Pith.

Robotics · arXiv:2610.11401 v1 · submitted 8 Oct 2026 · 46 references

WAM-Cache: Staleness-Bounded KV Reuse for Efficient World Action Models

Kai Ding, Yang He, Ruijie Quan, Yi Yang

The paper says

“World Action Models (WAMs) enable generalist robot manipulation by conditioning an action expert on representations from a pretrained video Diffusion Transformer (DiT).”

more of the abstract

In closed-loop control, the video DiT runs at every chunk to encode the current observation into layerwise key-value (KV) pairs that the action expert queries. This prefill dominates the per-chunk computational cost, yet existing training-free accelerations leave it fully dense. We present WAM-Cache, a training-free framework that retains layerwise key-value representations across chunks and recomputes only a sparse refresh set of tokens. Crucially, we find that the intuitive heuristic of refreshing visually drifted tokens plateaus far below the dense baseline, even with an oracle predicting ground-truth KV drift. Downstream action accuracy is instead governed by where the action expert attends, not by what moved. WAM-Cache therefore selects the refresh set by uniting the action expert's cross-attention with visual latent surprise, complemented by a strict age bound that suppresses compounding error. On Fast-WAM, WAM-Cache cuts video DiT prefill FLOPs by 32-42% across RoboTwin 2.0, LIBERO, and real-world experiments, while staying within 0.7-1.8 percentage points of the dense policy in simulation and 2.5 points on a real robot.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 10 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

WAM-Cache is a solid, well-executed engineering contribution that cuts WAM prefill FLOPs by 32-42% with negligible success loss, but its generality is untested beyond Fast-WAM. This line is wrong

the referee pass did not run

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 46 references say.

What would settle it

Design a benchmark of multi-phase tasks (e.g., reach → grasp → transport → place) with minimal visual change between phases but sharp shifts in action-relevant tokens. Evaluate WAM-Cache (G=2) against dense inference and against a variant that forces a full prefill at each phase transition (oracle phase detector). If the gap between WAM-Cache and the full-refresh variant exceeds 3 percentage points on these tasks, the attention lag is a practically significant limitation that the age bound does not fully address.

from the skeptic's reading

How the review went

3 readings by deepseek-v4-flash · 10 Oct 2026

  1. Reader holds up, with conditions confidence high
  2. Skeptic agrees 46 s · 23,683 tokens read, 4,228 written
  3. Editor wrote the verdict above 13 s · 24,514 tokens read, 1,395 written

Not run on this paper: the referee, the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

This paper solves a practical bottleneck in World Action Model inference: the video DiT prefill, which dominates per-chunk cost. Their insight is that refreshing tokens based on visual change plateaus; instead, you should refresh where the action expert attends, plus an age bound to cap staleness. That is a clean, non-obvious finding, and the ablations back it up convincingly.

The empirical evaluation is thorough for a single architecture. Three sim benchmarks, real-robot validation, multiple baselines calibrated to the same FLOP budget, and sweep over the three free parameters. The paper ships zero new model training; everything is inference-only. The numbers are credible: 32-42% FLOP reduction for 0.7-1.8 pp success drop in sim, 40% reduction in reality with one lost trial out of 20.

Soft spots: only Fast-WAM is tested, so we don't know how well the rules transfer to other WAMs. The age bound caps staleness duration but not error magnitude—the paper notes this explicitly (b110), and LIBERO Long's 3.4 pp gap shows the limitation is real, not hypothetical. Real-world trials are 20 per task; that is enough for a point estimate but not for statistical rigor. These are acknowledged limitations, so I don't consider them fatal, but they do bound the paper's claims.

The paper is a clear engineering contribution with a well-motivated design and honest reporting. I would send it to peer review: it deserves a referee round. If accepted conditional on addressing generality—even a discussion of which architectural choices would break the assumption—that would strengthen it. I would not bring it to my reading group (too applied for my group's taste), but colleagues working on robot inference efficiency should read it.

