Robotics · arXiv:2609.39973 v1 · submitted 30 Sep 2026 · 79 references
EWAM: Emergent Depth-Wise Specialization in a Unified Embodied Model – From Semantic Understanding through Visual Foresight to Action
The paper says
“Vision-language-action (VLA) policies emphasize semantic understanding, whereas world-action models (WAMs) learn predictive representations of environment dynamics. Systems that expose a policy to both sources often still concentrate action computation on a single expert.”
more of the abstract
We present EWAM, an action-centric unified embodied model whose asymmetric joint attention lets action tokens read semantic, current-visual, predicted-future, and action information at every layer while the perceptual experts retain their distinct roles. Without layer-wise supervision, EWAM develops an emergent depth-wise specialization: action queries attend mainly to vision-language features in shallow layers, to predicted future frames in intermediate layers, and to action tokens themselves in deep layers. This handoff replicates across tasks and is stable across denoising steps. Checkpoint tracking and causal interventions show that it is learned and that action generation depends on it. EWAM is pretrained in two separate regimes, one on cross-embodiment robot trajectories and one on human egocentric video. In simulation and real-robot experiments, it surpasses existing VLA, WAM, and hybrid baselines. Human egocentric data improve both cross-embodiment transfer and real-robot robustness, and subtask-phase supervision improves long-horizon completion. Together, these results suggest that unified embodied learning can induce an ordered internal progression from semantic understanding, through visual foresight, to action formation.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 1 Oct 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. A solid architecture plus a well-measured emergent routing phenomenon, with the strongest evidence supporting depth-wise functional dependence; the 'ordered understand-foresee-act' pipeline claim in the title and abstract outruns what the causal tests can establish. This line is wrong
3 major objections · 5 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 79 references say.
The paper says, at Abstract, §5.4, §8
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.94):
“Attention allocation identifies where action queries retrieve information, and the interventions establish functional dependence on selected sources, but neither provides a complete causal decomposition of the computation.”
Pith says: Objection 1 of 3 · major
The abstract's headline — an 'ordered internal progression from semantic understanding, through visual foresight, to action formation' — outruns the evidence. The fixed-checkpoint interventions (Table 1, right panel) show depth- and source-specific functional dependence: masking future keys in layers 10-19 drops success from 92.80% to 3.55%, masking VL keys in 0-9 costs 16.27 points, and masking all video in 20-29 costs 26.98 points.
more of objection 1
These tests establish that action generation depends concurrently on VL, future, and action sources at three depth ranges, but they do not establish that the intermediate future representations are computed from, or gated by, the shallow semantic stage, nor that deep action formation is downstream of the futures in a pipeline sense. Because the architecture executes all layers in a single forward pass, the observed handoff is equally consistent with three parallel, depth-specialized modules. The authors concede this in Section 8 ('neither provides a complete causal decomposition of the computation') and Appendix D ('We therefore restrict the claim to routing'), and they defer 'cross-layer causal mediation' to future work. Given that concession, the abstract and conclusion (b4, b10, b370) should either be reframed to claim depth-wise specialization without causal ordering, or the missing mediation evidence must be added.
The paper says, at §7.5, §7.6, Tables 7-9, 11
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.79):
“Differences between these conditions amount to one trial, so their ordering should not be overinterpreted.”
Pith says: Objection 2 of 3 · major
The real-world performance claims rest on 10 physical trials per task or per condition, with no uncertainty quantification or significance testing. In Table 9, the in-domain baseline is 8/10 and the held-out conditions are 7/10, 7/10, and 6/10; the authors themselves note that 'differences between these conditions amount to one trial' (b307), yet Section 7.5's conclusion that EWAM 'attains the highest success rate on every task' is asserted as a point comparison.
more of objection 2
Under a binomial model, a 10-point gap over 10 trials (e.g., Franka Stack Bowls, 80% vs 70%) carries a wide confidence interval and does not support a superiority claim. Please report exact binomial confidence intervals (or add trials) for all real-robot point comparisons, and temper the 'highest on every task' phrasing accordingly.
The paper says, at §7.2-7.4, Tables 4-6
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.94):
“Because pretraining data differ across methods, this comparison alone does not isolate the contribution of the attention architecture. 7.3 RoboTwin In-Domain Results We additionally report the in-domain setting, in which every method, including ours, is trained on both clean and randomized demonstrations and evaluated under each condition.”
Pith says: Objection 3 of 3 · major
The headline performance comparisons are confounded. For the clean-to-random protocol (Table 4) the authors acknowledge that 'pretraining data differ across methods... this comparison alone does not isolate the contribution of the attention architecture' (b277), but the abstract's unqualified 'surpasses existing VLA, WAM, and hybrid baselines' inherits that confound.
more of objection 3
The LIBERO results (Table 6) are reported with a note that baselines come 'from published evaluations,' meaning different training pipelines and data. The only setting that holds data conditions constant across methods is the RoboTwin in-domain evaluation (Table 5). Please restrict the summary-level performance claims to the matched in-domain condition (or add controlled pretraining-matched comparisons), and state explicitly in the abstract that the C2R and LIBERO margins reflect unmatched pretraining.
5 minor objections
The paper says, at §5.3, Table 13
Pith says: minor objection 1 of 5
The checkpoint-evolution analysis is measured on one representative sequence from a single thirty-two-GPU run; the authors acknowledge this ('The trajectory is measured on one representative sequence from one run'). The claim that the handoff 'emerges during training' would be considerably stronger with a second seed, but as stated this is a presentation limitation rather than a blocking one.
