Pith.

Robotics · arXiv:2609.04893 v1 · submitted 4 Sep 2026 · 18 references

Reasoning Without Inference Cost: Latent Semantic Scaffolding for Robot VLA Policies

Andrew Ting Yan Li, Zhuo Li, Zhelin Yang, Zhipeng Dong, Quentin Rouxel, Fei Chen

The paper says

“Vision-language-action (VLA) models are trained by imitation and capture what action to take but not why; adding causal reasoning improves manipulation, but current methods pay for it at inference time - generating reasoning tokens or rolling out predicted future states at every step, a cost that compounds over long horizons.”

more of the abstract

We ask whether this benefit can instead be captured during training and discarded before deployment. We introduce Latent Semantic Scaffolding (LSS), an auxiliary loss applied during human-demonstration pretraining that aligns a VLA's action-token representations to text embeddings of physical-reasoning rationales through a small projection head. The head is dropped at inference, leaving the unmodified base policy with zero added cost. Our central finding concerns alignment granularity: aligning each action token to the rationale of its own manipulation phase (Dense LSS) rather than to a single pooled episode-level embedding (Pooled LSS) yields representations that transfer markedly better to held-out tasks. Dense LSS attains both the best in-distribution success and the best transfer to tasks unseen during alignment, whereas pooled alignment over-specializes to the training task. A representational probe shows Dense LSS induces roughly twice the per-phase separability in the backbone, supporting that phase-local alignment is the operative mechanism.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 8 Sep 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. Solid training-time reasoning injection with zero inference cost; the dense-vs-pooled alignment finding is convincing but rests on a narrow empirical base (single task, single lambda). This line is wrong

3 major objections · 4 minor

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 18 references say.

The paper says, at Section IV-C, IV-D, Table I, Table II

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

The alignment weight λ is fixed at 0.1 throughout. Because the auxiliary loss is deliberately a small secondary pressure, the relative ordering of Dense vs. Pooled LSS may depend on the choice of λ. The paper should report a sweep over λ (e.g., 0.01, 0.05, 0.1, 0.2) for the in-distribution and at least one held-out task to demonstrate that the dense advantage is robust, not an artifact of this particular weight.

The paper says, at Section III-F

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

Dense LSS depends on VLM-generated phase boundaries and rationales with no formal quality control. While the authors flag this as a limitation, the central claim (dense vs. pooled) could be undermined if phase boundaries are systematically noisy. The paper would be significantly strengthened by: (a) reporting inter-annotator agreement between the VLM and a human on a random subset of episodes, …

more of objection 2

… or (b) a sensitivity analysis that perturbs phase boundaries (e.g., random shifts of ±10% of phase duration) to show the behavioral result is not driven by coincidentally correct segmentation.

The paper says, at Section IV-E, Figure 5

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The silhouette scores used in the mechanism probe are very small (0.047 vs. 0.021). While the authors note this is by design (λ=0.1), the supporting evidence would be more convincing if the probe were also applied to other tasks (e.g., the held-out tasks after Stage-3 finetune) to show that the separability difference persists through finetuning and correlates with transfer performance.

4 minor objections

The paper says, at Table I, Table II

Pith says: minor objection 1 of 4

The abbreviation 'A VP' is used without definition. Please define it when first introduced (e.g., 'human data from Apple Vision Pro (A VP)').

The paper says, at Section III-C, III-D

Pith says: minor objection 2 of 4

The term 'LSAHead' is introduced but not expanded. Either define it or rename to 'projection head' for clarity.

show 2 more

The paper says, at Section IV-E, Figure 5 caption

Pith says: minor objection 3 of 4

The phrase 'alignment budget' is non-standard. Replace with 'alignment weight λ' for consistency with the main text.

The paper says, at Section II-B

Pith says: minor objection 4 of 4

The reference to REPA [15] is appropriate but the paper could also cite more recent alignment-based methods in control, e.g., 'R3M' or 'VIP', to better position LSS in the representation-learning landscape.