The strongest objection

The paper's central mechanism relies on the action expert's cross-attention from the previous chunk as a proxy for which tokens matter in the current chunk. This assumption is weakest at phase transitions in multi-step tasks, where the attention from the previous phase may focus on tokens that become irrelevant, while missing critical new-phase tokens. The paper acknowledges this on LIBERO Long (b136) with a 3.4-percentage-point success gap (90.4% vs 93.8%), which is notably larger than the 0.65-point average gap. The age bound (G=2) limits the duration of staleness but not the magnitude of error that can accumulate before refresh. In tasks where critical tokens exhibit low visual drift and low inherited attention, such tokens could be omitted from the refresh set for up to G chunks, potentially causing cascading action errors. The paper's experiments involve relatively short episodes (5–10 chunks); longer or more phase-dense episodes could amplify this failure mode. The method's success in mitigating this is empirically strong on average, but the weakness is real and acknowledged.

The reader's weakest assumption correctly identifies the attention lag at phase transitions as the key risk. The paper's own data confirms that the largest per-suite gap occurs on LIBERO Long (multi-phase tasks). This is a load-bearing concern because the method's design centers on previous-chunk attention. However, the paper provides strong empirical evidence that the method works well on average across many tasks and that the age bound recovers most of the gap (Figure 2). The concrete test would help quantify the severity in worst-case phase-transition scenarios. I agree with the reader's assessment that the central claim is well-supported for typical manipulation but carries a condition for multi-phase tasks. Thus the CONDITIONAL verdict should remain unchanged.

Pith's reading of the paper (a machine's summary, not the authors')

World Action Models (WAMs) pair a video Diffusion Transformer (DiT) with an action expert for robot control. In closed-loop operation, the video DiT must re-encode every new observation, which dominates computation. WAM-Cache retains the DiT's layerwise key–value pairs across control cycles and recomputes only a sparse set of tokens each chunk. The key insight is that deciding which tokens to refresh based on visual change alone fails badly; instead, the refresh set must follow where the action expert attends. WAM-Cache selects tokens via a union of three signals: latent surprise (which tokens have drifted in embedding space), action-conditioned attention (which tokens the policy relied on in the previous chunk), and an age bound that forces refresh after a fixed number of reuses. On Fast-WAM, this cuts prefill FLOPs by 32–42% on RoboTwin 2.0, LIBERO, and a real robot, while losing at most 1.8 points in simulation and 2.5 points on hardware.

What carries the argument. The central mechanism is the refresh set R_t = S_t ∪ P_t ∪ A_t, where S_t contains tokens with the largest latent surprise (normalized distance from a cached reference embedding), P_t contains tokens with the largest action-expert cross-attention mass from the previous chunk (computed at zero extra cost), and A_t contains all tokens whose age counter has reached the bound G=2. This union ensures that tokens are refreshed if they have changed, if the policy cares about them, or if they have been cached too long. The prefill then runs only on this sparse set, updating the KV cache in place while leaving unselected tokens unchanged, and the action expert attends over the mixed cache without modification.

Core claim. The paper establishes that, for a World Action Model, the visual tokens whose key–value representations need refreshing across control chunks are not best chosen by measuring visual change or representation drift. Even an oracle with ground-truth knowledge of which cached tokens have drifted most leaves a 25-point gap in success rate. The correct criterion is where the action expert's cross-attention falls: the policy's own sensitivity to a stale token determines how badly the action degrades. WAM-Cache therefore combines the action expert's attention map from the previous chunk with a latent-surprise score that tracks embedding drift, and adds a strict age bound that forces every token to be recomputed after G=2 consecutive reuses. This union recovers 98% of the dense policy's success rate while removing 41.6% of the prefill FLOPs on RoboTwin 2.0, and the same rules transfer to LIBERO and a real AIRBOT Play arm without retuning.

The premise it rests on. The method assumes that the action expert's cross-attention from the previous chunk is a reliable proxy for which visual tokens will matter in the current chunk, an assumption that can fail at phase transitions in multi-step tasks where the policy suddenly shifts attention to previously unattended regions.

If the paper is right (Pith's inferences, not the authors')

WAM-Cache enables world action models to run at higher video resolutions or with more camera views without a proportional increase in compute, since the speedup relative to dense prefill grows with token count (reaching 1.67× at 9× native tokens).

The method composes with other orthogonal speedups such as fewer action-denoising steps, model distillation, and denoiser caching, because it modifies only the prefill stage and leaves the action expert untouched.

Deploying WAMs on edge robots becomes more practical: a 35% reduction in DiT FLOPs per chunk at N=2 steps translates to lower latency and energy use, closing the gap to faster but less general VLAs.

The core finding—that cache refresh should follow policy attention rather than visual drift—is a design principle that can transfer to other decoupled perception-action systems beyond the specific WAM architecture studied here.

If the action expert's attention is a good proxy for downstream sensitivity, then the same principle could apply to other multi-module robotic systems where a slow perception module feeds a fast controller. For example, a vision-language model serving as a scene descriptor for a motion planner might benefit from KV reuse guided by the planner's attention.

The age bound G=2 is a simple but effective heuristic; a more principled bound could be derived from per-token error accumulation models rather than a uniform counter, potentially recovering more FLOPs without sacrificing accuracy on multi-phase tasks.

The method's reliance on previous-chunk attention means it may struggle at the very first chunk after a task switch or environmental reset, but the paper acknowledges this and treats the first chunk as a dense prefill. In lifelong deployment, a reset detection mechanism could trigger a fresh dense prefill when attention lag is likely critical.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

World Action Models (WAMs) enable generalist robot manipulation by conditioning an action expert on representations from a pretrained video Diffusion Transformer (DiT). In closed-loop control, the video DiT runs at every chunk to encode the current observation into layerwise key-value (KV) pairs that the action expert queries. This prefill dominates the per-chunk computational cost, yet existing training-free accelerations leave it fully dense. We present WAM-Cache, a training-free framework that retains layerwise key-value representations across chunks and recomputes only a sparse refresh set of tokens. Crucially, we find that the intuitive heuristic of refreshing visually drifted tokens plateaus far below the dense baseline, even with an oracle predicting ground-truth KV drift. Downstream action accuracy is instead governed by where the action expert attends, not by what moved. WAM-Cache therefore selects the refresh set by uniting the action expert's cross-attention with visual latent surprise, complemented by a strict age bound that suppresses compounding error. On Fast-WAM, WAM-Cache cuts video DiT prefill FLOPs by 32-42% across RoboTwin 2.0, LIBERO, and real-world experiments, while staying within 0.7-1.8 percentage points of the dense policy in simulation and 2.5 points on a real robot.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-09

The authors' main claim, as the classifier picked it out 0.57

“On Fast-WAM, WAM-Cache cuts video DiT prefill FLOPs by 32-42% across RoboTwin 2.0, LIBERO, and real-world experiments, while staying within 0.7-1.8 percentage points of the dense policy in simulation and 2.5 points on a real robot.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.08 · states limitation 0.83 · human participants 0.07 · controlled comparison 0.95 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.96 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
World Action Models (WAMs) enable generalist robot manipulation by conditioning an action expert on representations from a pretrained video Diffusion Transformer (DiT). background 1.00 0.07 0.02 asserted 1.00
In closed-loop control, the video DiT runs at every chunk to encode the current observation into layerwise key-value (KV) pairs that the action expert queries. background 0.98 0.10 0.02 asserted 1.00
This prefill dominates the per-chunk computational cost, yet existing training-free accelerations leave it fully dense. background 0.51 0.14 0.03 asserted 1.00
We present WAM-Cache, a training-free framework that retains layerwise key-value representations across chunks and recomputes only a sparse refresh set of tokens. contribution 0.76 0.86 0.15 asserted 1.00
Crucially, we find that the intuitive heuristic of refreshing visually drifted tokens plateaus far below the dense baseline, even with an oracle predicting ground-truth KV drift. result 0.99 0.94 0.16 asserted 1.00
Downstream action accuracy is instead governed by where the action expert attends, not by what moved. interpretation 0.75 0.87 0.47 asserted 1.00
WAM-Cache therefore selects the refresh set by uniting the action expert's cross-attention with visual latent surprise, complemented by a strict age bound that suppresses compounding error. method 0.97 0.90 0.28 asserted 1.00
On Fast-WAM, WAM-Cache cuts video DiT prefill FLOPs by 32-42% across RoboTwin 2.0, LIBERO, and real-world experiments, while staying within 0.7-1.8 percentage points of the dense policy in simulation and 2.5 points on a real robot. result 1.00 0.96 0.40 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.95