The paper says, at Appendix B.1
Pith says: minor objection 2 of 5
The stage boundaries (0-9, 10-19, 20-29) were chosen after observing the attention transitions and are described as 'descriptive partitions rather than architecturally hard-coded modules.' Because the paper's Section 5.2 stage-level numbers inherit this post-hoc choice, please add a sensitivity check (e.g., reporting key allocations with boundaries shifted by ±2 layers) so readers can gauge how much the three-regime narrative depends on the chosen cut points.
show 3 more
The paper says, at Figure 1, Appendix A
Pith says: minor objection 3 of 5
The cross-paradigm attention maps are measured on a single RoboTwin ADJUSTBOTTLE episode at denoising step 0. The appendix states this and cites stability results, but the figure and its caption should flag 'single episode, denoising step 0' for readers who do not consult the appendix.
The paper says, at Table 8
Pith says: minor objection 4 of 5
Table 8's formatting appears broken: the π0.5 row reads '30 204060 60', which loses the column structure (e.g., the Pour Beans value is missing or merged). Please fix the table alignment so each model has seven task columns.
The paper says, at Throughout
Pith says: minor objection 5 of 5
The model name is rendered inconsistently as 'EWAM' and 'EW AM' (often with an errant space before 'AM'). Please harmonize to one spelling.
What would settle it
On the fixed robot-pretrained checkpoint, perturb shallow VL representations at inference by masking VL keys in layers 0-9 (or injecting counterfactual VL embeddings), then measure (a) the change in intermediate future-frame token representations at layers 10-19 (e.g., cosine similarity or reconstruction error of future latents) and (b) downstream task success. If future-frame embeddings are unchanged while success still drops, the shallow VL and intermediate future dependencies are parallel, not sequential, and the 'ordered progression' wording should be softened to 'depth-specialized routing'. If future embeddings change substantially, the ordering claim gains direct support.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 1 Oct 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 3 major objections, 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
Worth your time. The asymmetric action-centric mask is a real design choice, and the depth-wise attention handoff is genuinely observed: action queries read VL in shallow layers, future frames in the middle, and themselves deep, replicating across 50 tasks with tight bootstrap CIs and surviving token-count normalization. The checkpoint tracking shows the organization emerges rather than being present at init, and the causal tests are above the usual bar. Fixed-checkpoint masking is the right way to measure learned dependence, and the counterfactual future-feature injection (25/25 to 4/25, with interpretable arm-selection flips on ADJUSTBOTTLE) is good content-sensitivity evidence. The bidirectional-mask control in Appendix D is a genuinely thoughtful control: it shows the coarse depth organization is not an artifact of the mask, while the fine-grained image-release handoff and the VL stream's identity as a pure semantic source are. The paper is also honest in Section 8 and Appendix D that the analysis characterizes routing, not representation, and that a full causal decomposition is future work.
Now the soft spots, in proportion. The abstract and title claim an "ordered internal progression from semantic understanding through visual foresight to action formation." The evidence supports depth-specialized routing with coarse functional dependence — masking middle future keys collapses success to 3.55%, shallow VL masking costs 16 points — but nothing establishes a causal sequence between the stages. The model runs all layers in one forward pass; shallow VL masking does not show the middle future representations are computed from the semantic stage, and middle masking does not show deep action formation is downstream of futures in a pipeline sense. The intervention results are convincing as "these sources matter at these depths," not as "understanding precedes foresight precedes action." That is a framing overclaim contained in the title, not an incoherent paper; the fix is to reword the headline claims. Second, the benchmark wins run on unmatched pretraining data, which the authors concede ("this comparison alone does not isolate the contribution of the attention architecture"), so the performance story is suggestive rather than isolating. Third, the real-robot numbers rest on 10 trials per condition with no error bars — the "10% to 60%" robustness gain is vivid but thin. Fourth, no code or checkpoints are released, which matters more here than for many papers because the central phenomenon lives in the model's internals; nobody can verify or build on the routing analysis without weights.
Who gets value: embodied-AI researchers working on model internals and attention-based interpretability, and anyone designing unified VLA/WAM architectures. The causal-intervention methodology around attention routing is a contribution in itself. It deserves a serious referee; I'd send it out and push for aligned abstract claims and artifact release.
The strongest objection
The paper's central claim is an 'ordered internal progression' from semantic understanding through visual foresight to action formation (Abstract, b10). The observational evidence (Eq. 12 attention allocation, Figures 3-4) and the intervention tests (Table 1) establish that action generation depends concurrently on VL keys in layers 0-9 (masking costs 16.27 points), future-frame keys in layers 10-19 (masking costs 89.25 points), and video/action keys in deeper layers (masking all video in 20-29 costs 26.98 points). These are depth- and source-specific dependencies, and the counterfactual injection (Table 2) shows content sensitivity. However, none of these tests establishes a causal sequence between the stages. Masking shallow VL keys does not tell us whether intermediate future representations are computed from, or gated by, the shallow semantic stage; masking middle future keys does not tell us whether deep action formation is downstream of these futures in a pipeline sense. The architecture executes all layers in a single forward pass, so the observed attention handoff could equally reflect three parallel, depth-specialized modules rather than an ordered computational pipeline. The paper itself concedes in Section 8 and Appendix D that the analysis characterizes routing, not representation, and calls for 'cross-layer causal mediation' as future work. Given that the abstract's headline is the ordered progression, the current evidence supports depth-specialized routing with coarse functional dependence, but not the directionality implicit in 'understanding -> foresight -> action'.