What would settle it

Re-run the Table II transfer ladder with the same 50-episode / 22,000-step Stage-3 protocol for R3 (Pooled LSS) and R4 (Dense LSS) at lambda = 0.05 and lambda = 0.2, in addition to the reported lambda = 0.1, on adjust_bottle and shake_bottle. If Dense LSS remains at or above Pooled LSS on both tasks across all three lambda values, the central granularity claim is robust to the secondary-loss budget; if the ordering flips or becomes non-monotonic, the headline claim needs conditioning on lambda.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 8 Sep 2026

  1. Reader holds up, with conditions confidence high
  2. Skeptic partly agrees 18 s · 9,534 tokens read, 2,032 written
  3. Editor wrote the verdict above 17 s · 10,375 tokens read, 1,733 written
  4. Referee major revision: 3 major objections, 4 minor 25 s · 9,789 tokens read, 2,638 written

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

The paper makes a clean contribution: LSS aligns VLA action representations to reasoning text embeddings during training and discards the apparatus at inference, achieving zero runtime cost. The central finding—phase-local (dense) alignment transfers to held-out tasks while pooled alignment over-specializes—is well supported by the controlled finetuning ladder in Table II, where R2–R4 share the same human data and Stage-3 budget. The ablation ladder in Table I further isolates the reasoning content and alignment granularity. The paper is honestly written, flags its own limitations (single task, no formal VLM quality control, no code release), and treats the mechanism probe as supporting rather than primary evidence.

The soft spots are proportionate. The dense-vs-pooled advantage is shown only at λ=0.1; the mechanism probe reports small absolute silhouette values (0.047 vs. 0.021), and while the relative gap is consistent, sensitivity to the alignment weight is unexplored. This is a standard hyperparameter-robustness concern, not a fatal flaw. The reliance on VLM-generated phase boundaries without quality control is acknowledged; noisy boundaries would likely dilute the dense advantage rather than create a spurious one, so this is a minor worry. The single human-pretraining task limits generality, but the paper is upfront and the transfer results to structurally different tasks (move_can_pot) strengthen the claim.

Who is this for? Robot-learning researchers working on representation learning for VLAs and on reducing inference cost. The method is straightforward and clearly described.

My recommendation: this paper deserves a serious referee. The claims are specific, the experimental design is thoughtful, and the limitations are honestly stated. I would support conditional acceptance once the authors provide at least a sensitivity analysis on λ (and ideally a second task or domain). I'd bring it to reading group as a clean example of a focused contribution with a careful ladder design.

The strongest objection

After reading in good faith, I cannot identify a defect that invalidates the central claim. The paper argues that reasoning can be captured at training time by aligning action-token representations to per-phase rationale embeddings, and that phase-local (Dense LSS) alignment transfers better than episode-pooled alignment. The load-bearing premise is that the 90% vs. 85% in-distribution gap and the Table II transfer pattern (54/45/39, 26/20/19) are caused by alignment granularity rather than by uncontrolled differences between the Stage-2 backbones. The paper's design largely supports this: R2, R3, and R4 share the same human data and differ only in the alignment objective, with an identical Stage-3 budget, and all R1 baselines are reproduced on their hardware. The main residual threat is the one the paper itself flags in Section V and in the Caveat after Eq. (2): Dense LSS relies on VLM-generated phase boundaries with no formal quality control. If phase labels were noisy, token-to-phase assignments would be corrupted; however, noisy labels would presumably bias the result toward pooled-like or weaker behavior, not spuriously create a dense advantage. The more consequential open question is whether the dense advantage is robust to the choice of the alignment weight lambda. Both dense and pooled results are reported at lambda=0.1 only. Since the mechanism probe shows small absolute silhouette values (0.047 vs. 0.021), it is plausible that the dense-vs-pooled ordering is sensitive to lambda, and a different lambda could make the granularity claim non-monotonic or reverse it. This is a standard hyperparameter-robustness concern rather than an internal inconsistency, and it does not rise to a verdict change.