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-09: 1 request, 6,412 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (5)
Figure 1 of the paper
Figure 1. The WAM-Cache inference pipeline. (a) Sparse refresh selection at chunk t: three complementary signals construct the refresh set Rt. Latent surprise measures feature drift from a token’s reference embedding; action-conditioned selection identifies tokens heavily attended to by the action expert in the preceding chunk; and an age counter forces recomputation after G consecu￾tive reuses. (b) Sparse prefill and mixed-cache execution: the video DiT executes a sparse prefill exclusively over Rt, updating layerwise keys and values in-place. The action expert then denoises the next action…
Figure 2 of the paper
Figure 2. Sequential ablation of caching signals on clean RoboTwin 2.0. The first two configura￾tions run without the age bound. The next three subsections build the refresh rule step by step. We start from the intuitive rule of refreshing what changed, and each later signal addresses a failure of the preceding configura￾tion
Figure 3 of the paper
Figure 3. Prefill speedup in CUDA latency of the default clean and randomized configurations. Dashed lines mark the bound 1/ρ¯
Figure 4 of the paper
Figure 4. Hyperparameter sweeps on clean RoboTwin 2.0. (a) Age bound
Figure 5 of the paper
Figure 5. Real-world tasks on the AIRBOT Play arm, viewed from the external camera. Top: picking

Figures are the authors', reproduced from the paper.

References (46)
  1. [1] Arman Akbari, Ci Zhang, Arash Akbari, Lin Zhao, Yixiao Chen, Weiwei Chen, Xuan Zhang, Geng Yuan, and Yanzhi Wang. Flash-W AM: Modality-Aware Distillation for World Action Models. arXiv preprint arXiv:2606.05254, on PithPith's reviewarXiv
  2. [5] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token Merging: Your ViT But Faster.arXiv preprint arXiv:2210.09461, on PithPith's reviewarXiv
  3. [8] Chi-Lam Cheang, Guangzeng Chen, Ya Jing, Tao Kong, Hang Li, Yifeng Li, Yuxiao Liu, Hong- tao Wu, Jiafeng Xu, Yichu Yang, Hanbo Zhang, and Minzhao Zhu. GR-2: A generative video-language-action model with web-scale knowledge for robot manipulation.arXiv preprint arXiv:2410.06158, on PithPith's reviewarXiv
  4. [9] Jialei Chen, Kai Wang, Kang Chen, Shuaihang Chen, Feng Gao, Wenhao Tang, Zhiyuan Li, Weilin Liu, Zhuyu Yao, Boxun Li, Yuanbo Xu, and Chao Yu. LaW AM: Latent world action models for efficient dynamics-aware robot policies.arXiv preprint arXiv:2606.15768, 2026a. Liang Chen, Haoz... on PithPith's reviewarXiv
  5. [10] Tianxing Chen, Zanxin Chen, Baijun Chen, Zijian Cai, Yibin Liu, Zixuan Li, Qiwei Liang, Xian- liang Lin, Yiheng Ge, Zhenyu Gu, et al. RoboTwin 2.0: A scalable data generator and bench- mark with strong domain randomization for robust bimanual robotic manipulation.arXiv preprin... on PithPith's reviewarXiv
  6. [11] Yuran Chen, Xinye Cai, Zhonglin Gong, and Yang Huang. High-Fidelity One-Step Generative Vi- suomotor Policy via Recursive Correction, Frequency Consistency, and Contrastive Flow Match- ing.arXiv preprint arXiv:2607.03865, 2026b. 10 Preprint Ju Dong, Liding Zhang, Lei Zhang, Yu... on PithPith's reviewarXiv
  7. [12] Yilun Du, Mengjiao Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Joshua B. Tenenbaum, Dale Schu- urmans, and Pieter Abbeel. Learning universal policies via text-guided video generation.arXiv preprint arXiv:2302.00111, on PithPith's reviewarXiv
  8. [13] arXiv:2308.13494. Weilun Feng, Guoxin Fan, Haotong Qin, Mingqiang Wu, Yuqi Li, Xiangqi Li, Zhulin An, Libo Huang, Dingrui Wang, Longlong Liao, Michele Magno, Yongjun Xu, and Chuanguang Yang. WorldCache: Accelerating World Models for Free via Heterogeneous Token Caching. InInte... on PithPith's reviewarXiv
  9. [14] arXiv:2603.06331. Yuanchun Guo and Bingyan Liu. Reflex: Real-Time VLA Control through Streaming Inference. In International Conference on Machine Learning (ICML), on PithPith's reviewarXiv
  10. [15] arXiv:2607.14695. 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.arXiv preprint arXiv:2412.14803, on PithPith's reviewarXiv
  11. [16] Wei Huang, Bohan Zhang, Chenzhi Liu, Isabella Liu, Shuai Yang, Weian Mao, Luozhou Wang, Yicheng Xiao, Weifeng Lin, Qixin Hu, Bryan Chu, Sifei Liu, Linxi Fan, Xiaojuan Qi, Song Han, and Yukang Chen. Long-W AM: Scaling the context of world-action models.arXiv preprint arXiv:2610.10528, on PithPith's reviewarXiv
  12. [17] Joel Jang, Seonghyeon Ye, Zongyu Lin, Jiannan Xiang, Johan Bjorck, Yu Fang, Fengyuan Hu, Spencer Huang, Kaushil Kundalia, Yen-Chen Lin, Loic Magne, Ajay Mandlekar, Avnish Narayan, You Liang Tan, Guanzhi Wang, Jing Wang, Qi Wang, Yinzhen Xu, Xiaohui Zeng, Kaiyuan Zheng, Ruijie ... on PithPith's reviewarXiv
  13. [18] 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, on PithPith's reviewarXiv
  14. [19] Moo Jin Kim, Yihuai Gao, Tsung-Yi Lin, Yen-Chen Lin, Yunhao Ge, Grace Lam, Percy Liang, Shuran Song, Ming-Yu Liu, Chelsea Finn, and Jinwei Gu. Cosmos policy: Fine-tuning video models for visuomotor control and planning.arXiv preprint arXiv:2601.16163, on PithPith's reviewarXiv
  15. [20] 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-W AM: A 1B-Parameter World-Action Model with Low-Cost Future Imagination.arXiv preprint arXiv:2606.10040, 2026a. Lin Li,... on PithPith's reviewarXiv
  16. [21] 11 Preprint Xiang Li, Yupeng Zheng, Songen Gu, Huailiang Ma, Feng Yu, Yuhang Zheng, Xian Nie, Shanshuai Yuan, Yujie Zang, Weize Li, Shuai Tian, Moyang Liu, Ya-Qin Zhang, and Wenchao Ding. Latent Action as Intention Enables Efficient Future Imagination for World Action Models.a... on PithPith's reviewarXiv
  17. [22] 52202/079017-0722. Zekai Li, Jiaming Tang, and Zhijian Liu. FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference.arXiv preprint arXiv:2608.27384, 2026e. Ziang Li, Dongzhou Cheng, Yibin Wang, Shiyue Wang, Xiaoyang Xu, Lingxuan Weng, Juan Wang, and Jiaqi W... on PithPith's reviewarXiv
  18. [23] Yue Liao, Pengfei Zhou, Siyuan Huang, Donglin Yang, Shengcong Chen, Yuxin Jiang, Yue Hu, Jingbin Cai, Si Liu, Jianlan Luo, Liliang Chen, Shuicheng Yan, Maoqing Yao, and Guanghui Ren. Genie Envisioner: A unified world foundation platform for robotic manipulation.arXiv preprint ... on PithPith's reviewarXiv
  19. [24] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. LIBERO: Benchmarking knowledge transfer for lifelong robot learning.arXiv preprint arXiv:2306.03310, on PithPith's reviewarXiv
  20. [25] Qi Luo, Shuaijun Liu, Hao Zhao, Kunlin Li, Xiaobo Wang, Ningxing Su, Dongsheng Wang, and Yun Chen. The Gate, Not the Cache: Gate Provenance Bounds the Closed-Loop Reliability of Training-Free VLA Token Skipping.arXiv preprint arXiv:2608.00391, on PithPith's reviewarXiv
show 26 more
  1. [26] Jiangran Lyu, Ziming Li, Xuesong Shi, Chaoyi Xu, Yizhou Wang, and He Wang. DyW A: Dynamics- adaptive world action model for generalizable non-prehensile manipulation.arXiv preprint arXiv:2503.16806, on PithPith's reviewarXiv
  2. [27] Liheng Ma, Rui Heng Yang, Zhanguang Zhang, Mateo Clemente, Ziwen Hu, Tongtong Cao, and Yingxue Zhang. Faster-W AM: Do World Action Models Need Deep Action Modules?arXiv preprint arXiv:2608.02365, on PithPith's reviewarXiv
  3. [28] Xinyin Ma, Gongfan Fang, and Xinchao Wang. DeepCache: Accelerating Diffusion Models for Free.arXiv preprint arXiv:2312.00858, on PithPith's reviewarXiv
  4. [29] Weian Mao, Xi Lin, Wei Huang, Yuxin Xie, Tianfu Fu, Bohan Zhuang, Song Han, and Yukang Chen. TriAttention: Efficient Long Reasoning with Trigonometric KV Compression.arXiv preprint arXiv:2604.04921, on PithPith's reviewarXiv
  5. [30] NVIDIA, Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvi- jit Chattopadhyay, Yongxin Chen, Yin Cui, Yifan Ding, Daniel Dworakowski, Jiaojiao Fan, Michele Fenzi, Francesco Ferroni, Sanja Fidler, Dieter Fox, Songwei Ge, Yunhao Ge, Jinwei G... on PithPith's reviewarXiv
  6. [31] 12 Preprint Jonas Pai, Liam Achenbach, Victoriano Montesinos, Benedek Forrai, Oier Mees, and Elvis Nava. mimic-video: Video-action models for generalizable robot control beyond VLAs.arXiv preprint arXiv:2512.15692, on PithPith's reviewarXiv
  7. [32] Yirui Sun, Guangyu Zhuge, Keliang Liu, Jie Gu, Shiqin Dai, Xinyu Bing, Zhongxue Gan, and Chunxu Tian. SANTS: A State-Adaptive Scheduler for World Action Models.arXiv preprint arXiv:2605.27947, on PithPith's reviewarXiv
  8. [33] Yinzhou Tang, Jingbo Xu, Yu Shang, Zihao Song, Chen Gao, Wei Wu, and Yong Li. Dreaming when Necessary: Advancing World Action Models with Adaptive Multi-Modal Reasoning.arXiv preprint arXiv:2606.07089, on PithPith's reviewarXiv
  9. [34] GigaWorld Team, Angyuan Ma, Boyuan Wang, Bohan Li, Chaojun Ni, Guo Li, Guan Huang, Gu- osheng Zhao, Hao Li, Hengtao Li, Jingyu Liu, Jiwen Lu, Qiuping Deng, Tingdong Yu, Xu- ancheng Xu, Xinyu Zhou, Xiuwei Xu, Xinze Chen, Xiaofeng Wang, Xiaoyu Tian, Yang Wang, Yifan Chang, Yukun... on PithPith's reviewarXiv
  10. [35] Siyin Wang, Junhao Shi, Zhaoyang Fu, Xinzhe He, Feihong Liu, Chenchen Yang, Yikang Zhou, Zhaoye Fei, Jingjing Gong, Jinlan Fu, Mike Zheng Shou, Xuanjing Huang, Xipeng Qiu, and Yu-Gang Jiang. World action models: The next frontier in embodied AI.arXiv preprint arXiv:2605.12090,... on PithPith's reviewarXiv
  11. [36] 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, on PithPith's reviewarXiv
  12. [37] Yuzhou Wu, Yuxin Zheng, Muchun Niu, Yishan Yang, Tianhao Liu, Hanwen Kang, Jiajian Jing, Linfeng Zhang, and Chuan Wen. Reducing Temporal Redundancy for Efficient Vision-Language- Action Inference.arXiv preprint arXiv:2607.12287, 2026a. Zhijie Wu, Kento Kawaharazuka, and Kei Ok... on PithPith's reviewarXiv
  13. [38] Ge Yan, Jinghao Liu, Yuzhi Fan, Lei Cai, Minwen Liao, Jesse Zhang, and Dieter Fox. Flex-π: A Multi-Stream World-Action Model with Compute Flexibility.arXiv preprint arXiv:2608.10860, on PithPith's reviewarXiv
  14. [39] 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. ... on PitharXiv
  15. [40] Yixiao Zeng, Jianlei Zheng, Chaoda Zheng, Shijia Chen, Mingdian Liu, Tongping Liu, Teng- wei Luo, Yu Zhang, Boyang Wang, Linkun Xu, Siyuan Lu, Bo Tian, and Xianming Liu. X- Cache: Cross-Chunk Block Caching for Few-Step Autoregressive World Models Inference.arXiv preprint arXiv... on PithPith's reviewarXiv
  16. [41] Qihang Zhang, Lin Li, Luyao Zhang, Shuai Yang, Yiming Luo, Shuaiting Li, Ruilin Wang, Junke Wang, Jiahao Shao, Gangwei Xu, Jiaming Zhou, et al. Native video-action pretraining for gener- alizable robot control.arXiv preprint arXiv:2607.08639, 2026a. Yuyang Zhang, Wenyao Zhang,... on PithPith's reviewarXiv
  17. [42] doi: 10.52202/075280-1506. Weiheng Zhao, Haoyi Jiang, Xin Shi, Liu Liu, Fan Huang, Zhizhong Su, Wei Sui, and Xinggang Wang. Faster-W AM: Efficient Inference-Time Future Conditioning for Robust World Action Models.arXiv preprint arXiv:2608.04404, 2026a. Weisen Zhao, Lam Nguyen,... on PithPith's reviewarXiv
  18. [43] Jiaming Zhou, Qihang Zhang, Gangwei Xu, Cunxin Fan, Yujie Zhao, Ruilin Wang, Yiming Luo, Shuai Yang, Xing Zhu, Yujun Shen, Junwei Liang, and Yinghao Xu. Zero-W AM: In-context world-action modeling from human videos for open-ended task generalization.arXiv preprint arXiv:2608.26103, on PithPith's reviewarXiv
  19. [44] Siyuan Zhou, Yilun Du, Jiaben Chen, Yandong Li, Dit-Yan Yeung, and Chuang Gan. RoboDreamer: Learning compositional world models for robot imagination.arXiv preprint arXiv:2404.12377, on PithPith's reviewarXiv
  20. [45] Chuning Zhu, Raymond Yu, Siyuan Feng, Benjamin Burchfiel, Paarth Shah, and Abhishek Gupta. Unified world models: Coupling video and action diffusion for pretraining on large robotic datasets.arXiv preprint arXiv:2504.02792, on PithPith's reviewarXiv
  21. [46] Chang Zou, Xuyang Liu, Ting Liu, Siteng Huang, and Linfeng Zhang. Accelerating Diffusion Transformers with Token-wise Feature Caching.arXiv preprint arXiv:2410.05317, on PithPith's reviewarXiv
  22. [2022] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choro- manski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, Pete Florence, Chuyuan Fu, Montse Gonzalez Arenas, Keerthana Gopalakrishnan, Kehang Han, Karol Hausman, Alexander Herzog, Ja... on PithPith's reviewarXiv
  23. [2023] Jun Cen, Chaohui Yu, Hangjie Yuan, Yuming Jiang, Siteng Huang, Jiayan Guo, Xin Li, Yibing Song, Hao Luo, Fan Wang, Deli Zhao, and Hao Chen. WorldVLA: Towards autoregressive action world model.arXiv preprint arXiv:2506.21539, on PithPith's reviewarXiv
  24. [2024] Hongzhe Bi, Hengkai Tan, Shenghao Xie, Zeyuan Wang, Shuhe Huang, Haitian Liu, Ruowen Zhao, Yao Feng, Chendong Xiang, Yinze Rong, Hongyan Zhao, Hanyu Liu, Zhizhong Su, Lei Ma, Hang Su, and Jun Zhu. Motus: A unified latent action world model.arXiv preprint arXiv:2512.13030, on PithPith's reviewarXiv
  25. [2025] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al.π 0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164, on PithPith's reviewarXiv
  26. [2026] Homanga Bharadhwaj, Debidatta Dwibedi, Abhinav Gupta, Shubham Tulsiani, Carl Doersch, Ted Xiao, Dhruv Shah, Fei Xia, Dorsa Sadigh, and Sean Kirmani. Gen2Act: Human video generation in novel scenarios enables generalizable robot manipulation.arXiv preprint arXiv:2409.16283, on PithPith's reviewarXiv

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment