Robotics · arXiv:2610.21389 v1 · submitted 15 Oct 2026 · 25 references
Fast-VLA: You Only Think Once for Robot Manipulation
The paper says
“Vision-language-action (VLA) policies commonly place a multi-billion-parameter model inside the feedback loop, repeatedly recomputing task context whenever they predict a new action chunk. Yet task understanding changes far more slowly than the robot state.”
more of the abstract
We introduce Fast-VLA, which invokes a frozen 4B VLM once per episode to contextualize the instruction and initial views as fixed, multi-layer key-value task memory. During execution, a lightweight online actor fuses learned time- and arm-indexed queries with current visual and proprioceptive feedback to form action queries. These queries retrieve task context from the fixed memory and predict action chunks without re-running the foundation model. Fast-VLA achieves 92.46% and 92.40% success on RoboTwin 2.0 Clean and Randomized, and 95.95% and 79.11% on LIBERO and LIBERO-Plus evaluation. Against three continually conditioned baselines on the same hardware, it reduces RoboTwin amortized policy inference time by 10.9-17.1x and requires 1.22 GiB of post-prefill online resident memory. Task-memory swaps further show that the actor relies on episode-matched task memory. These results show that task context can be established once while closed-loop control remains responsive throughout execution. Project page: https://garfield-kh.github.io/FastVLA
from the abstract
Pith says
Machine review · deepseek-v4-flash · 16 Oct 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
The passes disagreed. Reader: holds up. Skeptic: holds up only with conditions. Referee: major revision.
Fast-VLA is a well-engineered architecture that cleanly separates one-time task encoding from online control, backed by thorough benchmarks and convincing efficiency gains; its main unvalidated assumption—that all task-relevant objects are visible at episode start—deserves explicit discussion but doesn't undermine the core contribution. This line is wrong
1 major objection · 5 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 25 references say.
The paper says, at §3.2, §5, Table 4
The review did not quote this passage.
Pith says: Objection 1 of 1 · major
The paper relies on the unvalidated assumption that all task-relevant objects are visible in the initial multi-camera views. The task memory is built once from the initial frames; objects that are initially occluded and appear later (e.g., inside a drawer that opens during the episode) have no corresponding KV entries, so the cross-memory attention cannot retrieve them.
more of objection 1
The ablation in Table 4 confirms the criticality of initial visual content: swapping the initial-view segment collapses success to 2.64%. However, the benchmarks (RoboTwin, LIBERO) inherently place all required objects in the initial field of view; the paper does not evaluate tasks where objects become visible only after the memory is fixed. The Limitations section (E) only mentions dynamic task goals, not this occlusion issue. The authors should either evaluate on a benchmark that includes initially occluded objects, or explicitly qualify the central claim that 'task context can be established once' to acknowledge that the current design assumes full initial visibility of all relevant scene elements. This is a load-bearing limitation for generalization to many real-world scenarios.
5 minor objections
The paper says, at Table 8
Pith says: minor objection 1 of 5
The memory refresh ablation uses last-layer memory only, not the full 34-layer memory. While the conclusion that once-per-episode memory is the most efficient trade-off is plausible, the trend could differ with full-memory refreshes. Please clarify why last-layer memory was chosen for this ablation and whether the comparison has been verified with full memory.
The paper says, at Figure 2
Pith says: minor objection 2 of 5
The figure would benefit from a clearer visual distinction between data paths that execute once per episode (VLM prefill, memory construction) and those that execute every action chunk (online actor). Consider using a distinct line style or color for the once-per-episode arrow.
show 3 more
The paper says, at §3.3, Query fusion
Pith says: minor objection 3 of 5
The description of the gated residual MLP in query fusion does not specify the number of layers or hidden dimension. Including these architectural details would aid reproducibility.
The paper says, at Appendix A, Table 10
Pith says: minor objection 4 of 5
The DINOv2 learning-rate multiplier is listed as 0.1 for both RoboTwin and LIBERO. Please confirm that this multiplier applies uniformly to all visual encoders, including those for wrist cameras, and specify whether the encoders are fine-tuned from the same initial checkpoint.
The paper says, at §4.1, amortized ms/step definition
Pith says: minor objection 5 of 5
The amortized latency metric is clearly defined in §3.4 and Table 2 footnotes, but the main text (e.g., §4.1) could briefly restate that this metric includes the once-per-episode VLM prefill cost averaged over average episode length. This would aid readers who skip the footnotes.
What would settle it
Take a LIBERO task where a key object starts in a closed drawer (e.g., 'put the black bowl in the bottom drawer and close it'). Prepare two initial observations: (A) drawer open, bowl visible; (B) drawer closed, bowl occluded. Use the same actor weights (trained on standard LIBERO). Construct task memory from each initial observation and run Fast-VLA for 200 episodes per condition. Compare success rates. If condition (B) success falls below ~70% (the no-memory baseline from Table 5), the fixed-memory assumption is load-bearing.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 16 Oct 2026
- Reader holds up
- Skeptic partly agrees: says it holds up only with conditions Fixed task memory from initial views assumes all relevant objects are visible at episode start; paper does not test tasks where occluded objects appear later, leaving a key assumption unvalidated.
- Editor wrote the verdict above
- Referee major revision: 1 major objection, 5 minor
Not run on this paper: 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
Fast-VLA does something genuinely useful: it decouples the expensive VLM prefill from the closed-loop control loop while keeping task context explicit and retrievable. The architecture is clean—frozen VLM builds layer-wise KV memory once, lightweight actor fuses learned queries with live feedback to read that memory—and the results are solid. On RoboTwin 2.0 they match DM0.5 within 1 percentage point while cutting amortized inference time by 11x, and the memory-swap ablations convincingly show the actor depends on episode-matched content, not spurious correlations.
The engineering is thorough: same-hardware profiling, multi-benchmark evaluation, and careful ablations (memory construction, online inputs, refresh frequency). The thanks to the (non-reviewed) DM0.5 foundation model is fine—it's a frozen component, and the paper's claim is about the architecture around it.
The soft spot the stress test flags is real: Fast-VLA builds task memory from initial views and never updates it, which assumes all relevant objects are visible at episode start. The authors mention dynamic task goals as unexplored but don't address the static-occlusion case. This is a load-bearing assumption for any deployment where objects appear later (e.g., opening a drawer to reveal a mug). The benchmarks happen to place everything in the initial field of view, so the paper doesn't test this at all. It's a limitation that should be stated prominently and ideally probed with a targeted experiment.
That said, the core result—that one-time task encoding can support competitive closed-loop control—holds for the tested regimes. The paper is honest about its constraints, and the efficiency gains are large enough to matter for real-world deployment.
This paper deserves a serious referee. I'd recommend requesting an explicit limitation statement about the static-occlusion assumption and, if feasible, a small experiment on a task where a required object is initially hidden. No need to hold up acceptance over it; it's a future-work item, not a fatal flaw.
The strongest objection
Fast-VLA constructs layer-wise KV task memory once from the instruction and initial multi-view images. The memory is fixed for the episode. While the online actor observes current views for state estimation, the task memory provides semantic context that the cross-memory attention retrieves. An object not visible in any initial camera frame (e.g., inside a drawer that opens later) has no corresponding keys or values in the memory. The paper's own ablation (Table 4) shows that swapping the initial-view segment collapses success to 2.64%, proving the initial visual content is critical. Yet the benchmarks (RoboTwin, LIBERO) inherently place all required objects in the initial field of view; they do not force a situation where an object is initially occluded and becomes visible only during execution. The paper does not evaluate this scenario or acknowledge it as a limitation (Section E only mentions dynamic task goals). If the actor cannot retrieve context for objects that appear later, the central claim that 'task context can be established once' would not generalize to many real-world tasks where objects are initially hidden. This is a load-bearing unvalidated assumption.
The reader's weakest assumption correctly flagged the fixed-memory sufficiency, but focused on dynamic task goals (e.g., mid-episode instruction changes). The paper's limitation section only discusses that case. However, a distinct, unacknowledged assumption is that all task-relevant objects are visible in the initial views. The central claim's generality is weakened if tasks with initially occluded objects are common. The paper's ablations confirm that the initial-view segment is crucial for success; yet the benchmarks never require the actor to handle an object that appears only after the initial observation. A simple controlled experiment (e.g., occluding an object behind a closed drawer) would directly test whether the fixed memory can still guide action when the scene's semantic content evolves. If it fails, then 'think once' is insufficient for a broad class of real manipulation tasks. This does not invalidate the paper's contributions, but it qualifies the scope. Therefore, I recommend a CONDITIONAL verdict: accept on condition that the authors either provide such an evaluation or explicitly delimit the assumption in the paper.
The referee's report
If the approach generalizes, it offers a compelling efficiency-accuracy trade-off for deploying VLA policies in latency-sensitive robotics applications. The empirical evaluation is extensive: same-hardware latency profiling across three benchmarks with over 10,000 rollouts, thorough ablations covering memory construction, refresh frequency, encoder choice, and input modalities, and controlled memory interventions that verify the actor relies on episode-matched context. The work addresses a practical bottleneck in existing VLA methods by moving the large backbone outside the repeated control loop without sacrificing closed-loop responsiveness on the tested tasks.
Pith's reading of the paper (a machine's summary, not the authors')
Vision-language-action (VLA) policies typically run a large vision-language model at every control step, re-encoding the instruction and images each time. Fast-VLA argues this is wasteful because the task understanding changes far more slowly than the robot state. The paper introduces an architecture that invokes a frozen 4B VLM once per episode to build fixed, layer-wise key-value task memory from the instruction and initial views. During execution, a lightweight online actor fuses learned queries with current visual and proprioceptive feedback to retrieve task context from this fixed memory and predict action chunks. This separation yields competitive success rates on RoboTwin 2.0, LIBERO, and LIBERO-Plus while reducing amortized inference time by 10.9–17.1× and requiring only 1.22 GiB of online memory.
What carries the argument. The key machinery is the fixed layer-wise key-value task memory, built once per episode by a frozen 4B vision-language model (DM0.5). This memory stores the encoded instruction and initial multi-view frames across all 34 Transformer layers and remains unchanged during execution. During online control, the actor uses learnable time- and arm-indexed queries, fused with current arm state and wrist-view CLS tokens via a gated residual MLP, to form action queries. These action queries retrieve task context from the fixed memory through cross-attention at multiple decoder layers, enabling the actor to predict a 50-step action chunk in one forward pass without iterative denoising.
Core claim. Fast-VLA demonstrates that one-time task context encoding suffices for closed-loop robot manipulation. The core discovery is that by constructing fixed key-value memory from the instruction and initial multi-camera views using a frozen VLM at episode start, and then having a lightweight online actor query this memory with current observations, a VLA policy can maintain competitive performance while drastically reducing computational cost. The actor's queries incorporate learned time- and arm-indexed embeddings fused with proprioception and wrist-view features, allowing it to read relevant task context from all VLM layers. Controlled memory swaps show that the actor is indeed using the episode-matched memory.
The premise it rests on. The paper assumes that task understanding encoded in the initial instruction and views remains sufficient for the entire episode, i.e., task goals do not change mid-episode.
If the paper is right (Pith's inferences, not the authors')
Fast-VLA shows that large VLMs can be moved outside the online control loop, enabling resource-constrained robots to benefit from rich semantic understanding without high per-step cost.
The fixed memory design allows precomputing task memory offline, reducing online computation to the lightweight actor only.
The approach is compatible with different frozen VLMs as memory producers (PaliGemma, π0.5, Qwen2.5-VL, DM0.5), offering a trade-off between memory footprint and robustness to perturbations.
Because task memory is fixed, the method is limited to episodes with static task goals, as the authors acknowledge.
The actor's heavy reliance on wrist-view visual feedback (Table 6) identifies wrist cameras as a critical sensor modality for closed-loop control in this architecture.
If task understanding can be encoded once, the initial reasoning for task planning (e.g., object localization, subgoal decomposition) might also be amortized, potentially allowing even larger models for initial reasoning without runtime penalty.
The sharp performance drop under memory swaps (from 92% to 0.5–35%) suggests the actor is tightly coupled to the exact initial observation; this may limit robustness to initial perception errors or significant scene changes during an episode.
The method's low memory footprint (1.22 GiB for RoboTwin) and latency (1.44 ms/step) open the possibility of deploying high-capacity VLM knowledge on low-cost embedded hardware that can afford one VLM prefill.
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
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.
The paper
Abstract
Vision-language-action (VLA) policies commonly place a multi-billion-parameter model inside the feedback loop, repeatedly recomputing task context whenever they predict a new action chunk. Yet task understanding changes far more slowly than the robot state. We introduce Fast-VLA, which invokes a frozen 4B VLM once per episode to contextualize the instruction and initial views as fixed, multi-layer key-value task memory. During execution, a lightweight online actor fuses learned time- and arm-indexed queries with current visual and proprioceptive feedback to form action queries. These queries retrieve task context from the fixed memory and predict action chunks without re-running the foundation model. Fast-VLA achieves 92.46% and 92.40% success on RoboTwin 2.0 Clean and Randomized, and 95.95% and 79.11% on LIBERO and LIBERO-Plus evaluation. Against three continually conditioned baselines on the same hardware, it reduces RoboTwin amortized policy inference time by 10.9-17.1x and requires 1.22 GiB of post-prefill online resident memory. Task-memory swaps further show that the actor relies on episode-matched task memory. These results show that task context can be established once while closed-loop control remains responsive throughout execution. Project page: https://garfield-kh.github.io/FastVLA
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 · 2026-10-16
The authors' main claim, as the classifier picked it out 1.00
“These results show that task context can be established once while closed-loop control remains responsive throughout execution.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.07 · states limitation 0.12 · human participants 0.03 · controlled comparison 0.97 · randomized assignment 0.07 · states quantitative result 0.99 · compares with prior results 0.96 · mentions code or data availability 0.38
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.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Vision-language-action (VLA) policies commonly place a multi-billion-parameter model inside the feedback loop, repeatedly recomputing task context whenever they predict a new action chunk. | background 1.00 | 0.08 | 0.03 | asserted 1.00 |
| Yet task understanding changes far more slowly than the robot state. | background 0.78 | 0.18 | 0.07 | asserted 1.00 |
| We introduce Fast-VLA, which invokes a frozen 4B VLM once per episode to contextualize the instruction and initial views as fixed, multi-layer key-value task memory. | contribution 0.62 | 0.93 | 0.12 | asserted 1.00 |
| During execution, a lightweight online actor fuses learned time- and arm-indexed queries with current visual and proprioceptive feedback to form action queries. | method 0.99 | 0.86 | 0.04 | asserted 0.99 |
| These queries retrieve task context from the fixed memory and predict action chunks without re-running the foundation model. | method 0.99 | 0.82 | 0.16 | asserted 0.99 |
| Fast-VLA achieves 92.46% and 92.40% success on RoboTwin 2.0 Clean and Randomized, and 95.95% and 79.11% on LIBERO and LIBERO-Plus evaluation. | result 1.00 | 0.96 | 0.10 | asserted 1.00 |
| Against three continually conditioned baselines on the same hardware, it reduces RoboTwin amortized policy inference time by 10.9-17.1x and requires 1.22 GiB of post-prefill online resident memory. | result 1.00 | 0.96 | 0.08 | asserted 1.00 |
| Task-memory swaps further show that the actor relies on episode-matched task memory. | result 0.85 | 0.94 | 0.05 | asserted 0.99 |
| These results show that task context can be established once while closed-loop control remains responsive throughout execution. | interpretation 1.00 | 0.87 | 0.88 | asserted 1.00 |
| Project page: https://garfield-kh.github.io/FastVLA | future work 0.51 | 0.15 | 0.03 | unclear 0.98 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.89
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, 2026-10-16: 1 request, 7,571 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 (6)
Figures are the authors', reproduced from the paper.
References (25)
- [1] Joshua Ainslie, James Lee-Thorp, Michiel de Jong, Yury Zemlyanskiy, Federico Lebr´on, and Sumit Sanghai. Gqa: Training generalized multi-query transformer models from multi-head checkpoints. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processi... on Pith
- [6] Qingwen Bu, Yanting Yang, Jisong Cai, Shenyuan Gao, Guanghui Ren, Maoqing Yao, Ping Luo, and Hongyang Li. Univla: Learning to act anywhere with task-centric latent actions.arXiv preprint arXiv:2505.06111, on PithPith's reviewarXiv