The reader's verdict (CONDITIONAL, high confidence) appears appropriate: the paper is a coherent training-time alignment method with a controlled ablation ladder, honest limitation statements, and an internally consistent mechanism story. The concern I would stress-test is not exactly the reader's chosen weakest assumption. The VLM phase-quality limitation is real, acknowledged in Section III-D (Caveat), Section III-F, and Section V, and the reader correctly identifies it as unverified. However, a formal quality check of phase labels would mainly establish that the Dense LSS alignment targets are correct; it is less central to the paper's claim than the robustness of the dense-vs-pooled ordering to the alignment weight lambda. The paper reports all results at lambda=0.1, and the mechanism probe (Section IV-E, Fig. 5) describes the representational shift as 'small by design' (silhouette 0.047 vs. 0.021). Since the auxiliary loss explicitly must not override the flow-matching objective, the scale of lambda is a deliberately sensitive knob: too small and LSS has no effect; too large and it degrades action prediction. The claim that Dense LSS, not Pooled LSS, is the transferable regime should hold across a reasonable range of lambda, not just at one value. This is a concrete, falsifiable check and does not require re-deriving any equation. I also note the paper's independent supports: a full ablation ladder including a frozen-backbone control (Run F, 0%), a trivial-target control (58%), and a reasoning-as-input control (Run C, 71%), plus the explicit statement that all Table II baselines were reproduced under an identical protocol. The absence of code and data release and the single pretraining task are scoping limitations already reflected in the CONDITIONAL verdict. A lambda sweep, one histogram of phase-boundary agreement, and one run with an independently re-annotated phase subset would settle the remaining concerns. None of these observations warrant changing the reader's verdict, so I recommend UNCHANGED.

The referee's report

If the results hold, LSS is a practical method to inject reasoning into VLA policies without inference-time overhead, which is a clear contribution given the high cost of test-time reasoning methods (e.g., embodied chain-of-thought, world action models). The alignment-granularity finding (dense vs. pooled) is novel and has practical implications: episode-level alignment over-specializes, while phase-local alignment transfers. The experimental design is well-controlled, with all Stage-3 runs using identical budgets and backbones differing only in the Stage-2 alignment objective. The paper is clearly written and the claims are appropriately scoped.

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

Latent Semantic Scaffolding (LSS) is a training-only auxiliary loss that aligns a vision-language-action (VLA) model's action-token representations to text embeddings of physical-reasoning rationales from a frozen encoder. The alignment head and encoder are discarded at inference, leaving the base policy with no added computation. The paper's central finding concerns alignment granularity: aligning each action token to the rationale of its own manipulation phase (Dense LSS) transfers to held-out tasks markedly better than aligning a single pooled episode-level embedding (Pooled LSS). In a controlled finetuning ladder, Dense LSS achieves the highest success on both in-distribution and unseen tasks, whereas Pooled LSS over-specializes to the training task. A representational probe shows Dense LSS induces roughly twice the per-phase separability in the backbone, supporting phase-local shaping as the mechanism.

What carries the argument. The central object is Latent Semantic Scaffolding (LSS), a cosine-distance auxiliary loss added during human-demonstration pretraining that aligns a VLA's action-token hidden states (via a two-layer MLP projection head g_phi) to precomputed T5 embeddings of VLM-generated physical-reasoning text. The crucial variant is Dense LSS, which aligns each token to the embedding of its own manipulation phase (approach, grip, rotate, withdraw) rather than to a single mean-pooled episode embedding. The LSAHead and reasoning encoder are used only during training and discarded at inference.

Core claim. The paper establishes a training-time method that captures the benefit of causal reasoning for robot VLA policies without incurring inference cost. The key result is that aligning each action token (through a small projection head) to the T5 embedding of its own manipulation-phase reasoning rationale—rather than to a single pooled episode embedding—produces representations that generalize to tasks not seen during alignment. The deployed policy carries zero extra computation because the projection head and reasoning encoder are discarded; the reasoning influence persists only in the trained backbone weights.

The premise it rests on. Dense LSS depends on the quality of VLM-generated phase boundaries and rationales, for which the paper applies no formal quality control; noisy segmentation could misalign tokens and degrade the advantage over pooled alignment.

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

LSS enables deployment of a standard VLA policy—with no runtime reasoning generation or state rollouts—while still benefiting from reasoning-shaped representations.

Alignment granularity is a decisive design lever: per-token phase-local alignment transfers to unfamiliar tasks, whereas episode-pooled alignment over-specializes and can regress on held-out tasks.

The insight that reasoning can be 'baked into' weights rather than generated step-by-step opens a new axis for efficient manipulation: training-time computation can replace test-time computation without loss of task performance.

The representational probe (silhouette score) provides a diagnostic that stronger per-phase separability in the backbone correlates with better transfer, suggesting a general design principle for auxiliary alignment losses.

LSS is compatible with any VLA backbone that produces a sequence of action tokens and can be inserted between human pretrain and robot finetune stages with only an additional lightweight supervised stage.

The paper’s reliance on VLM-generated phase boundaries for Dense LSS is a brittleness point: future work could test whether human-annotated boundaries or a learned segmentation improve (or degrade) the gap between dense and pooled alignment.

The results raise a broader question: if reasoning can be captured at train time via alignment, might other forms of privileged information (object pose, contact forces, future states) also be 'scaffolded' into the backbone without runtime overhead?

The fact that Pooled LSS still helps in-distribution but fails to transfer suggests that collapsing temporal structure destroys phase-specific cues that are essential for generalization—a caution for any method that uses episode-level pooled representations.

Zero-cost inference reasoning could enable longer-horizon tasks where test-time reasoning would compound, and might be combined with light test-time refinements if the task requires adaptation beyond the training distribution.

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

Vision-language-action (VLA) models are trained by imitation and capture what action to take but not why; adding causal reasoning improves manipulation, but current methods pay for it at inference time - generating reasoning tokens or rolling out predicted future states at every step, a cost that compounds over long horizons. We ask whether this benefit can instead be captured during training and discarded before deployment. We introduce Latent Semantic Scaffolding (LSS), an auxiliary loss applied during human-demonstration pretraining that aligns a VLA's action-token representations to text embeddings of physical-reasoning rationales through a small projection head. The head is dropped at inference, leaving the unmodified base policy with zero added cost. Our central finding concerns alignment granularity: aligning each action token to the rationale of its own manipulation phase (Dense LSS) rather than to a single pooled episode-level embedding (Pooled LSS) yields representations that transfer markedly better to held-out tasks. Dense LSS attains both the best in-distribution success and the best transfer to tasks unseen during alignment, whereas pooled alignment over-specializes to the training task. A representational probe shows Dense LSS induces roughly twice the per-phase separability in the backbone, supporting that phase-local alignment is the operative mechanism.

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 · 2026-09-23

The authors' main claim, as the classifier picked it out 0.93

“Our central finding concerns alignment granularity: aligning each action token to the rationale of its own manipulation phase (Dense LSS) rather than to a single pooled episode-level embedding (Pooled LSS) yields representations that transfer markedly better to held-out tasks.”

How the main claim is stated: asserted 0.99

What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.23 · human participants 0.30 · controlled comparison 0.82 · randomized assignment 0.02 · states quantitative result 0.91 · compares with prior results 0.87 · 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
Vision-language-action (VLA) models are trained by imitation and capture what action to take but not why; adding causal reasoning improves manipulation, but current methods pay for it at inference time - generating reasoning tokens or rolling out predicted future states at every step, a cost that compounds over long horizons. background 0.90 0.07 0.04 asserted 1.00
We ask whether this benefit can instead be captured during training and discarded before deployment. gap 0.84 0.10 0.04 unclear 0.38
We introduce Latent Semantic Scaffolding (LSS), an auxiliary loss applied during human-demonstration pretraining that aligns a VLA's action-token representations to text embeddings of physical-reasoning rationales through a small projection head. method 0.79 0.93 0.05 asserted 0.99
The head is dropped at inference, leaving the unmodified base policy with zero added cost. method 0.89 0.85 0.06 asserted 1.00
Our central finding concerns alignment granularity: aligning each action token to the rationale of its own manipulation phase (Dense LSS) rather than to a single pooled episode-level embedding (Pooled LSS) yields representations that transfer markedly better to held-out tasks. result 0.97 0.94 0.84 asserted 0.99
Dense LSS attains both the best in-distribution success and the best transfer to tasks unseen during alignment, whereas pooled alignment over-specializes to the training task. result 1.00 0.95 0.61 asserted 1.00
A representational probe shows Dense LSS induces roughly twice the per-phase separability in the backbone, supporting that phase-local alignment is the operative mechanism. result 0.67 0.96 0.09 asserted 0.87

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.97

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, 2026-09-23: 1 request, 5,775 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. LSS injects reasoning at training time and discards it at inference. Left (training): the backbone is trained on the flow-matching objective while an auxiliary projection head gϕ aligns each action token to the T5 embedding of its manipulation-phase reasoning—an auxiliary branch that shapes the backbone but does not lie on the action-production path. Right (inference): the head and reasoning encoder are removed; the deployed policy is the unmodified base VLA with reasoning “baked in,” at zero added cost. I. INTRODUCTION Vision-language-action models (VLAs) map perception to action…
Figure 2 of the paper
Figure 2. Dense LSS architecture (Stage 2, training only). A frozen DINO+SigLIP encoder and a frozen T5-XXL command encoder condition the 2B diffusion-transformer backbone, which produces K=16 action tokens Zt ∈ R2176 supervised by the flow-matching loss Lflow. In parallel, the LSAHead projection gϕ maps each action token into T5 space and a cosine-distance loss aligns it to the reasoning embedding of its own manipulation phase, ephase(t) (straddling tokens masked by mt). The combined objective is Ltotal = Lflow + λLLSS. Pooled LSS uses the identical architecture but replaces the per-phase…
Figure 3 of the paper
Figure 3. Dense LSS token-to-phase alignment. Each of the 16 action tokens Zt is assigned to the manipulation phase spanned by its frames (tokens straddling a boundary, e.g. the greyed slot, are masked by mt); color indicates the assigned phase (orange = grip, green = rotate; two representative phases shown). The per-phase reasoning text is encoded by a frozen T5 and mean-pooled into a target ephase(t) . The head gϕ projects each token into T5 space and Ldense aligns it to the target of its own phase. frame ranges. Phase boundaries are validated primarily by VLM-vs-human video comparison; a…
Figure 4 of the paper
Figure 4. Human demonstration data (top) and simulation evaluation environ
Figure 5 of the paper
Figure 5. Per-phase separability of the LSS-induced representational change on

Figures are the authors', reproduced from the paper.

References (18)
  1. [1] A. Brohanet al., “RT-2: Vision-language-action models transfer web knowledge to robotic control,”arXiv preprint arXiv:2307.15818, 2023 on PithPith's reviewarXiv
  2. [2] M. J. Kim, K. Pertsch, S. Karamchetiet al., “OpenVLA: An open-source vision-language-action model,”arXiv preprint arXiv:2406.09246, 2024 on PithPith's reviewarXiv
  3. [3] K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichteret al., “π 0: A vision-language-action flow model for general robot control,”arXiv preprint arXiv:2410.24164, 2024, rSS 2025 on PithPith's reviewarXiv
  4. [4] A. Khazatsky, K. Pertschet al., “DROID: A large-scale in-the-wild robot manipulation dataset,”arXiv preprint arXiv:2403.12945, 2024 on PithPith's reviewarXiv
  5. [5] H. Bi, L. Wu, T. Lin, H. Tan, Z. Su, H. Su, and J. Zhu, “H-RDT: Human manipulation enhanced bimanual robotic manipulation,”arXiv preprint arXiv:2507.23523, 2025 on PithPith's reviewarXiv
  6. [6] R. Yang, Q. Yu, Y . Wu, R. Yan, B. Li, A.-C. Cheng, X. Zou, Y . Fang, H. Yin, S. Liu, S. Han, Y . Lu, and X. Wang, “EgoVLA: Learning vision- language-action models from egocentric human videos,”arXiv preprint arXiv:2507.12440, 2025 on PithPith's reviewarXiv
  7. [7] H. Luo, Y . Feng, W. Zhang, S. Zheng, Y . Wang, H. Yuan, J. Liu, C. Xu, Q. Jin, and Z. Lu, “Being-H0: Vision-language-action pretraining from large-scale human videos,”arXiv preprint arXiv:2507.15597, 2025 on PithPith's reviewarXiv
  8. [8] Q. Zhao, Y . Lu, M. J. Kim, Z. Fu, Z. Zhang, Y . Wu, Z. Li, Q. Ma, S. Han, C. Finn, A. Handa, M.-Y . Liu, D. Xiang, G. Wetzstein, and T.-Y . Lin, “CoT-VLA: Visual chain-of-thought reasoning for vision-language- action models,”arXiv preprint arXiv:2503.22020, 2025, cVPR 2025 on PithPith's reviewarXiv
  9. [9] M. Zawalski, W. Chen, K. Pertsch, O. Mees, C. Finn, and S. Levine, “Robotic control via embodied chain-of-thought reasoning,”arXiv preprint arXiv:2407.08693, 2024, coRL 2024 on PithPith's reviewarXiv
  10. [10] J. Cen, C. Yu, H. Yuan, Y . Jiang, S. Huang, J. Guo, X. Li, Y . Song, H. Luo, F. Wang, D. Zhao, and H. Chen, “WorldVLA: Towards autoregressive action world model,”arXiv preprint arXiv:2506.21539, 2025 on PithPith's reviewarXiv
  11. [11] Z. Zhang, Z. Li, B. Rahmati, R. H. Yang, Y . Ma, A. Rasouli, S. Pak- damansavoji, Y . Wu, L. Zhang, T. Cao, F. Wen, X. Wang, X. Quan, and Y . Zhang, “Do world action models generalize better than VLAs? a robustness study,”arXiv preprint arXiv:2603.22078, 2026 on PithPith's reviewarXiv
  12. [12] Z. Duan, Y . Zhang, S. Geng, G. Liu, J. Boedecker, and C. X. Lu, “Fast ECoT: Efficient embodied chain-of-thought via thoughts reuse,”arXiv preprint arXiv:2506.07639, 2025 on PitharXiv
  13. [13] Open X-Embodiment Collaboration, “Open X-embodiment: Robotic learning datasets and RT-X models,” inIEEE International Conference on Robotics and Automation (ICRA), 2024 on Pith
  14. [14] Z. Zhang, Y . He, Y . Sun, J. Shi, L. Liu, and Q. Nie, “RoboAct-CLIP: Video-driven pre-training of atomic action understanding for robotics,” arXiv preprint arXiv:2504.02069, 2025, iROS 2025 on PithPith's reviewarXiv
  15. [15] S. Yu, S. Kwak, H. Jang, J. Jeong, J. Huang, J. Shin, and S. Xie, “Rep- resentation alignment for generation: Training diffusion transformers is easier than you think,”arXiv preprint arXiv:2410.06940, 2024, iCLR 2025 on PithPith's reviewarXiv
  16. [16] T. Chen, Z. Chen, B. Chen, Z. Cai, Y . Liu, Z. Li, Q. Liang, X. Lin, Y . Ge, Z. Guet al., “RoboTwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation,”arXiv preprint arXiv:2506.18088, 2025 on PithPith's reviewarXiv
  17. [17] R. Hoque, P. Huang, D. J. Yoon, M. Sivapurapu, and J. Zhang, “EgoDex: Learning dexterous manipulation from large-scale egocentric video,” arXiv preprint arXiv:2505.11709, 2025 on PithPith's reviewarXiv
  18. [18] R.-Z. Qiu, S. Yang, X. Cheng, C. Chawla, J. Li, T. He, G. Yan, D. J. Yoon, R. Hoque, L. Paulsen, G. Yang, J. Zhang, S. Yi, G. Shi, and X. Wang, “Humanoid policy∼human policy,”arXiv preprint arXiv:2503.13441, 2025 on PitharXiv
Cited by 2 Pith papers

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