Pith. sign in

REVIEW 4 major objections 4 minor 15 references

A world action model can match or exceed deeper baselines with a single-layer action head, provided the head fuses representations from every layer of a pretrained video backbone after aligning their rotary position embeddings.

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

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

A single-layer action head with cross-layer KV fusion and RoPE realignment on a pretrained video DiT matches deeper World Action Models while cutting inference latency to 66.5 ms.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A genuinely useful efficiency result for WAMs, but the OOD headline number is selected on the same benchmark and should not be taken at face value. the 4 major comments →

arxiv 2608.02365 v1 pith:QROYAXFH submitted 2026-08-03 cs.AI cs.LGcs.RO

Faster-WAM: Do World Action Models Need Deep Action Modules?

Liheng Ma , Rui Heng Yang , Zhanguang Zhang , Mateo Clemente , Ziwen Hu , Tongtong Cao , Yingxue Zhang This is my paper
classification cs.AI cs.LGcs.RO
keywords world action modelsvideo diffusion transformersdocking interfaceskey-value fusionrotary position embedding alignmentrobot manipulationinference latencyout-of-distribution generalization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper asks whether World Action Models need action modules as deep as their video backbones, and answers no. It introduces Dock of Transformers (DoT), a video-centric design in which a pretrained video transformer is treated as a representation hub and lightweight output heads are attached through a docking interface. Its instantiation Faster-WAM uses a single-layer action transformer docked onto a 30-layer video backbone through KV-Fusion and video–action RoPE alignment. Without embodied pretraining it reaches 98.5% on LIBERO, 89.2% on RoboTwin 2.0, and 75.0% on LIBERO-Plus, while generating a 32-step action chunk in 66.5 ms, a 3.2× speedup over Fast-WAM. The point is that deep action-specific computation may be unnecessary if the video world model already encodes physical dynamics.

Core claim

That a one-layer action head can match or beat deeper action modules by reading fused representations from every layer of a full-depth video backbone. The docking interface aggregates keys and values across all 30 video layers (KV-Fusion) and reconciles the positional coordinate systems: video keys stored in 3D rotary coordinates are unrotated, fused in canonical space, then rerotated with 1D rotary embeddings so attention with action queries has proper relative-position semantics. With this interface, the action head's depth is decoupled from the video backbone's depth; the video backbone processes only the conditioning frame, and the head denoises the action chunk. The empirical claim is t

What carries the argument

Dock of Transformers (DoT): a design principle that makes the video backbone a representation hub and connects task heads through docking interfaces, exposing multi-layer key/value caches without requiring one-to-one layer correspondence. Faster-WAM instantiates it with KV-Fusion, which channel-mixes video keys/values into the action head's feature space and linearly mixes them across layers with a learnable cross-layer aggregation matrix, and video–action RoPE alignment, which strips 3D rotary embeddings, fuses in canonical space, and re-applies 1D rotary embeddings so mixed-context attention remains relative-position-correct. The mechanism's work is to let a single action layer draw on dis

Load-bearing premise

That a pretrained video world model already contains the physical-dynamics information needed to predict actions, so a shallow head only needs to read it out; if the video hub lacks this information, the single-layer action head cannot compensate.

What would settle it

Train Faster-WAM with the video backbone frozen and its weights progressively shuffled or randomly reinitialized. If action success stays high, the action head is not actually exploiting learned video dynamics and the 'video hub already captures physics' explanation fails; if success collapses as backbone quality degrades, the hub premise is supported.

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

If this is right

  • Action-module depth is not a bottleneck: a single-layer action head suffices for strong LIBERO and RoboTwin success.
  • WAM latency can drop to VLA-level: 66.5 ms per 32-step action chunk, 3.2× faster than Fast-WAM, on a 24 GB consumer GPU.
  • Docking all-layer representations improves out-of-distribution generalization: LIBERO-Plus rises from 51.5% to 75.0% over Fast-WAM, with large gains on camera and sensor perturbations.
  • Removing text cross-attention from the action head can help when instruction conditioning is routed through the language-conditioned video hub.

Where Pith is reading between the lines

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

  • The DoT docking pattern may generalize to other output heads — navigation waypoints, affordance maps, reward models — so the same pretrained video hub could serve multiple downstream policies without re-running deep action towers.
  • The learnable cross-layer aggregation matrix can be read as an adaptive depth selector; the layer-signal analysis in the ablation hints that studying how these weights shift across perturbation types could yield a more principled alternative to fixed middle-layer heuristics.
  • The unrotate–fuse–rerotate RoPE alignment is a general cure for cross-modal positional mismatch and could transfer to other multimodal transformers that mix tokens with incompatible rotary embeddings.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes Dock of Transformers (DoT), a video-centric architecture in which a pretrained video world model acts as a representation hub and lightweight task heads are attached through explicit docking interfaces. The instantiation Faster-WAM docks a single-layer action DiT onto a 30-layer video backbone using two mechanisms: KV-Fusion, which aggregates keys and values across all video layers, and video-action RoPE alignment, which corrects the rotary-position basis mismatch between the video keys and action queries. The authors report competitive success rates on LIBERO (98.5%), RoboTwin 2.0 (89.17%), and LIBERO-Plus (75.0%), with a 66.5 ms end-to-end latency for a 32-step action chunk, a 3.2× speedup over Fast-WAM. They conclude that a deep action module is not necessary if the video backbone already encodes physical dynamics.

Significance. If the claims hold, the paper makes a useful architectural contribution: it decouples action-head depth from video-backbone depth, which is a practical bottleneck in existing MoT-style WAMs. The mathematical analysis in Appendix A.1 of the 3D-RoPE/1D-RoPE mismatch is clean and self-contained, and the controlled latency comparison is a strength. The proposed design is simple and likely reproducible. However, the paper's most distinctive quantitative claim — strong out-of-distribution generalization on LIBERO-Plus — is currently supported by a single run that was selected by ablating on the same benchmark, which materially weakens the evidence. The clean LIBERO and RoboTwin comparisons are not invalidated, but the OOD-generalization conclusion needs stronger empirical support before it can be regarded as established.

major comments (4)
  1. [§4.3, Fig. 3a, Table 2] The headline LIBERO-Plus result, 75.0%, is the endpoint of a sequential ablation performed on the same LIBERO-Plus benchmark. Each of the four design choices — final-layer-only access, KV-Fusion, RoPE alignment, and removal of text cross-attention — was retained because it improved LIBERO-Plus success in Fig. 3a. Table 2 then reports the final configuration's 75.0% as the method's OOD generalization performance. This is a selected maximum over the explored design path, not an unbiased estimate; the 23.5-point gap over Fast-WAM may be inflated by within-benchmark model selection. This issue is load-bearing for the strongest claim. Please provide a held-out OOD benchmark not used in any design decision, or a nested/hold-out split of LIBERO-Plus for selection and evaluation, or at least report the unselected design variants with variance to show the final choice is not an argmax over noise.
  2. [§4.2 / Table 2] No error bars or multiple-seed results are reported for Faster-WAM or for baselines. LIBERO-Plus baselines are quoted from Zhang et al. (2026), and the variance of both the baselines and the reported Faster-WAM number is unknown. Since the paper's most distinctive quantitative claim is the 23.5-point OOD improvement, single runs without variance and without a protocol check against the baseline source are insufficient. I request mean and standard deviation over at least three seeds for Faster-WAM, and ideally for the compared methods under the same evaluation code, or a clear statement of why the benchmark is deterministic and how the baseline numbers were reproduced.
  3. [§4.2.1, Table 1] The latency comparison's consistency is unclear. The Table 1 footnote says Fast-WAM latency sums separately measured T5 and model latency, while end-to-end latency for other methods is described as including 'input feature encoding' but the exact inclusion of the language encoder is not stated for each method. The abstract claims the lowest end-to-end latency in a controlled comparison and a 3.2× speedup, so the measurement protocol (GPU model, batch size, dtype, text-encoder inclusion, denoising steps, and memory constraints) must be documented identically for all compared methods.
  4. [§3.1.2, Eqs. (19)–(20)] The RoPE alignment derivation in Appendix A.1 is mathematically sound, but the definition of the 1D position b_j assigned to each fused video token is missing. If b_j is a single temporal index shared by all tokens of the conditioning frame, then spatial (y,x) coordinates are not represented in the action head's positional encoding; if b_j is a flattened scalar, the paper's own Eq. (5) argues that flattening does not resolve the mismatch. Please specify b_j, justify the choice, and discuss what positional information is retained or discarded. This is needed to assess whether the alignment truly restores a common basis or silently removes spatial positional information.
minor comments (4)
  1. [§4.1] The statement 'we jointly optimize the video backbone' while 'keeping the video VAE frozen' is clear, but the paper should state explicitly whether the video backbone is initialized with the Wan2.2 pretrained weights and whether any text-conditioning cross-attention blocks in the backbone are frozen or fine-tuned. This affects the interpretation of 'without additional embodied pretraining.'
  2. [Fig. 3b] The color scale and the meaning of 'Signal Intensity' are not defined. Please describe how the learned layer-mixing coefficients A_h are visualized and whether the values are normalized across heads or layers.
  3. [§4.2.1, Table 1] The table footnotes are confusing: the * marker appears next to 'N/A' for LingBot-VA and Motus, but the footnote says 'LIBERO latency not measured: exceeds memory constraints of 24 GB.' It would be clearer to state separately that these models could not be run on the 24 GB GPU and that their latency numbers are therefore not reported.
  4. [Eq. (2)] The mode-n product notation in Eq. (2), '×4 W^K', is defined by Definition 1, but the reshape operations are written compactly. A short example or a clarification of the tensor index order would improve readability.

Circularity Check

2 steps flagged

LIBERO-Plus OOD claim is a selected maximum (Fig. 3a ablation on the same benchmark) and its baselines come from an overlapping-author paper; core LIBERO/RoboTwin results are independent.

specific steps
  1. fitted input called prediction [Sec. 4.3, Fig. 3a; Sec. 4.2.3, Table 2]
    "Figure 3a presents a sequential ablation of the design choices leading to Faster-WAM on LIBERO-Plus. ... Finally, removing text cross-attention reduces complexity of the action head while further improving the success rate to 75.0%."

    The headline OOD success (75.0%) in Table 2 is the endpoint of a sequential ablation whose selection criterion is LIBERO-Plus success (60.3 -> 66.8 -> 71.3 -> 75.0). The same benchmark is then offered as evidence of out-of-distribution generalization. The four design choices (final-layer access, KV-Fusion, RoPE alignment, no text cross-attention) were chosen by maximizing LIBERO-Plus accuracy, so the 23.5-point gain over Fast-WAM is a selected maximum over the explored design path, not an unbiased prediction of generalization. This is a fitted input (design) reported as a prediction (OOD score) on the same data.

  2. self citation load bearing [Table 2 caption; Sec. 4.2.3]
    "Table 2: LIBERO-Plus success (%) across perturbations; baselines are from (Zhang et al., 2026)."

    Zhang et al. (2026) is co-authored by Zhanguang Zhang, Rui Heng Yang, Tongtong Cao, and Yingxue Zhang, who are also authors of the present paper. The Fast-WAM baseline (51.5%) and all other Table 2 baselines are taken from this overlapping-author prior work and are not re-measured here. The central OOD comparison (75.0 vs 51.5, a 23.5-point gap) therefore rests on a load-bearing self-citation rather than an independent evaluation in this paper. This does not make the result false, but it is not independent support under the stated rules.

full rationale

The paper's mathematical core is self-contained: the Appendix A.1 unrotate-fuse-rerotate argument is a genuine algebraic identity, and the DoT equations (1)-(3), (18)-(20) do not define the claimed result in terms of its inputs. No uniqueness theorem is imported, and no existing result is merely renamed. The main architectural claim—that a one-layer action head docked to a 30-layer video hub is competitive on LIBERO (98.5%) and RoboTwin 2.0 (89.17%)—is supported by Table 1 with external baselines and is not circular. The circularity concern is concentrated in the distinctive OOD-generalization claim. The 75.0% LIBERO-Plus number is the same number reached at the end of a sequential ablation performed on LIBERO-Plus (Sec. 4.3, Fig. 3a), so the design was selected using the benchmark that is then reported as the evidence of generalization; this is a selected maximum, not an unbiased held-out prediction. In addition, the baseline numbers defining the 23.5-point OOD gap are sourced from a prior robustness study with overlapping authorship, making the comparison load-bearing on self-citation. These issues weaken the OOD claim but do not invalidate the clean benchmark results or the latency measurements, so the overall circularity is partial rather than total.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 1 invented entities

The central claim rests mainly on learned parameters (channel mixing, layer aggregation, action head) that are optimized on the benchmarks. No new physical entities are postulated. The key domain assumption is that the video hub already contains the needed dynamics.

free parameters (3)
  • W^K, W^V (channel mixing matrices) = learned, D'×D' each; ~20M params total in KV-Fusion
    Remap video keys/values into action-head feature space (Eq. 2). Trainable parameters optimized on the target benchmarks.
  • A_h (cross-layer aggregation matrix per head) = learned, L_axL_v per head
    Fuses representations across video layers for each attention head (Eq. 3). Learned during training.
  • Action head parameters = single Transformer layer, hidden 1024, attention width 3072, ~30M params
    The single-layer action DiT, optimized on target benchmarks.
axioms (4)
  • domain assumption Pretrained video world models already capture physical dynamics sufficient for action prediction
    Stated in Sec. 3 as the 'central premise of WAMs'. If false, the DoT design loses its motivation and the single-layer head cannot compensate.
  • standard math RoPE orthogonality and relative-position identity
    Used in Appx. A.1 to justify undoing 3D RoPE and re-applying 1D RoPE; standard property of rotary embeddings.
  • domain assumption Evaluation baselines from cited works are directly comparable
    Tables 1-2 take baseline success and latency from prior papers. Comparability across different eval protocols and hardware is assumed.
  • domain assumption Removing text cross-attention from the action head does not lose information
    The paper assumes language conditioning is fully routed through the video hub. Validated in ablation on LIBERO-Plus, but this is a design assumption.
invented entities (1)
  • DoT (Dock of Transformer) / docking interface no independent evidence
    purpose: Architectural framework enabling lightweight task heads to read from all layers of a video backbone via KV-Fusion and RoPE realignment.
    Not a physical entity; it is a new architecture. Its evidence is the paper's benchmark results, which are internal. No external falsifiable prediction is made beyond the reported success rates.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Faster-WAM: Do World Action Models Need Deep Action Modules?." pith.science (2026). https://pith.science/paper/QROYAXFH

@misc{pith2026260802365,
  author       = {Pith},
  title        = {Pith review of: Faster-WAM: Do World Action Models Need Deep Action Modules?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QROYAXFH}},
  note         = {Machine review of arXiv:2608.02365}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

World Action Models (WAMs) couple robot action prediction with video world models. Existing WAMs with shared-backbone and Mixture-of-Transformers designs generally tie the depth of the action module to that of the video backbone, resulting in substantial computational overhead and high inference latency. To address this limitation, we introduce Dock of Transformer (DoT), a video-centric design principle that treats a pretrained video Transformer as a representation hub and connects lightweight output-heads through docking interfaces. This enables flexible output-head design while providing direct access to representations from all layers of the backbone. We then introduce \textbf{Faster-WAM}, an instantiation of DoT for WAMs, which docks a single-layer action head onto a 30-layer video backbone. The docking interface fuses keys and values from all video layers and applies RoPE realignment. Without additional embodied pretraining, Faster-WAM achieves competitive performance on LIBERO and RoboTwin 2.0 while demonstrating strong out-of-distribution generalization on LIBERO-Plus. Faster-WAM also achieves the lowest end-to-end latency in our controlled comparison, requiring only 66.5 ms per inference --- a \(3.2\times\) speedup over Fast-WAM. Overall, these results demonstrate that the video-centric DoT architecture supports flexible task-specific head design while delivering low inference latency, strong action-prediction performance, and robust generalization.

Figures

Figures reproduced from arXiv: 2608.02365 by Liheng Ma, Mateo Clemente, Rui Heng Yang, Tongtong Cao, Yingxue Zhang, Zhanguang Zhang, Ziwen Hu.

Figure 1
Figure 1. Figure 1: Comparison of representative WAM architectures. MoT uses one-to-one video–action layer correspondence, while H-Bridge (Team et al., 2026) restricts cross-stream interaction to fixed intermediate layers. In contrast, DoT treats the video backbone as a representation hub that task-specific nodes access through flexible docking interfaces. action inference to computation through the full video backbone. Mixtu… view at source ↗
Figure 2
Figure 2. Figure 2: DoT architecture treats a pretrained video DiT as a representation hub and attaches [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Analysis of Faster-WAM on LIBERO-Plus: (a) sequential design contributions [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] 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

15 extracted references · 1 canonical work pages

  1. [3]

    Tianxing Chen, Zanxin Chen, Baijun Chen, Zijian Cai, Yibin Liu, Zixuan Li, Qiwei Liang, Xianliang Lin, Yiheng Ge, Zhenyu Gu, et al

    URL https://openreview.net/forum?id=Oz335dV48X. Tianxing Chen, Zanxin Chen, Baijun Chen, Zijian Cai, Yibin Liu, Zixuan Li, Qiwei Liang, Xianliang Lin, Yiheng Ge, Zhenyu Gu, et al. Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation. arXiv preprint arXiv:2506.18088,

    Pith/arXiv arXiv

  2. [5]

    Yucheng Hu, Yanjiang Guo, Pengchao Wang, Xiaoyu Chen, Yen-Jen Wang, Jianke Zhang, Koushil Sreenath, Chaochao Lu, and Jianyu Chen

    URL https://arxiv.org/abs/2602.10717. Yucheng Hu, Yanjiang Guo, Pengchao Wang, Xiaoyu Chen, Yen-Jen Wang, Jianke Zhang, Koushil Sreenath, Chaochao Lu, and Jianyu Chen. Video prediction policy: A generalist robot policy with predictive visual representations. In Aarti Singh, Maryam Fazel, Daniel Hsu, Simon Lacoste-Julien, Felix Berkenkamp, Tegan Maharaj, K...

    arXiv

  3. [7]

    Light-wam: Efficient world action models with state-fusion action decoding.arXiv preprint arXiv:2606.08242, 2026d

    Ziang Li, Dongzhou Cheng, Yibin Wang, Shiyue Wang, Xiaoyang Xu, Lingxuan Weng, Juan Wang, and Jiaqi Wang. Light-wam: Efficient world action models with state-fusion action decoding.arXiv preprint arXiv:2606.08242, 2026d. Ziang Li, Dongzhou Cheng, Yibin Wang, Shiyue Wang, Xiaoyang Xu, Lingxuan Weng, Juan Wang, and Jiaqi Wang. Light-wam: Efficient world act...

    Pith/arXiv arXiv

  4. [8]

    Teli Ma, Jia Zheng, Zifan Wang, Chunli Jiang, Andy Cui, Junwei Liang, and Shuo Yang

    URL https://proceedings.neurips.cc/paper_files/paper/2023/file/ 8c3c666820ea055a77726d66fc7d447f-Paper-Datasets_and_Benchmarks.pdf. Teli Ma, Jia Zheng, Zifan Wang, Chunli Jiang, Andy Cui, Junwei Liang, and Shuo Yang. Dit4dit: Jointly modeling video dynamics and actions for generalizable robot control. arXiv preprint arXiv:2603.10448,

    arXiv 2023

  5. [9]

    Jonas Pai, Liam Achenbach, Victoriano Montesinos, Benedek Forrai, Oier Mees, and Elvis Nava

    URL https://arxiv.org/abs/2503.14734. Jonas Pai, Liam Achenbach, Victoriano Montesinos, Benedek Forrai, Oier Mees, and Elvis Nava. mimic-video: Video-action models for generalizable robot control beyond vlas,

    Pith/arXiv arXiv

  6. [10]

    William Peebles and Saining Xie

    URLhttps://arxiv.org/abs/2512.15692. William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF international conference on computer vision, pp. 4195–4205,

    Pith/arXiv arXiv

  7. [11]

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu

    URL https://arxiv.org/abs/ 2506.01844. Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063,

    Pith/arXiv arXiv

  8. [12]

    URL https://arxiv.org/abs/2604.27792. Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jingren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang, Pandeng Li, Pingyu Wu, Ruihang Chu, Ruili Feng,...

    Pith/arXiv arXiv

  9. [13]

    URL https://arxiv.org/abs/2503.20314. Angen Ye, Boyuan Wang, Chaojun Ni, Guan Huang, Guosheng Zhao, Hao Li, Hengtao Li, Jie Li, Jindi Lv, Jingyu Liu, Min Cao, Peng Li, Qiuping Deng, Wenjun Mei, Xiaofeng Wang, Xinze Chen, Xinyu Zhou, Yang Wang, Yifan Chang, Yifan Li, Yukun Zhou, Yun Ye, Zhichao Liu, and Zheng Zhu. Gigaworld-policy: An efficient action-cent...

    Pith/arXiv arXiv 2026

  10. [14]

    URLhttps://arxiv.org/abs/2603.22078. Jinliang Zheng, Jianxiong Li, Zhihao Wang, Dongxiu Liu, Xirui Kang, Yuchun Feng, Yi- nan Zheng, Jiayin Zou, Yilun Chen, Jia Zeng, Tai Wang, Ya-Qin Zhang, Jingjing Liu, and Xianyuan Zhan. X-VLA: Soft-prompted transformer as scalable cross-embodiment vision-language-action model. InThe Fourteenth International Conference...

    Pith/arXiv arXiv

  11. [15]

    doi: 10.15607/RSS.2025.XXI.015. Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, Quan Vuong, Vincent Vanhoucke, Huong Tran, Radu Soricut, Anikait Singh, Jaspiar Singh, Pierre Sermanet, Pannag R. Sanketi, Grecia Salazar, Michael S. Ryoo, Krista Reymann, Kanishka Rao, Karl Pertsch, I...

    work page doi:10.15607/rss.2025.xxi.015 2025

  12. [2009]

    Efficient-wam: A 1b-parameter world-action model with low-cost future imagination.arXiv preprint arXiv:2606.10040, 2026a

    Jiajun Li, Tiecheng Guo, Yifan Ye, Rongyu Zhang, Xiaowei Chi, Qianpu Sun, Ying Li, Yunfan Lou, Yan Huang, Zhihe Lu, Meng Guo, and Shanghang Zhang. Efficient-wam: A 1b-parameter world-action model with low-cost future imagination.arXiv preprint arXiv:2606.10040, 2026a. URLhttps://arxiv.org/abs/2606.10040. Jiajun Li, Tiecheng Guo, Yifan Ye, Rongyu Zhang, Xi...

    Pith/arXiv arXiv

  13. [2023]

    Senyu Fei, Siyin Wang, Junhao Shi, Zihao Dai, Jikun Cai, Pengfang Qian, Li Ji, Xinzhe He, Shiduo Zhang, Zhaoye Fei, Jinlan Fu, Jingjing Gong, and Xipeng Qiu

    URL https://proceedings.neurips.cc/paper_files/paper/2023/file/ 1d5b9233ad716a43be5c0d3023cb82d0-Paper-Conference.pdf. Senyu Fei, Siyin Wang, Junhao Shi, Zihao Dai, Jikun Cai, Pengfang Qian, Li Ji, Xinzhe He, Shiduo Zhang, Zhaoye Fei, Jinlan Fu, Jingjing Gong, and Xipeng Qiu. Libero-plus: A progressive robustness benchmark for visual-language-action model...

    2023

  14. [2025]

    Jian Chen, Yesheng Liang, and Zhijian Liu

    URLhttps://arxiv.org/abs/2507.15493. Jian Chen, Yesheng Liang, and Zhijian Liu. DFlash: Block diffusion for flash speculative decoding. InForty-third International Conference on Machine Learning,

    Pith/arXiv arXiv

  15. [2026]

    Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Robert Equi, Chelsea Finn, Niccolo Fusai, Manuel Y. Galliker, Dibya Ghosh, Lachy Groom, Karol Hausman, brian ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Devin LeBlanc, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Allen Z. Ren, ...

    2025

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.