The reader's weakest_assumption identified that attention mass is a routing-level proxy and that the interventions do not establish the ordering itself; my reading converges on this as the most load-bearing gap. The paper's own limitations (Section 8, Appendix D) acknowledge the absence of a full causal decomposition, and Appendix B.1 admits the depth partitions are post-hoc descriptive ranges. Since the empirical results themselves — cross-task attention replication, checkpoint tracking, fixed-checkpoint masking, and counterfactual injection — are internally consistent and reasonably controlled, the concern is not about fabrication but about the strength of the interpretive claim. The verdict CONDITIONAL already reflects this: the ordered-progression claim is a condition that additional cross-layer mediation tests, or a wording revision, would settle. I recommend keeping the conditional verdict rather than moving to accept or reject, because the core architectural novelty and the supporting evidence are substantial, while the 'ordered progression' phrasing exceeds what the current tests demonstrate.
The referee's report
If the depth-wise specialization result holds, it is a genuinely interesting phenomenon: there are few documented cases of an emergent layer-wise division of labor in a unified policy trained without layer-wise supervision. The paper's measurement methodology is a real strength that should be credited: task-bootstrap 95% confidence intervals are reported for attention allocations (Table 14), the analysis controls for token-count differences via enrichment statistics (Eq. 20), the handoff replicates across all 50 RoboTwin tasks with a tight crossover at median layer 13, the bidirectional-mask ablation (Appendix D) isolates the role of the asymmetric mask, and the fixed-checkpoint masking plus counterfactual injection is a well-designed causal test of depth- and source-specific dependence. The observation that masking future-frame keys in layers 10-19 collapses success from 92.80% to 3.55% while masking VL keys in 0-9 costs 16.27 points is a striking, falsifiable functional signature. The egocentric-pretraining and subtask-phase results are of secondary significance but are competently executed.
Pith's reading of the paper (a machine's summary, not the authors')
The paper claims that a single unified robot model, exposed simultaneously to semantic, visual-prediction, and action computation, spontaneously organizes itself into an ordered depth-wise pipeline: action queries draw on vision-language understanding in shallow layers, on predicted future frames in intermediate layers, and on their own action context in deep layers. This 'understand–foresee–act' handoff is claimed to emerge without any layer-wise supervision, to replicate across 50 tasks, and to be functionally load-bearing — removing predicted-future features in the intermediate layers drops success from 92.80% to 3.55%, and injecting deliberately wrong future features turns 25/25 successful rollouts into 4/25. If this is right, it shows a concrete architectural mechanism by which a unified policy avoids the single-expert bias the paper documents in VLA, world-action, and hybrid baselines, and it suggests that the division of labor between grounding, prediction, and control can be learned rather than prescribed.
What carries the argument. The load-bearing mechanism is the asymmetric joint attention mask, $A\leftarrow\{V,A,L\}$, $V\leftarrow V$, $L\leftarrow L$ (Eq. 3), enforced through a single concatenated masked-attention operation with per-stream normalization and projection parameters. Because action queries are the only cross-source readers, their post-softmax attention allocation (Eq. 12, $R^m_l$: the fraction of action-query attention mass falling on source $m$ at block $l$) becomes a well-defined measurement of which information source the policy retrieves at each depth. This is what turns the architectural claim into a testable quantity: the depth profile of $R^m_l$ across five sources (VL-text, VL-image, condition frame, future frame, action) is the evidence for the three-stage handoff, and it is the target of the causal interventions (layer-targeted masking, counterfactual future-feature injection). The joint video–action flow-matching objective (Eq. 5) supplies the training pressure that lets the organization emerge rather than being hand-designed.
Core claim. On the paper's own terms, the discovery is that an asymmetric joint-attention design — in which action tokens are the only queries allowed to read vision-language, video, and action keys, while the two perceptual experts attend only to themselves — induces an emergent depth-wise specialization in a 30-layer unified transformer. Measured by action-query attention mass, the dominant information source changes with depth: vision-language tokens (text plus current image) receive 75.4% of shallow action attention; the intermediate stage distributes attention across all three sources, with predicted future frames overtaking the current image at a median crossover layer 13; and deep layers raise action self-attention to 59.5%. Checkpoint tracking shows the organization is absent at 50 training steps and consolidates by 40,000. Layer-targeted interventions show action generation depends on it: masking future features in layers 10–19 collapses a fixed pretrained checkpoint from 92.80% to 3.55% success, while masking shallow video costs only 0.82 points. Same-task counterfactual injection of a donor rollout's future features degrades success from 25/25 to 4/25 across five tasks, including steering ADJUSTBOTTLE rollouts toward the donor's arm choice. The paper frames this as evidence that unified embodied learning can induce an ordered internal progression from semantic grounding, through visual foresight, to action formation.
The premise it rests on. Action-query attention mass (Eq. 12) is treated as a faithful proxy for the information action generation actually retrieves and uses at each depth, and the shallow/intermediate/deep boundaries (layers 0–9, 10–19, 20–29) are descriptive partitions chosen after observing the attention transitions; if attention allocation tracks statistical regularity rather than control-relevant information flow, the specific claim of an ordered understand–foresee–act handoff is not established, though the coarse source-dependence interventions would still stand.
If the paper is right (Pith's inferences, not the authors')
An action-centric asymmetric attention bridge is sufficient to induce depth-wise specialization without explicit layer-wise supervision; the ordered handoff emerges during training and stabilizes by 40,000 optimizer steps.
Predicted-future representations are functionally load-bearing for action generation specifically in the intermediate layers: masking them at layers 10–19 collapses success from 92.80% to 3.55%, far beyond the shallow (0.82-point) or deep (26.98-point) video-mask costs.
Human egocentric pretraining transfers to a held-out robot embodiment (40.1% to 66.9% cross-embodiment success), and egocentric co-training raises real-robot robustness under scene changes (10% to 60%), because both share the same camera-relative wrist-motion action space.
Subtask-phase supervision — a textual label of the active phase plus a normalized phase index — raises long-horizon completion (instructed ranking 83.5% to 91.0%, stacking 44.5% to 58.5%, dual-basket sorting 1.0 to 2.2 of 4 objects) without an external planner.
The same model competes well against VLA, WAM, and hybrid baselines on RoboTwin clean-to-random (72.1%), in-domain (92.9%), LIBERO (98.8%), and across all seven real-world tasks.
If the ordered handoff is a general property of action-centric attention rather than of this specific 30-layer stack, it should reappear at other scales; a testable extension is whether the semantic-to-foresight-to-action ordering and its crossover depth shift predictably with layer count, token ratios, and task horizon.
The measured specialization suggests a concrete optimization route the paper mentions only briefly: allocate compute over the expert–depth grid by functional contribution, pruning the near-dead deep VL-image pathway (1.3% attention in layers 16–29) instead of compressing all three experts uniformly.
The paper's own limitation statement leaves open whether the routing change is accompanied by representational change; until deep-layer activations are probed or patched directly, 'emergence' is established for attention allocation, not for the semantics that allocation carries.
The counterfactual injection results hint that future-frame representations encode control-relevant content (arm selection, spatial targeting) rather than serving only as an auxiliary prediction target, which strengthens the case for letting the action stream read intermediate predictions during denoising rather than consuming a separately generated subgoal.
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 emphasize semantic understanding, whereas world-action models (WAMs) learn predictive representations of environment dynamics. Systems that expose a policy to both sources often still concentrate action computation on a single expert. We present EWAM, an action-centric unified embodied model whose asymmetric joint attention lets action tokens read semantic, current-visual, predicted-future, and action information at every layer while the perceptual experts retain their distinct roles. Without layer-wise supervision, EWAM develops an emergent depth-wise specialization: action queries attend mainly to vision-language features in shallow layers, to predicted future frames in intermediate layers, and to action tokens themselves in deep layers. This handoff replicates across tasks and is stable across denoising steps. Checkpoint tracking and causal interventions show that it is learned and that action generation depends on it. EWAM is pretrained in two separate regimes, one on cross-embodiment robot trajectories and one on human egocentric video. In simulation and real-robot experiments, it surpasses existing VLA, WAM, and hybrid baselines. Human egocentric data improve both cross-embodiment transfer and real-robot robustness, and subtask-phase supervision improves long-horizon completion. Together, these results suggest that unified embodied learning can induce an ordered internal progression from semantic understanding, through visual foresight, to action formation.
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 · 2026-10-01
The authors' main claim, as the classifier picked it out 1.00
“Together, these results suggest that unified embodied learning can induce an ordered internal progression from semantic understanding, through visual foresight, to action formation.”
How the main claim is stated: hedged possibility 1.00
What the abstract itself states, as probabilities: claims first 0.07 · states limitation 0.13 · human participants 0.17 · controlled comparison 0.88 · randomized assignment 0.02 · states quantitative result 0.05 · compares with prior results 0.97 · 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.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Vision-language-action (VLA) policies emphasize semantic understanding, whereas world-action models (WAMs) learn predictive representations of environment dynamics. | background 1.00 | 0.05 | 0.02 | asserted 1.00 |
| Systems that expose a policy to both sources often still concentrate action computation on a single expert. | gap 0.82 | 0.08 | 0.02 | asserted 0.97 |
| We present EWAM, an action-centric unified embodied model whose asymmetric joint attention lets action tokens read semantic, current-visual, predicted-future, and action information at every layer while the perceptual experts retain their distinct roles. | contribution 0.65 | 0.89 | 0.10 | asserted 1.00 |
| Without layer-wise supervision, EWAM develops an emergent depth-wise specialization: action queries attend mainly to vision-language features in shallow layers, to predicted future frames in intermediate layers, and to action tokens themselves in deep layers. | result 0.98 | 0.95 | 0.32 | asserted 1.00 |
| This handoff replicates across tasks and is stable across denoising steps. | result 1.00 | 0.92 | 0.07 | asserted 1.00 |
| Checkpoint tracking and causal interventions show that it is learned and that action generation depends on it. | result 0.94 | 0.95 | 0.08 | asserted 1.00 |
| EWAM is pretrained in two separate regimes, one on cross-embodiment robot trajectories and one on human egocentric video. | method 1.00 | 0.93 | 0.02 | asserted 1.00 |
| In simulation and real-robot experiments, it surpasses existing VLA, WAM, and hybrid baselines. | result 1.00 | 0.95 | 0.11 | asserted 1.00 |
| Human egocentric data improve both cross-embodiment transfer and real-robot robustness, and subtask-phase supervision improves long-horizon completion. | result 1.00 | 0.94 | 0.07 | asserted 1.00 |
| Together, these results suggest that unified embodied learning can induce an ordered internal progression from semantic understanding, through visual foresight, to action formation. | interpretation 1.00 | 0.75 | 0.86 | hedged possibility 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.98
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, 2026-10-01: 1 request, 7,548 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 (19)
Figures are the authors', reproduced from the paper.
References (79)
- [1] Kenneth J. W. Craik.The Nature of Explanation. Cambridge University Press, 1943 on Pith
- [2] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, et al. RT-1: Robotics transformer for real-world control at scale. InRobotics: Science and Systems, 2023 on Pith
- [3] Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, et al. PaLM-E: An embodied multimodal language model. InInternational Conference on Machine Learning, 2023 on Pith
- [4] Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, et al. RT-2: Vision-language-action models transfer web knowledge to robotic control. InConference on Robot Learning, pages 2165–2183, 2023 on Pith
- [5] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan P. Foster, Pannag R. Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. OpenVLA: An ... on Pith
- [6] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al.π0: A vision-language-action flow model for general robot control. InRobotics: Science and Systems, 2025 on Pith
- [7] 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 L... on Pith
- [8] Seonghyeon Ye, Yunhao Ge, Kaiyuan Zheng, Shenyuan Gao, Sihyun Yu, George Kurian, Suneel Indupuru, You Liang Tan, Chuning Zhu, et al. World action models are zero-shot policies.arXiv preprint arXiv:2602.15922, 2026. URLhttps://arxiv.org/abs/2602.15922 on PithPith's reviewarXiv
- [9] 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, 2026. URLhttps://arxiv.org/abs/2601.21998 on PithPith's reviewarXiv
- [10] 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, 2026. URL https://arxiv.org/ abs/2603.16666 on PithPith's reviewarXiv
- [11] Zhanguang Zhang, Zhiyuan Li, Behnam Rahmati, Rui Heng Yang, Yintao Ma, Amir Rasouli, Sajjad Pakdamansavoji, Yangzheng Wu, Lingfeng Zhang, Tongtong Cao, Feng Wen, Xinyu Wang, Xingyue Quan, and Yingxue Zhang. Do world action models generalize better than VLAs? a robustness study... on PithPith's reviewarXiv
- [12] Wei Wu, Fangjing Wang, Fan Lu, He Sun, Shi Liu, Yunnan Wang, Yibin Yan, Yong Wang, Shuailei Ma, Xinyang Wang, Yibin Liu, Shuai Yang, Tianxiang Zhou, Kejia Zhang, Lei Zhou, Cheng Su, Nan Xue, Bin Tan, Han Zhang, Youchao Zhang, Fei Liao, Xing Zhu, Yujun Shen, and Kecheng Zheng. ... on PithPith's reviewarXiv
- [13] Jianke Zhang, Yucheng Hu, Yanjiang Guo, Xiaoyu Chen, Yichen Liu, Wenna Chen, Chaochao Lu, and Jianyu Chen. UniJEPA: Enhancing robot policy via unified continuous and discrete representation learning. InInternational Conference on Machine Learning, 2026 on Pith
- [14] Jingwen Sun, Wenyao Zhang, Zekun Qi, Shaojie Ren, Zezhi Liu, Hanxin Zhu, Guangzhong Sun, Xin Jin, and Zhibo Chen. VLA-JEPA: Enhancing vision-language-action model with latent world model. InEuro- pean Conference on Computer Vision (ECCV), pages 478–497, 2026. doi: 10.1007/978-... on Pithdoi:10.1007/978-3-032-37196-6
- [15] Physical Intelligence, Bo Ai, Ali Amin, Raichelle Aniceto, Ashwin Balakrishna, Greg Balke, Kevin Black, George Bokinsky, Shihao Cao, et al.π0.7: A steerable generalist robotic foundation model with emergent capabilities.arXiv preprint arXiv:2604.15483, 2026. URL https://arxiv.... on PithPith's reviewarXiv
- [16] Qian Long, Yueze Wang, Jiaxi Song, Junbo Zhang, Peiyan Li, Wenxuan Wang, Yuqi Wang, Haoyang Li, Shaoxuan Xie, Guocai Yao, Hanbo Zhang, Xinlong Wang, Zhongyuan Wang, Xuguang Lan, Huaping Liu, and Xinghang Li. Scaling world model for hierarchical manipulation policies.arXiv prep... on PitharXiv
- [17] 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. In IEEE/CVF Conference on Computer V... on Pith
- [18] Dibya Ghosh, Homer Rich Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, Jianlan Luo, You Liang Tan, Lawrence Yunliang Chen, Quan Vuong, Ted Xiao, Pannag R Sanketi, Dorsa Sadigh, Chelsea Finn, and Sergey Levine. Octo: An open-... on Pithdoi:10.15607/rss.2024.xx.090
- [19] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. InRobotics: Science and Systems, 2023 on Pith
- [20] Chaorui Deng, Deyao Zhu, Kunchang Li, Chenhui Gou, Feng Li, Zeyu Wang, Shu Zhong, Weihao Yu, Xiaonan Nie, Ziang Song, Guang Shi, and Haoqi Fan. Emerging properties in unified multimodal pretraining.arXiv preprint arXiv:2505.14683, 2025 on PithPith's reviewarXiv
show 59 more
- [21] Adriana Aida, Walid Amer, Katarina Bankovic, Dhruv Behl, Fabian Busch, Annie Bhalla, Minh Duong, Florian Gienger, Rohan Godse, Denis Grachev, Ralf Gulde, Elisa Hagensieker, Junpeng Hu, Shivam Joshi, Tobias Knobloch, Likith Kumar, Damien LaRocque, Keerthana Lokesh, Omar Moured,... on PithPith's reviewarXiv
- [22] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, et al. Ego4D: Around the world in 3,000 hours of egocentric video. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18995–19012, 2022 on Pith
- [23] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, et al. Scaling egocentric vision: The EPIC-KITCHENS dataset. InProceedings of the European Conference on Computer Vision (ECCV), pages 753–771, 2018 on Pith
- [24] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, et al. The “something something” video database for learning and evaluating visual common sense. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 5... on Pith
- [25] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3M: A universal visual representation for robot manipulation. InConference on Robot Learning (CoRL), 2022 on Pith
- [26] Arjun Majumdar, Karmesh Yadav, Sergio Arnaud, Yecheng Jason Ma, Claire Chen, et al. Where are we in the search for an artificial visual cortex for embodied intelligence? InAdvances in Neural Information Processing Systems (NeurIPS), pages 655–677, 2023 on Pith
- [27] Hongtao Wu, Ya Jing, Chilam Cheang, Guangzeng Chen, Jiafeng Xu, et al. Unleashing large-scale video generative pre-training for visual robot manipulation. InInternational Conference on Learning Representations (ICLR), pages 10641–10662, 2024 on Pith
- [28] Seonghyeon Ye, Joel Jang, Byeongguk Jeon, Sejune Joo, Jianwei Yang, et al. Latent action pretraining from videos. InInternational Conference on Learning Representations (ICLR), pages 28213–28239, 2025 on Pith
- [29] Qixiu Li, Yu Deng, Yaobo Liang, Lin Luo, Lei Zhou, Chengtang Yao, Lingqi Zeng, Zhiyuan Feng, Huizhi Liang, Sicheng Xu, et al. Scalable vision-language-action model pretraining for robotic manipulation with real-life human activity videos.arXiv preprint arXiv:2510.21571, 2025 on PitharXiv
- [30] Ryan Hoque, Peide Huang, David Yoon, Mouli Sivapurapu, and Jian Zhang. EgoDex: Learning dexterous manipulation from large-scale egocentric video. InInternational Conference on Learning Representations, pages 4218–4237, 2026. 26 YINWANG on Pith
- [31] Simar Kareer, Dhruv Patel, Ryan Punamiya, Pranay Mathur, Shuo Cheng, Chen Wang, Judy Hoffman, and Danfei Xu. EgoMimic: Scaling imitation learning via egocentric video. InIEEE International Conference on Robotics and Automation (ICRA), pages 13226–13233, 2025 on Pith
- [32] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. InInternational Conference on Learning Representations, 2023 on Pith
- [33] Wan Team et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025 on PithPith's reviewarXiv
- [34] Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-VL technical report.arXiv preprint arXiv:2511.21631, 2025 on PithPith's reviewarXiv
- [35] Ropedia. Xperience-10M: A large-scale egocentric multimodal dataset with structured 3d/4d annotations. https://huggingface.co/datasets/ropedia-ai/xperience-10m, 2026. Dataset on Pith
- [36] Hongwu Wang, Chenhao Yu, Youhao Hu, Jiachen Zhang, Yuanyuan Li, and Shaqi Luo. BifrostUMI: Bridg- ing robot-free demonstrations and humanoid whole-body manipulation.arXiv preprint arXiv:2605.03452, 2026 on PithPith's reviewarXiv
- [37] Macrodata Labs. Turning egocentric video into robot actions. https://macrodata.co/blog/ turning-egocentric-video-into-3d-hand-actions , 2026. Technical blog, August 6, on Pith
- [38] Jinglei Zhang, Jiankang Deng, Chao Ma, and Rolandos Alexandros Potamias. HaWoR: World-space hand motion reconstruction from egocentric videos. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1805–1815, 2025 on Pith
- [39] Accessed September 10, 2026 on Pith
- [40] Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Didac Suris Coll- Vinent, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang, et al. SAM 3: Segment anything with concepts. InInternational Conference on Learning Representations, pa... on Pith
- [41] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together.ACM Transactions on Graphics, 36(6):245:1–245:17, 2017 on Pith
- [42] Gaotian Wang, Kejia Ren, Andrew Morgan, Yiting Chen, Howard H. Qian, Podshara Chanrungmaneekul, and Kaiyu Hang. EgoInfinity: A web-scale 4d hand-object interaction data engine for any-view robot retargeting and video-to-action robot learning.arXiv preprint arXiv:2606.17385, 2026 on PithPith's reviewarXiv
- [43] Bruce D. Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. InProceedings of the International Joint Conference on Artificial Intelligence (IJCAI), pages 674–679, 1981 on Pith
- [44] Haoqi Yuan, Zhixuan Liang, Anzhe Chen, Ye Wang, Haoyang Li, Pei Lin, Yiyang Huang, Zixing Lei, Tong Zhang, Jiazhao Zhang, et al. Qwen-RobotManip technical report: Alignment unlocks scale for robotic manipulation foundation models.arXiv preprint arXiv:2606.17846, 2026. 27 YINWANG on PithPith's reviewarXiv
- [45] Bhawna Paliwal, Haritheja Etukuru, William Liang, Pieter Abbeel, Nur Muhammad Shafiullah, and Jitendra Malik. Do as i do: Dexterous manipulation data from everyday human videos.arXiv preprint arXiv:2606.19333, 2026 on PithPith's reviewarXiv
- [46] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. LIBERO: Bench- marking knowledge transfer for lifelong robot learning. InAdvances in Neural Information Processing Systems, pages 44776–44791, 2023 on Pith
- [47] Tianxing Chen, Zanxin Chen, Baijun Chen, Zijian Cai, Yibin Liu, Qiwei Liang, Zixuan Li, 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. InInternational ... on PithPith's reviewarXiv
- [48] Tao Jiang, Tianyuan Yuan, Yicheng Liu, Chenhao Lu, Jianning Cui, Xiao Liu, Shuiqi Cheng, Jiyang Gao, Huazhe Xu, and Hang Zhao. Galaxea open-world dataset and G0 dual-system VLA model.arXiv preprint arXiv:2509.00576, 2025. URLhttps://arxiv.org/abs/2509.00576 on PithPith's reviewarXiv
- [49] StarVLA Community. StarVLA: A lego-like codebase for vision-language-action model developing. arXiv preprint arXiv:2604.05014, 2026 on PithPith's reviewarXiv
- [50] Jinliang Zheng, Jianxiong Li, Zhihao Wang, Dongxiu Liu, Xirui Kang, Yuchun Feng, Yinan 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. ... on Pith
- [51] Rui Cai, Jun Guo, Xinze He, Piaopiao Jin, Jie Li, Bingxuan Lin, Futeng Liu, Wei Liu, Fei Ma, Kun Ma, Feng Qiu, Heng Qu, Yifei Su, Qiao Sun, Dong Wang, Donghao Wang, Yunhong Wang, Rujie Wu, Diyun Xiang, Yu Yang, Hangjun Ye, Yuan Zhang, and Quanyun Zhou. Xiaomi-robotics-0: An op... on PitharXiv
- [52] Yuyang Zhang, Wenyao Zhang, Zekun Qi, He Zhang, Haitao Lin, et al. ImageWAM: Do world action models really need video generation, or just image editing?arXiv preprint arXiv:2606.19531, 2026 on PithPith's reviewarXiv
- [53] Yandan Yang, Shuang Zeng, Tong Lin, Xinyuan Chang, Dekang Qi, et al. ABot-M0: Vla foundation model for robotic manipulation with action manifold learning.arXiv preprint arXiv:2602.11236, 2026 on PithPith's reviewarXiv
- [54] Lishan Yang, Wenxuan Song, Xi Wang, Pingyue Sheng, Zheng Fang, Ziyang Zhou, Junjie He, Haodong Yan, Jiayi Chen, Nan Sun, Qiao Sun, Pengwei Wang, Lingqiao Liu, Yan Wang, Yuxiang Gao, Feras Dayoub, and Haoang Li. 4D-W AM: Infusing spatiotemporal awareness into world action model... on PithPith's reviewarXiv
- [55] Jun Guo, Qiwei Li, Peiyan Li, Zilong Chen, Nan Sun, Yifei Su, Heyun Wang, Yuan Zhang, Xinghang Li, and Huaping Liu. Unified 4D world action modeling from video priors with asynchronous denoising. arXiv preprint arXiv:2604.26694, 2026 on PithPith's reviewarXiv
- [56] Wei Wu, Fan Lu, Yunnan Wang, Shuai Yang, Shi Liu, et al. A pragmatic vla foundation model.arXiv preprint arXiv:2601.18692, 2026 on PithPith's reviewarXiv
- [57] GigaBrain Team et al. Gigabrain-0.7: Scaling embodied foundation models to emergent capabilities with a three-system architecture.arXiv preprint arXiv:2608.15875, 2026 on PithPith's reviewarXiv
- [58] He Zhang, Lingzhu Xiang, Haitao Lin, Zeyu Huang, Minghui Wang, et al. Hy-Embodied-0.5-VLA: From vision-language-action models to a real-world robot learning stack.arXiv preprint arXiv:2606.14409, 2026. 28 YINWANG on PithPith's reviewarXiv
- [59] Tianle Zhang, Zhihao Yuan, Dafeng Chi, Peidong Liu, Dongwei Li, et al. JoyAI-RA 0.1: A foundation model for robotic autonomy.arXiv preprint arXiv:2604.20100, 2026 on PithPith's reviewarXiv
- [60] Angen Ye, Boyuan Wang, Chaojun Ni, Guan Huang, Guosheng Zhao, Hao Li, Hengtao Li, Jie Li, Jindi Lv, Jingyu Liu, et al. Gigaworld-policy: An efficient action-centered world-action model.arXiv preprint arXiv:2603.17240, 2026 on PitharXiv
- [61] Hao Li, Ganlong Zhao, Yufei Liu, Haotian Hou, Guoquan Ye, et al. ACE-Ego-0: Unifying egocentric human and robotic data for vla pretraining.arXiv preprint arXiv:2606.17200, 2026 on PithPith's reviewarXiv
- [62] Junhao Cai, Zetao Cai, Jiafei Cao, Yilun Chen, Zeyu He, Lei Jiang, Hang Li, Hengjie Li, Yang Li, Yufei Liu, et al. InternVLA-A1: Unifying understanding, generation and action for robotic manipulation.arXiv preprint arXiv:2601.02456, 2026 on PitharXiv
- [63] Jialei Chen, Kai Wang, Kang Chen, Shuaihang Chen, Feng Gao, et al. LaWAM: Latent world action models for efficient dynamics-aware robot policies.arXiv preprint arXiv:2606.15768, 2026 on PithPith's reviewarXiv
- [64] Qiuyue Wang, Mingsheng Li, Jian Guan, Jinhui Ye, Sicheng Xie, Yitao Liu, Junhao Chen, Zhixuan Liang, Jie Zhang, Xintong Hu, et al. Qwen-VLA: Unifying vision-language-action modeling across tasks, environments, and robot embodiments.arXiv preprint arXiv:2605.30280, 2026. URL ht... on PithPith's reviewarXiv
- [65] Yi Yang, Zhihong Liu, Siqi Kou, Yiyang Chen, Yanzhe Hu, et al. World-language-action model for unified world modeling, language reasoning, and action synthesis.arXiv preprint arXiv:2606.05979, 2026 on PithPith's reviewarXiv
- [66] Xiaomi Robotics Team, Jun Guo, Piaopiao Jin, Jason Li, Peiyan Li, Yingyan Li, Futeng Liu, Wanli Peng, Optimus Qin, Yifei Su, et al. Xiaomi-Robotics-1: Scaling vision-language-action models with over 100k hours of real-world trajectories.arXiv preprint arXiv:2607.15330, 2026 on PithPith's reviewarXiv
- [67] Kevin Black, Manuel Y . Galliker, and Sergey Levine. Real-time execution of action chunking flow policies. InAdvances in Neural Information Processing Systems, pages 33383–33407, 2025 on Pith
- [68] Shihan Wu, Xuecheng Liu, Shaoxuan Xie, Pengwei Wang, Xinghang Li, Bowen Yang, Zhe Li, Kai Zhu, Hongyu Wu, Yiheng Liu, et al. RoboCOIN: An open-sourced bimanual robotic data collection for integrated manipulation.arXiv preprint arXiv:2511.17441, 2025. URL https://arxiv.org/abs/... on PithPith's reviewarXiv
- [69] Ziyu Wang, Chenyuan Liu, Yushun Xiang, Runhao Zhang, Qingbo Hao, Hongliang Lu, Houyu Chen, Zhizhong Feng, Kaiyue Zheng, Dehao Ye, et al. The Great March 100: 100 detail-oriented tasks for evaluating embodied ai agents.arXiv preprint arXiv:2601.11421, 2026. URL https://arxiv.or... on PitharXiv
- [70] Adina Yakefu, Bin Xie, Chongyang Xu, Enwen Zhang, Erjin Zhou, Fan Jia, Haitao Yang, Haoqiang Fan, Haowei Zhang, Hongyang Peng, et al. RoboChallenge: Large-scale real-robot evaluation of embodied policies.arXiv preprint arXiv:2510.17950, 2025. URLhttps://arxiv.org/abs/2510.17950 on PitharXiv
- [71] Kun Wu, Chengkai Hou, Jiaming Liu, Zhengping Che, Xiaozhu Ju, Zhuqin Yang, Meng Li, Yinuo Zhao, Zhiyuan Xu, Guang Yang, et al. RoboMIND: Benchmark on multi-embodiment intelligence normative data for robot manipulation. InRobotics: Science and Systems (RSS), 2025. URL https: //... on Pith
- [72] Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. RDT-1B: a diffusion foundation model for bimanual manipulation. InInternational Conference on Learning Representations, pages 29982–30009, 2025. 29 YINWANG on Pith
- [73] Yunchao Ma, Ze Chen, Erjin Zhou, Ziming Liu, Haowei Zhang, Kai Liu, and Haoqiang Fan. Table30 V2: Evaluating generalized models by real robots at scale. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 4461–4467, 2026 on Pith
- [74] Rolandos Alexandros Potamias, Jinglei Zhang, Jiankang Deng, and Stefanos Zafeiriou. WiLoR: End-to- end 3d hand localization and reconstruction in-the-wild. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12242–12254, 2025 on Pith
- [75] Chengkai Hou, Kun Wu, Jiaming Liu, Zhengping Che, Di Wu, Fei Liao, Guangrun Li, Jingyang He, Qiuxuan Feng, Zhao Jin, et al. RoboMIND 2.0: A multimodal, bimanual mobile manipulation dataset for generalizable embodied intelligence.arXiv preprint arXiv:2512.24653, 2025. URL https... on PitharXiv
- [76] Abraham Savitzky and Marcel J. E. Golay. Smoothing and differentiation of data by simplified least squares procedures.Analytical Chemistry, 36(8):1627–1639, 1964. doi: 10.1021/ac60214a047. 30 YINWANG A Cross-Paradigm Attention Measurement Figure 1 compares action-query attenti... on Pithdoi:10.1021/ac60214a047
- [77] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. ByteTrack: Multi-object tracking by associating every detection box. InProceedings of the European Conference on Computer Vision (ECCV), pages 1–21, 2022 on Pith
- [79] = (¯qL 0,...,¯q L 27,¯qL 26,¯qL 27).(16) At each unified block, Eq. 4 combines the aligned Qwen3-VL state with the accumulated VL state. The normalized fused tokens are mapped by block-specific query, key, and value projections from 2,048 dimensions to 24 heads of width 128. V... on Pith
- [2026] URL https://openaccess.thecvf.com/content/CVPR2026/html/Bi_Motus_ A_Unified_Latent_Action_World_Model_CVPR_2026_paper.html. 25 YINWANG on Pith
Cited by 1 Pith paper
Reviewed papers on Pith that cite this one. The full citation record
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