- [7] Jisong Cai, Long Ling, Shiwei Chu, Zhongshan Liu, Jiayue Kang, Zhixuan Liang, Wenjie Xu, Yinan Mao, Weinan Zhang, Xiaokang Yang, Ru Ying, Ran Zheng, and Yao Mu. Aha-wam: Asynchronous horizon-adaptive world-action modeling with observation-guided context routing. arXiv preprint... on PithPith's reviewarXiv
- [8] 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 ar... on PithPith's reviewarXiv
- [9] URLhttps://www.dexmal.com/blog/dm0.5/index_en.html. 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: In-depth robustness analysis of vision-language-action... on PithPith's reviewarXiv
- [10] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla... on PithPith's reviewarXiv
- [11] Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-tuning vision-language-action models: Optimiz- ing speed and success.arXiv preprint arXiv:2502.19645, on PithPith's reviewarXiv
- [12] Lin Li, Qihang Zhang, Yiming Luo, Shuai Yang, Ruilin Wang, Fei Han, Mingrui Yu, Zelin Gao, Nan Xue, Xing Zhu, Yujun Shen, and Yinghao Xu. Causal world modeling for robot control.arXiv preprint arXiv:2601.21998, on PithPith's reviewarXiv
- [13] 11 Preprint 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
- [14] Liheng Ma, Rui Heng Yang, Zhanguang Zhang, Mateo Clemente, Ziwen Hu, Tongtong Cao, and Yingxue Zhang. Faster-wam: Do world action models need deep action modules?arXiv preprint arXiv:2608.02365, on PithPith's reviewarXiv
- [15] Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy Vo, Marc Szafraniec, et al. Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, on PithPith's reviewarXiv
- [16] Karl Pertsch, Kyle Stachowicz, Brian Ichter, Danny Driess, Suraj Nair, Quan Vuong, Oier Mees, Chelsea Finn, and Sergey Levine. Fast: Efficient action tokenization for vision-language-action models.arXiv preprint arXiv:2501.09747, on PithPith's reviewarXiv
- [17] Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Manuel Y. Galliker, Dibya Ghosh, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Devin Le... on PithPith's reviewarXiv
- [18] StarVLA Community. Starvla: A lego-like codebase for vision-language-action model developing. arXiv preprint arXiv:2604.05014, on PithPith's reviewarXiv
- [19] Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding.arXiv preprint arXiv:2104.09864, on PithPith's reviewarXiv
- [21] Wei Wu, Fan Lu, Yunnan Wang, Shuai Yang, Shi Liu, Fangjing Wang, Shuailei Ma, He Sun, Yong Wang, Zhenqi Qiu, Houlong Xiong, Ziyu Wang, Shuai Zhou, Yiyu Ren, Kejia Zhang, Hui Yu, Jingmei Zhao, Qian Zhu, Ran Cheng, Yong-Lu Li, Yongtao Huang, Xing Zhu, Yujun Shen, and Kecheng Zhe... on PithPith's reviewarXiv
- [22] Tianyuan Yuan, Zibin Dong, Yicheng Liu, and Hang Zhao. Fast-wam: Do world action models need test-time future imagination?arXiv preprint arXiv:2603.16666, on PithPith's reviewarXiv
- [23] Weiheng Zhao, Haoyi Jiang, Xin Shi, Liu Liu, Fan Huang, Zhizhong Su, Wei Sui, and Xinggang Wang. Faster-wam: Efficient inference-time future conditioning for robust world action models. arXiv preprint arXiv:2608.04404, on PithPith's reviewarXiv
- [24] Jinliang Zheng, Jianxiong Li, Zhihao Wang, Dongxiu Liu, Xirui Kang, Yuchun Feng, Yinan Zheng, Jiayin Zou, Yilun Chen, Jia Zeng, et al. X-vla: Soft-prompted transformer as scalable cross- embodiment vision-language-action model.arXiv preprint arXiv:2510.10274, on PithPith's reviewarXiv
- [25] Warmup and timing 3 warmup calls followed by 10 timed calls; report mean latency. Device synchronization Synchronized immediately before and after each timed call. Preprocessing and trans- fer Timing starts from raw uint8 HWC NumPy images and includes resizing, nor- malization... on Pith
show 5 more
- [2021] Qiuyue Wang, Mingsheng Li, Jian Guan, Jinhui Ye, Sicheng Xie, Yitao Liu, Junhao Chen, Zhixuan Liang, Jie Zhang, Xintong Hu, Xuhong Huang, Pei Lin, et al. Qwen-vla: Unifying vision- language-action modeling across tasks, environments, and robot embodiments.arXiv preprint arXiv:... on PithPith's reviewarXiv
- [2023] Arman Akbari, Ci Zhang, Arash Akbari, Lin Zhao, Yixiao Chen, Weiwei Chen, Xuan Zhang, Geng Yuan, and Yanzhi Wang. Flash-wam: Modality-aware distillation for world action models.arXiv preprint arXiv:2606.05254, on PithPith's reviewarXiv
- [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
- [2025] Lucas Beyer, Andreas Steiner, Andr´e Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer.arXiv preprint arXiv:2407.07726, on PithPith's reviewarXiv
- [2026] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, et al. Qwen2.5-vl technical report.arXiv preprint arXiv:2502.13923, 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.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment