hierarchical action-structured decoding rule, can generate robot action sequences more accurately than existing auto-regressive and flow-matching VLAs. Instead of classifying over the full vocabulary, the model learns to choose among 32 special action tokens representing discretized position, rotation, and gripper commands. During decoding, actions are ranked by summed token confidence; the highest-confidence action is partially revealed token-by-token while lower-confidence actions are remasked for later diffusion steps. On CALVIN ABC-D, the model com

What carries the argument

The core mechanism is the masked-diffusion predict-remask cycle, in which a transformer mask predictor fills masked tokens and low-confidence predictions are remasked for iterative refinement. LLaDA-VLA adds two modifications: (1) localized special-token classification, which restricts the cross-entropy loss and inference to the V_a special action tokens rather than the full vocabulary, and (2) hierarchical action-structured decoding, which first ranks actions by summed token confidence and then ranks tokens within the top action, preserving both intra-action and inter-action structure. The action chunk is K×7 tokens, where 7 represents three positional deltas, three rotational deltas, and o

Load-bearing premise

The central claim rests on the assumption that the baselines were evaluated under matched conditions—same training data, task definitions, and robot setups—so the quoted gains reflect true superiority rather than comparison artifacts.

What would settle it

Run OpenVLA, CogACT, and pi0 on exactly the same training data, action tokenization, and evaluation protocol used for LLaDA-VLA on CALVIN ABC-D and SimplerEnv WidowX, across multiple seeds; if the average CALVIN length and SimplerEnv success rates do not reproduce the reported margins, the state-of-the-art claim is unsupported.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Masked-diffusion VLMs become a viable third backbone family for robot policies, alongside auto-regressive VLMs and flow-matching models.
  • Parallel, confidence-based decoding can replace left-to-right token generation for action sequences, potentially improving long-horizon coherence.
  • Restricting classification to task-relevant special tokens makes it easier to adapt large pretrained diffusion models to structured output domains.
  • Action chunk size has a sweet spot—here chunk size 5 beats both smaller and larger chunks—because larger chunks make mask prediction harder while smaller chunks reduce trajectory smoothness.

Where Pith is reading between the lines

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

  • An extension the paper leaves implicit: the localized special-token trick could transfer to other VLM-controlled structured outputs, such as API calls, GUI commands, or code generation, where full-vocabulary classification is wasteful.
  • Hierarchical confidence-based remasking is a general way to impose group structure on discrete diffusion; it could be tested on other group-structured sequence problems, such as protein or scene generation.
  • Because the paper quotes baseline numbers from prior publications rather than rerunning them, a direct matched re-implementation on identical data and robot setups would be the decisive test of the claimed margins.
  • A further untested combination would be to apply the hierarchical decoding rule to continuous action heads or flow-matching policies, to isolate whether the gains come from the diffusion backbone or the structured decoding strategy.

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

5 major / 4 minor

Summary. The paper introduces LLaDA-VLA, a Vision-Language-Diffusion-Action model built on a pretrained masked diffusion vision-language model (LLaDA-V). To adapt the diffusion VLM to low-level robotic control, the authors propose (1) localized special-token classification, which restricts prediction to a small set of action tokens, and (2) hierarchical action-structured decoding, which remasks at both the action and token levels according to confidence scores. The model is evaluated on CALVIN, SimplerEnv, and a real WidowX robot, where it is reported to outperform OpenVLA, CogACT, and π0. Ablations on CALVIN attribute performance gains to each of the two proposed components, and an additional ablation studies the action chunk size.

Significance. If the reported results hold, the paper opens a new direction: applying large masked diffusion VLMs to robotic action generation. The two design choices are simple and well-motivated, and the ablation study isolates their contributions. The work is timely given the concurrent DiscreteDiffusionVLA and the broader interest in non-autoregressive VLAs. However, the central claim is empirical state-of-the-art performance, and that claim currently rests on comparison protocols that are not fully matched or described. The paper would be strengthened by releasing code, fine-tuning data specifications, and matched baseline evaluations.

major comments (5)
  1. [§4.1.1, Table 1] The SimplerEnv training/evaluation protocol is unspecified. The text only says 'We train and evaluate LLaDA-VLA on SimplerEnv'; it does not state the demonstration count, task setup, or whether comparison methods (RT-1-X, Octo, OpenVLA, CogACT) were fine-tuned on the same data or evaluated zero-shot. SimplerEnv is commonly used to test pretrained policies, so if baselines are quoted from prior checkpoints without matched fine-tuning, the reported +51.3 percentage-point gain over OpenVLA is not a controlled comparison. Please specify the training data and either rerun baselines under identical conditions or clearly label zero-shot versus fine-tuned.
  2. [Table 1 and §4.2.1] Table 1 is internally inconsistent. The DiscreteDiffusionVLA row reports three per-task values (37.5, 20.8, 29.2) with average 29.2, which is the mean of those three, while every other row averages all four tasks; treating the missing entry as zero would give 21.9. The text gives LLaDA-VLA average as 55.1, while the table row says 55.5; the introduction claims a 51.3% gain over OpenVLA, whereas §4.2.1 states 50.9%. These discrepancies make the headline margins ambiguous and need correction.
  3. [§4.2.1, Tables 3–4] The real-robot comparisons are not controlled. π0 and CogACT numbers appear to be taken from external papers rather than run on the authors' WidowX setup; the manuscript does not state baseline trial counts, task definitions, or observation/action settings. With only 10 trials per LLaDA-VLA task and no confidence intervals, a difference of 20–30 percentage points on small samples is not established. Please either re-evaluate baselines in the same harness or report the external source conditions and uncertainty estimates.
  4. [§4.2.3, Table 5] The ablation baseline text says 'average episode length of only 2.54' but Table 5 reports 2.64. More substantively, the ablations are single-run numbers with no multiple seeds or confidence intervals; the +0.79 and +0.58 Avg. Len. improvements may be within run-to-run variability on CALVIN. Since these two components are the paper's core technical contribution, the ablation evidence should include variance estimates or repeated-seed results.
  5. [§4.1.2, Table 6] The design choices K=5 and 10 diffusion steps are selected on CALVIN and then transferred without discussion to SimplerEnv and real-robot experiments. The paper does not report whether the performance gain is robust to chunk-size or step-count variations across environments. This limits the generality of the main results, though it is secondary to the comparison-fairness issue.
minor comments (4)
  1. [§3.2.3] The 'confidence' used for action-level and token-level remasking is never formally defined. Please state explicitly whether it is the softmax probability of the predicted token, a logit, or some normalized quantity.
  2. [§4.2.3 and §4.2.1] There are numerical inconsistencies in the text: '2.54' vs '2.64' for the baseline Avg. Len., and '50.9%' vs '51.3%' for the OpenVLA gain. These should be reconciled with the tables.
  3. [Abstract and Related Work] The abstract and introduction claim 'the first Vision-Language-Diffusion-Action model' despite the concurrent DiscreteDiffusionVLA [30]. Please clarify the exact novelty claim and how LLaDA-VLA differs from [30] beyond using a pretrained d-VLM.
  4. [Table 2] The table caption says 'average success rate over 1000 rollouts per task' but does not specify the number of seeds or evaluation episodes used for the baselines. Please provide the protocol and, if possible, error bars.

Circularity Check

0 steps flagged

No circular derivation: the method's components are ablated against external benchmarks, pretrained weights are external, and no fitted quantity is presented as a prediction.

full rationale

The paper's central claims are empirical: LLaDA-VLA is built on pretrained d-VLMs (LLaDA-V, which is external to the authors) and evaluated on external benchmarks (CALVIN, SimplerEnv) and a real robot. The two proposed components, localized special-token classification and hierarchical action-structured decoding, are not derived from the target results; they are introduced as modeling choices and validated through ablations on CALVIN. No parameter is fitted to the reported benchmark numbers and then renamed as a prediction. Hyperparameters such as action chunk size (K=5) and diffusion steps (10) are selected via ablations or stated as fixed design choices, not optimized to force the SOTA outcome. The pretrained weights and benchmark definitions come from outside this paper, so there is no self-citation chain that makes the conclusion load-bearing. The main weaknesses are concerns about comparison fairness (e.g., SimplerEnv baseline numbers are quoted without a fully specified matched training protocol, and real-robot comparisons use only 10 trials per task) and internal numerical inconsistencies (e.g., Table 1's average for DiscreteDiffusionVLA is computed over three tasks, and the text's 55.1 average differs from the table's 55.5). These are validity/reliability issues, not circularity: they do not show that any claimed result reduces by definition to its own inputs. Therefore, no circular step is identified.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

Free parameters are the hyperparameters selected through ablations or fixed without reported sensitivity analysis. Axioms list the background assumptions inherited from the masked diffusion framework, the action discretization choice, the benchmark protocol, and the fine-tuning regime. No new physical or theoretical entities are introduced; the added action tokens are a technical vocabulary extension, not a postulated entity.

free parameters (6)
  • action chunk size K = 5
    Selected via ablation on CALVIN (Table 6); K=5 gives the best Avg. Len. of 4.01.
  • number of diffusion steps = 10
    Set for inference; '2 iterations per action' implies chunk size 5; not ablated.
  • action token vocabulary size Va = 32
    32 additional special tokens introduced to discretize continuous actions; chosen without reported ablation.
  • learning rate = 2e-5
    Fine-tuning hyperparameter.
  • batch size = 128
    Fine-tuning hyperparameter.
  • training epochs = 3
    Fine-tuning schedule.
axioms (4)
  • domain assumption Masked diffusion with confidence-based remasking can generate coherent discrete token sequences (from LLaDA).
    The decoding strategy in Section 3.1 relies on the LLaDA reverse process where low-confidence tokens are remasked; this is assumed to transfer to action tokens.
  • domain assumption Discretizing continuous robot actions into 32 shared bins preserves enough precision for successful manipulation.
    Section 3.2.1 discretizes actions into bins with Va=32 special tokens; no analysis of discretization error is provided.
  • domain assumption The evaluation benchmarks (CALVIN ABC-D, SimplerEnv Visual Matching, WidowX tasks) are valid and the baselines are comparable.
    Section 4.1.1 describes the benchmarks but does not specify the training data for SimplerEnv or confirm baseline parity.
  • domain assumption Fine-tuning the pretrained d-VLM LLaDA-V for 3 epochs adapts it to action prediction without catastrophic forgetting.
    Section 4.1.2 sets the fine-tuning schedule; no analysis of forgetting is given.

pith-pipeline@v1.3.0-alltime-deepseek · 12769 in / 16466 out tokens · 151504 ms · 2026-08-04T22:51:02.739119+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of LLaDA-VLA: Vision Language Diffusion Action Models." pith.science (2026). https://pith.science/paper/KA2VX6DZ

@misc{pith2026250906932,
  author       = {Pith},
  title        = {Pith review of: LLaDA-VLA: Vision Language Diffusion Action Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KA2VX6DZ}},
  note         = {Machine review of arXiv:2509.06932}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The rapid progress of auto-regressive vision-language models (VLMs) has inspired growing interest in vision-language-action models (VLA) for robotic manipulation. Recently, masked diffusion models, a paradigm distinct from autoregressive models, have begun to demonstrate competitive performance in text generation and multimodal applications, leading to the development of a series of diffusion-based VLMs (d-VLMs). However, leveraging such models for robot policy learning remains largely unexplored. In this work, we present LLaDA-VLA, the first Vision-Language-Diffusion-Action model built upon pretrained d-VLMs for robotic manipulation. To effectively adapt d-VLMs to robotic domain, we introduce two key designs: (1) a localized special-token classification strategy that replaces full-vocabulary classification with special action token classification, reducing adaptation difficulty; (2) a hierarchical action-structured decoding strategy that decodes action sequences hierarchically considering the dependencies within and across actions. Extensive experiments demonstrate that LLaDA-VLA significantly outperforms state-of-the-art VLAs on both simulation and real-world robots.

Figures

Figures reproduced from arXiv: 2509.06932 by Hebei Li, Kefan Gu, Tiancai Wang, Xiaoyan Sun, Yucheng Zhao, Yuqing Wen.

Figure 1
Figure 1. Figure 1: Comparison between Autoregressive-based VLA Model [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of LLaDA-VLA. (a) Overall architecture. Visual features extracted by the vision encoder are projected into the text space and concatenated with text tokens. Together with masked tokens, they are fed into a large language diffusion model to generate action sequences via Localized Special-Token Classification and further refined with Hierarchical Action-Structured Decoding. (b) Hierarchical Action-S… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative results of LLaDA-VLA on CALVIN tasks. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative results of LLaDA-VLA on SimplerEnv tasks. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative results of LLaDA-VLA on real-world in-domain tasks. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative results of LLaDA-VLA on real-world out-of-domain tasks. [PITH_FULL_IMAGE:figures/full_fig_p008_6.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 16 Pith papers

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

  1. Rethinking Muon Beyond Pretraining: Spectral Failures and High-Pass Remedies for VLA and RLVR

    cs.LG 2026-05 conditional novelty 7.0

    Pion modifies Muon's Newton-Schulz iterations into a controllable high-pass filter that anchors dominant singular values at 1 while suppressing noisy tails, outperforming Muon and AdamW in VLA and RLVR regimes.

  2. BlockVLA: Accelerating Autoregressive VLA via Block Diffusion Finetuning

    cs.RO 2026-05 unverdicted novelty 7.0

    BlockVLA accelerates autoregressive VLA models by 3.3x using block diffusion finetuning, with faster training convergence and better early performance on long-horizon robotic tasks.

  3. ReflectDrive-2: Reinforcement-Learning-Aligned Self-Editing for Discrete Diffusion Driving

    cs.RO 2026-05 unverdicted novelty 7.0

    ReflectDrive-2 achieves 91.0 PDMS on NAVSIM with camera input by training a discrete diffusion model to self-edit trajectories via RL-aligned AutoEdit.

  4. DiscreteRTC: Discrete Diffusion Policies are Natural Asynchronous Executors

    cs.RO 2026-04 unverdicted novelty 7.0

    Discrete diffusion policies act as natural asynchronous executors for robotics by treating action generation as iterative unmasking, yielding higher success rates and lower computation than flow-matching real-time chu...

  5. DiscreteRTC: Discrete Diffusion Policies are Natural Asynchronous Executors

    cs.RO 2026-04 unverdicted novelty 7.0

    Discrete diffusion policies support native asynchronous execution via unmasking for real-time chunking, delivering higher success rates and 0.7x inference cost versus flow-matching RTC on dynamic robotics benchmarks a...

  6. DMax: Aggressive Parallel Decoding for dLLMs

    cs.LG 2026-04 conditional novelty 7.0

    DMax uses On-Policy Uniform Training and Soft Parallel Decoding to enable aggressive parallelism in dLLMs, raising TPF on GSM8K from 2.04 to 5.47 and on MBPP from 2.71 to 5.86 while preserving accuracy.

  7. SEAM: Smooth Execution of Action-Chunked Motion for Vision-Language-Action Policies

    cs.RO 2026-07 conditional novelty 6.0

    SEAM steers flow-matching VLA chunk denoising with a closed-form overlap correction from the previous unexecuted tail, cutting boundary jerk ~28% and discontinuity ~27% at near-baseline cost.

  8. Escaping the Diversity Trap in Robotic Manipulation via Anchor-Centric Adaptation

    cs.RO 2026-05 unverdicted novelty 6.0

    Anchor-Centric Adaptation escapes the diversity trap by prioritizing repeated demonstrations at core anchors over broad coverage, yielding higher success rates under fixed data budgets in robotic manipulation.

  9. ReflectDrive-2: Reinforcement-Learning-Aligned Self-Editing for Discrete Diffusion Driving

    cs.RO 2026-05 unverdicted novelty 6.0

    ReflectDrive-2 combines masked discrete diffusion with RL-aligned self-editing to generate and refine driving trajectories, reaching 91.0 PDMS on NAVSIM camera-only and 94.8 in best-of-6.

  10. dWorldEval: Scalable Robotic Policy Evaluation via Discrete Diffusion World Model

    cs.RO 2026-04 unverdicted novelty 6.0

    A discrete diffusion model tokenizes multimodal robotic data and uses a progress token to predict future states and task completion for scalable policy evaluation.

  11. Learning Native Continuation for Action Chunking Flow Policies

    cs.RO 2026-02 unverdicted novelty 6.0

    Legato trains flow-based VLA policies with schedule-shaped action-noise mixtures and randomized conditions to achieve smoother trajectories and ~10% faster task completion than real-time chunking across five real-worl...

  12. AsyncVLA: Asynchronous Flow Matching for Vision-Language-Action Models

    cs.RO 2025-11 unverdicted novelty 6.0

    AsyncVLA adds asynchronous flow matching and a confidence rater to VLA models so they can generate actions on flexible schedules and selectively refine low-confidence tokens before execution.

  13. TS-Mask VLA: 2D Temporal-Spatial Masking for Vision-Language-Action Model with Effective Bridging

    cs.RO 2026-07 conditional novelty 5.0

    A 0.5B VLA with bridge-conditioned discrete diffusion and 2D temporal–spatial action masking reaches 95.7% LIBERO success and 4.19 CALVIN average length.

  14. TBD-VLA: Temporal Block Diffusion Vision Language Action Model

    cs.CV 2026-06 unverdicted novelty 5.0

    TBD-VLA partitions action sequences into temporal blocks, performs masked discrete diffusion within blocks, and autoregressive generation across blocks to unify parallel decoding with temporal coherence in discrete VL...

  15. DMax: Aggressive Parallel Decoding for dLLMs

    cs.LG 2026-04 unverdicted novelty 5.0

    DMax enables faster parallel decoding in diffusion language models by using on-policy training to recover from errors and soft embedding interpolations for iterative revision, boosting tokens per forward pass roughly ...

  16. Discrete Diffusion Models: A Unified Framework from Tokenization to Generation

    cs.LG 2026-07 unverdicted novelty 4.0

    Discrete diffusion models are re-framed as instances of a tokenization-centric, four-component design space (corruption, denoiser, objective, sampler) in a broad survey with no new experimental or theoretical results.

Reference graph

Works this paper leans on

59 extracted references · 13 canonical work pages · cited by 13 Pith papers

  1. [1]

    Flamingo: a visual language model for few-shot learning.Advances in neural information processing systems, 35:23716–23736,

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Men- sch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning.Advances in neural information processing systems, 35:23716–23736,

  2. [2]

    Openflamingo: An open- source framework for training large autoregressive vision- language models.arXiv preprint arXiv:2308.01390, 2023

    Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An open- source framework for training large autoregressive vision- language models.arXiv preprint arXiv:2308.01390, 2023. 1

    Pith/arXiv arXiv 2023

  3. [3]

    Rt-h: Action hierarchies using language.arXiv preprint arXiv:2403.01823, 2024

    Suneel Belkhale, Tianli Ding, Ted Xiao, Pierre Sermanet, Quon Vuong, Jonathan Tompson, Yevgen Chebotar, De- bidatta Dwibedi, and Dorsa Sadigh. Rt-h: Action hierarchies using language.arXiv preprint arXiv:2403.01823, 2024. 1, 2

    Pith/arXiv arXiv 2024

  4. [4]

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al.pi 0: A vision-language- action flow model for general robot control.arXiv preprint arXiv:2410.24164, 2024. 1, 2, 6

    Pith/arXiv arXiv 2024

  5. [5]

    Zero-shot robotic manipulation with pretrained image- editing diffusion models.arXiv preprint arXiv:2310.10639,

    Kevin Black, Mitsuhiko Nakamoto, Pranav Atreya, Homer Walke, Chelsea Finn, Aviral Kumar, and Sergey Levine. Zero-shot robotic manipulation with pretrained image- editing diffusion models.arXiv preprint arXiv:2310.10639,

    Pith/arXiv arXiv

  6. [6]

    Rt-2: Vision-language-action models transfer web knowledge to robotic control.arXiv preprint arXiv:2307.15818, 2023

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control.arXiv preprint arXiv:2307.15818, 2023. 1, 2

    Pith/arXiv arXiv 2023

  7. [7]

    Rt-1: Robotics transformer for real-world control at scale

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakr- ishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 2

    Pith/arXiv arXiv 2022

  8. [8]

    Analog bits: Generating discrete data using diffusion models with self-conditioning.arXiv preprint arXiv:2208.04202, 2022

    Ting Chen, Ruixiang Zhang, and Geoffrey Hinton. Analog bits: Generating discrete data using diffusion models with self-conditioning.arXiv preprint arXiv:2208.04202, 2022. 2

    Pith/arXiv arXiv 2022

  9. [9]

    Llada-medv: Exploring large language diffusion models for biomedical image understanding.arXiv preprint arXiv:2508.01617,

    Xuanzhao Dong, Wenhui Zhu, Xiwen Chen, Zhipeng Wang, Peijie Qiu, Shao Tang, Xin Li, and Yalin Wang. Llada-medv: Exploring large language diffusion models for biomedical image understanding.arXiv preprint arXiv:2508.01617,

    arXiv

  10. [10]

    Palm- e: An embodied multimodal language model.arXiv preprint arXiv:2303.03378, 2023

    Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palm- e: An embodied multimodal language model.arXiv preprint arXiv:2303.03378, 2023. 2

    Pith/arXiv arXiv 2023

  11. [11]

    Fast ecot: Ef- ficient embodied chain-of-thought via thoughts reuse.arXiv preprint arXiv:2506.07639, 2025

    Zhekai Duan, Yuan Zhang, Shikai Geng, Gaowen Liu, Joschka Boedecker, and Chris Xiaoxuan Lu. Fast ecot: Ef- ficient embodied chain-of-thought via thoughts reuse.arXiv preprint arXiv:2506.07639, 2025. 2

    arXiv 2025

  12. [12]

    Scaling recti- fied flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling recti- fied flow transformers for high-resolution image synthesis. InForty-first international conference on machine learning,

  13. [13]

    Scaling diffusion language models via adaptation from autoregressive models.arXiv preprint arXiv:2410.17891, 2024

    Shansan Gong, Shivam Agarwal, Yizhe Zhang, Jiacheng Ye, Lin Zheng, Mukai Li, Chenxin An, Peilin Zhao, Wei Bi, Jiawei Han, et al. Scaling diffusion language models via adaptation from autoregressive models.arXiv preprint arXiv:2410.17891, 2024. 2, 3

    Pith/arXiv arXiv 2024

  14. [14]

    Diffuseq: Sequence to sequence text generation with diffusion models.arXiv preprint arXiv:2210.08933, 2022

    Shansan Gong, Mukai Li, Jiangtao Feng, Zhiyong Wu, and LingPeng Kong. Diffuseq: Sequence to sequence text generation with diffusion models.arXiv preprint arXiv:2210.08933, 2022. 2

    Pith/arXiv arXiv 2022

  15. [15]

    Rvt-2: Learning precise manipulation from few demonstrations.arXiv preprint arXiv:2406.08545, 2024

    Ankit Goyal, Valts Blukis, Jie Xu, Yijie Guo, Yu-Wei Chao, and Dieter Fox. Rvt-2: Learning precise manipulation from few demonstrations.arXiv preprint arXiv:2406.08545, 2024. 2

    Pith/arXiv arXiv 2024

  16. [16]

    Rvt: Robotic view transformer for 3d object manipulation

    Ankit Goyal, Jie Xu, Yijie Guo, Valts Blukis, Yu-Wei Chao, and Dieter Fox. Rvt: Robotic view transformer for 3d object manipulation. InConference on Robot Learning, pages 694–

    work page

  17. [17]

    Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020. 2

    2020

  18. [18]

    Coarse-to-fine q-attention: Efficient learn- ing for visual robotic manipulation via discretisation

    Stephen James, Kentaro Wada, Tristan Laidlow, and An- drew J Davison. Coarse-to-fine q-attention: Efficient learn- ing for visual robotic manipulation via discretisation. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13739–13748, 2022. 2

    work page 2022

  19. [19]

    Bc-z: Zero-shot task generalization with robotic imitation learning

    Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Fred- erik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning. InConference on Robot Learning, pages 991–

    work page

  20. [20]

    Vima: General robot manipulation with multimodal prompts.arXiv preprint arXiv:2210.03094, 2(3):6, 2022

    Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anand- kumar, Yuke Zhu, and Linxi Fan. Vima: General robot manipulation with multimodal prompts.arXiv preprint arXiv:2210.03094, 2(3):6, 2022. 2

    Pith/arXiv arXiv 2022

  21. [21]

    Prismatic vlms: Investigating the design space of visually-conditioned language models

    Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh. Prismatic vlms: Investigating the design space of visually-conditioned language models. InForty-first International Conference on Machine Learning, 2024. 1, 2

    work page 2024

  22. [22]

    3d diffuser actor: Policy diffusion with 3d scene rep- resentations.arXiv preprint arXiv:2402.10885, 2024

    Tsung-Wei Ke, Nikolaos Gkanatsios, and Katerina Fragki- adaki. 3d diffuser actor: Policy diffusion with 3d scene rep- resentations.arXiv preprint arXiv:2402.10885, 2024. 5

    Pith/arXiv arXiv 2024

  23. [23]

    Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024. 1, 2, 5

    Pith/arXiv arXiv 2024

  24. [24]

    Llava-onevision: Easy visual task transfer

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Zi- wei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 2 9

    Pith/arXiv arXiv 2024

  25. [25]

    Cogact: A foundational vision- language-action model for synergizing cognition and action in robotic manipulation.arXiv preprint arXiv:2411.19650,

    Qixiu Li, Yaobo Liang, Zeyu Wang, Lin Luo, Xi Chen, Mozheng Liao, Fangyun Wei, Yu Deng, Sicheng Xu, Yizhong Zhang, et al. Cogact: A foundational vision- language-action model for synergizing cognition and action in robotic manipulation.arXiv preprint arXiv:2411.19650,

    Pith/arXiv arXiv

  26. [26]

    Lavida: A large diffu- sion language model for multimodal understanding.arXiv preprint arXiv:2505.16839, 2025

    Shufan Li, Konstantinos Kallidromitis, Hritik Bansal, Akash Gokul, Yusuke Kato, Kazuki Kozuka, Jason Kuen, Zhe Lin, Kai-Wei Chang, and Aditya Grover. Lavida: A large diffu- sion language model for multimodal understanding.arXiv preprint arXiv:2505.16839, 2025. 2

    Pith/arXiv arXiv 2025

  27. [27]

    Evaluating real-world robot manipulation policies in simulation.arXiv preprint arXiv:2405.05941, 2024

    Xuanlin Li, Kyle Hsu, Jiayuan Gu, Karl Pertsch, Oier Mees, Homer Rich Walke, Chuyuan Fu, Ishikaa Lunawat, Isabel Sieh, Sean Kirmani, et al. Evaluating real-world robot manipulation policies in simulation.arXiv preprint arXiv:2405.05941, 2024. 5

    Pith/arXiv arXiv 2024

  28. [28]

    Towards generalist robot policies: What matters in building vision-language-action models.arXiv preprint arXiv:2412.14058, 2024

    Xinghang Li, Peiyan Li, Minghuan Liu, Dong Wang, Jirong Liu, Bingyi Kang, Xiao Ma, Tao Kong, Hanbo Zhang, and Huaping Liu. Towards generalist robot policies: What matters in building vision-language-action models.arXiv preprint arXiv:2412.14058, 2024. 2

    Pith/arXiv arXiv 2024

  29. [29]

    Vision-language foundation models as effective robot imitators.arXiv preprint arXiv:2311.01378,

    Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, et al. Vision-language foundation models as effective robot imitators.arXiv preprint arXiv:2311.01378,

    Pith/arXiv arXiv

  30. [30]

    Discrete diffusion vla: Bringing dis- crete diffusion to action decoding in vision-language-action policies.arXiv preprint arXiv:2508.20072, 2025

    Zhixuan Liang, Yizhuo Li, Tianshuo Yang, Chengyue Wu, Sitong Mao, Liuao Pei, Xiaokang Yang, Jiangmiao Pang, Yao Mu, and Ping Luo. Discrete diffusion vla: Bringing dis- crete diffusion to action decoding in vision-language-action policies.arXiv preprint arXiv:2508.20072, 2025. 3, 5

    Pith/arXiv arXiv 2025

  31. [31]

    Improved baselines with visual instruction tuning

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296–26306, 2024. 1, 2

    work page 2024

  32. [32]

    Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023. 1, 2

    2023

  33. [33]

    Longllada: Unlocking long context capabilities in diffusion llms.arXiv preprint arXiv:2506.14429, 2025

    Xiaoran Liu, Zhigeng Liu, Zengfeng Huang, Qipeng Guo, Ziwei He, and Xipeng Qiu. Longllada: Unlocking long context capabilities in diffusion llms.arXiv preprint arXiv:2506.14429, 2025. 2

    arXiv 2025

  34. [34]

    dllm-cache: Accelerating diffusion large language models with adaptive caching.arXiv preprint arXiv:2506.06295,

    Zhiyuan Liu, Yicun Yang, Yaojie Zhang, Junjie Chen, Chang Zou, Qingyuan Wei, Shaobo Wang, and Linfeng Zhang. dllm-cache: Accelerating diffusion large language models with adaptive caching.arXiv preprint arXiv:2506.06295,

    Pith/arXiv arXiv

  35. [35]

    Discrete diffusion modeling by estimating the ratios of the data distri- bution.arXiv preprint arXiv:2310.16834, 2023

    Aaron Lou, Chenlin Meng, and Stefano Ermon. Discrete diffusion modeling by estimating the ratios of the data distri- bution.arXiv preprint arXiv:2310.16834, 2023. 1

    Pith/arXiv arXiv 2023

  36. [36]

    Tess: Text-to-text self-conditioned simplex diffu- sion.arXiv preprint arXiv:2305.08379, 2023

    Rabeeh Karimi Mahabadi, Hamish Ivison, Jaesung Tae, James Henderson, Iz Beltagy, Matthew E Peters, and Arman Cohan. Tess: Text-to-text self-conditioned simplex diffu- sion.arXiv preprint arXiv:2305.08379, 2023. 2

    Pith/arXiv arXiv 2023

  37. [37]

    Octo: An open- source generalist robot policy

    Oier Mees, Dibya Ghosh, Karl Pertsch, Kevin Black, Homer Rich Walke, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, Jianlan Luo, et al. Octo: An open- source generalist robot policy. InFirst Workshop on Vision- Language Models for Navigation and Manipulation at ICRA 2024, 2024. 2, 5

    work page 2024

  38. [38]

    Calvin: A benchmark for language- conditioned policy learning for long-horizon robot manip- ulation tasks.IEEE Robotics and Automation Letters, 7(3):7327–7334, 2022

    Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wol- fram Burgard. Calvin: A benchmark for language- conditioned policy learning for long-horizon robot manip- ulation tasks.IEEE Robotics and Automation Letters, 7(3):7327–7334, 2022. 5

    work page 2022

  39. [39]

    Large language diffusion models.arXiv preprint arXiv:2502.09992, 2025

    Shen Nie, Fengqi Zhu, Zebin You, Xiaolu Zhang, Jingyang Ou, Jun Hu, Jun Zhou, Yankai Lin, Ji-Rong Wen, and Chongxuan Li. Large language diffusion models.arXiv preprint arXiv:2502.09992, 2025. 1, 2, 3

    Pith/arXiv arXiv 2025

  40. [40]

    Llarva: Vision-action instruction tuning enhances robot learning

    Dantong Niu, Yuvan Sharma, Giscard Biamby, Jerome Quenum, Yutong Bai, Baifeng Shi, Trevor Darrell, and Roei Herzig. Llarva: Vision-action instruction tuning enhances robot learning. In8th Annual Conference on Robot Learn- ing. 2

    work page

  41. [41]

    Your absorbing dis- crete diffusion secretly models the conditional distributions of clean data.arXiv preprint arXiv:2406.03736, 2024

    Jingyang Ou, Shen Nie, Kaiwen Xue, Fengqi Zhu, Jiacheng Sun, Zhenguo Li, and Chongxuan Li. Your absorbing dis- crete diffusion secretly models the conditional distributions of clean data.arXiv preprint arXiv:2406.03736, 2024. 2, 3

    Pith/arXiv arXiv 2024

  42. [42]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 4195–4205,

  43. [43]

    Robot learning with sen- sorimotor pre-training

    Ilija Radosavovic, Baifeng Shi, Letian Fu, Ken Goldberg, Trevor Darrell, and Jitendra Malik. Robot learning with sen- sorimotor pre-training. InConference on Robot Learning, pages 683–693. PMLR, 2023. 2

    work page 2023

  44. [44]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 2

    2022

  45. [45]

    Simple and effective masked dif- fusion language models.Advances in Neural Information Processing Systems, 37:130136–130184, 2024

    Subham Sahoo, Marianne Arriola, Yair Schiff, Aaron Gokaslan, Edgar Marroquin, Justin Chiu, Alexander Rush, and V olodymyr Kuleshov. Simple and effective masked dif- fusion language models.Advances in Neural Information Processing Systems, 37:130136–130184, 2024. 1

    work page 2024

  46. [46]

    Simplified and generalized masked diffu- sion for discrete data.Advances in neural information pro- cessing systems, 37:103131–103167, 2024

    Jiaxin Shi, Kehang Han, Zhe Wang, Arnaud Doucet, and Michalis Titsias. Simplified and generalized masked diffu- sion for discrete data.Advances in neural information pro- cessing systems, 37:103131–103167, 2024. 1

    work page 2024

  47. [47]

    Perceiver- actor: A multi-task transformer for robotic manipulation

    Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver- actor: A multi-task transformer for robotic manipulation. InConference on Robot Learning, pages 785–799. PMLR,

    work page

  48. [48]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020. 2

    Pith/arXiv arXiv 2010

  49. [49]

    Siglip 2: Multilingual vision-language en- coders with improved semantic understanding, localization, and dense features.arXiv preprint arXiv:2502.14786, 2025

    Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muham- mad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, et al. Siglip 2: Multilingual vision-language en- coders with improved semantic understanding, localization, and dense features.arXiv preprint arXiv:2502.14786, 2025. 3 10

    Pith/arXiv arXiv 2025

  50. [50]

    Open x-embodiment: Robotic learning datasets and rt-x models

    Quan Vuong, Sergey Levine, Homer Rich Walke, Karl Pertsch, Anikait Singh, Ria Doshi, Charles Xu, Jianlan Luo, Liam Tan, Dhruv Shah, et al. Open x-embodiment: Robotic learning datasets and rt-x models. InTowards General- ist Robots: Learning Paradigms for Scalable Skill Acquisi- tion@ CoRL2023, 2023. 1, 5

    work page 2023

  51. [51]

    Unleashing large-scale video generative pre- training for visual robot manipulation.arXiv preprint arXiv:2312.13139, 2023

    Hongtao Wu, Ya Jing, Chilam Cheang, Guangzeng Chen, Jiafeng Xu, Xinghang Li, Minghuan Liu, Hang Li, and Tao Kong. Unleashing large-scale video generative pre- training for visual robot manipulation.arXiv preprint arXiv:2312.13139, 2023. 5, 6

    Pith/arXiv arXiv 2023

  52. [52]

    Mmada: Mul- timodal large diffusion language models.arXiv preprint arXiv:2505.15809, 2025

    Ling Yang, Ye Tian, Bowen Li, Xinchen Zhang, Ke Shen, Yunhai Tong, and Mengdi Wang. Mmada: Mul- timodal large diffusion language models.arXiv preprint arXiv:2505.15809, 2025. 1, 2

    Pith/arXiv arXiv 2025

  53. [53]

    Dream 7b: Diffusion large language models.arXiv preprint arXiv:2508.15487, 2025

    Jiacheng Ye, Zhihui Xie, Lin Zheng, Jiahui Gao, Zirui Wu, Xin Jiang, Zhenguo Li, and Lingpeng Kong. Dream 7b: Diffusion large language models.arXiv preprint arXiv:2508.15487, 2025. 2, 3

    Pith/arXiv arXiv 2025

  54. [54]

    Dinoiser: Diffused conditional se- quence learning by manipulating noises.arXiv preprint arXiv:2302.10025, 2023

    Jiasheng Ye, Zaixiang Zheng, Yu Bao, Lihua Qian, and Mingxuan Wang. Dinoiser: Diffused conditional se- quence learning by manipulating noises.arXiv preprint arXiv:2302.10025, 2023. 2

    Pith/arXiv arXiv 2023

  55. [55]

    Llada-v: Large language diffusion models with visual instruction tuning

    Zebin You, Shen Nie, Xiaolu Zhang, Jun Hu, Jun Zhou, Zhiwu Lu, Ji-Rong Wen, and Chongxuan Li. Llada-v: Large language diffusion models with visual instruction tuning. arXiv preprint arXiv:2505.16933, 2025. 1, 2, 3, 6

    Pith/arXiv arXiv 2025

  56. [56]

    Robotic control via embodied chain-of-thought reasoning.arXiv preprint arXiv:2407.08693, 2024

    Michał Zawalski, William Chen, Karl Pertsch, Oier Mees, Chelsea Finn, and Sergey Levine. Robotic control via embodied chain-of-thought reasoning.arXiv preprint arXiv:2407.08693, 2024. 2

    Pith/arXiv arXiv 2024

  57. [57]

    Diffa: Large language diffusion models can lis- ten and understand.arXiv preprint arXiv:2507.18452, 2025

    Jiaming Zhou, Hongjie Chen, Shiwan Zhao, Jian Kang, Jie Li, Enzhi Wang, Yujie Guo, Haoqin Sun, Hui Wang, Aobo Kong, et al. Diffa: Large language diffusion models can lis- ten and understand.arXiv preprint arXiv:2507.18452, 2025. 2

    arXiv 2025

  58. [58]

    Chatvla: Unified multimodal un- derstanding and robot control with vision-language-action model.arXiv preprint arXiv:2502.14420, 2025

    Zhongyi Zhou, Yichen Zhu, Minjie Zhu, Junjie Wen, Ning Liu, Zhiyuan Xu, Weibin Meng, Ran Cheng, Yaxin Peng, Chaomin Shen, et al. Chatvla: Unified multimodal un- derstanding and robot control with vision-language-action model.arXiv preprint arXiv:2502.14420, 2025. 1, 2

    Pith/arXiv arXiv 2025

  59. [59]

    Llada 1.5: Variance-reduced prefer- ence optimization for large language diffusion models.arXiv preprint arXiv:2505.19223, 2025

    Fengqi Zhu, Rongzhen Wang, Shen Nie, Xiaolu Zhang, Chunwei Wu, Jun Hu, Jun Zhou, Jianfei Chen, Yankai Lin, Ji-Rong Wen, et al. Llada 1.5: Variance-reduced prefer- ence optimization for large language diffusion models.arXiv preprint arXiv:2505.19223, 2025. 2 11

    Pith/arXiv arXiv 2025