Pith. sign in

REVIEW 1 major objections 6 minor 38 references

Half the robot demos, half the training, better policy

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

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

T0 review · glm-5.2

2026-07-08 05:35 UTC pith:43YIADF6

load-bearing objection SIEVE selects 50% of VLA training data via primitive-composition structure and matches or beats full-data training — but all results are single-seed, and the margins are small enough to be noise. the 1 major comments →

arxiv 2607.06442 v1 pith:43YIADF6 submitted 2026-07-07 cs.RO

SIEVE: Structure-Aware Data Selection for Imitation Learning with VLA Models

Changti Wu , Bin Yu , Zhaolong Shen , Shijie Lian , Xiaopeng Lin , Cong Huang , Zhirui Zhang , Lei Zhang
show 1 more author
Kai Chen
This is my paper
classification cs.RO
keywords dataselectionsievedemonstrationsimitationlearningmodelsprimitives
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

SIEVE is a data selection method for training robot policies by imitation learning. Its central claim is that robot demonstration trajectories are not monolithic sequences but compositions of reusable behavioral primitives (like 'reach, grasp, release') connected by transition interfaces. By segmenting trajectories at gripper state changes, clustering the resulting segments into a vocabulary of primitives, and then selecting demonstrations that maximally expose these reusable structures under diminishing returns, SIEVE can train better policies with less data and less computation. The method works in three stages: first, it discovers primitives by segmenting trajectories at end-effector grasp/release flips and clustering the segment representations; second, it allocates a selection budget across composition patterns (ordered sequences of primitives) using a greedy objective that rewards broad exposure to diverse primitives and transitions rather than redundant reinforcement of common ones; third, within each composition-pattern bucket, it selects medoid trajectories (those closest to the cluster center) to provide stable, predictable supervision for behavior cloning. The paper shows that on the Bridge-V2 dataset with a 4-billion-parameter VLA model, SIEVE achieves 56.3% average task success using 50% of demonstrations and 25K training steps, compared to 51.8% for full-data training with 50K steps. Similar gains hold across two additional datasets and a second VLA model architecture. The result implies that reusable compositional structure is a stronger signal for data utility than either holistic trajectory similarity or individual state-action quality.

Core claim

The paper establishes that selecting training demonstrations based on their coverage of reusable primitive-transition composition patterns yields better imitation learning policies than training on the full dataset, while using only half the data and half the training steps. The key mechanism is a three-stage pipeline: trajectory segmentation at gripper state boundaries produces visuo-motor primitives via clustering; a structural exposure objective with logarithmic diminishing returns allocates selection budgets across composition patterns to maximize coverage of diverse primitives and their transition interfaces; and medoid selection within each pattern bucket picks stable, central realizex

What carries the argument

The central machinery is the composition-pattern representation of trajectories. Each trajectory is segmented at end-effector grasp/release state flips into visuo-motor segments, which are encoded with a pretrained video encoder (V-JEPA2), dimensionality-reduced via PCA, and clustered via Mini-Batch K-Means into a vocabulary of K primitives (where K is auto-selected by maximizing a reuse-discriminability criterion). Each trajectory then becomes an ordered sequence of primitive IDs called a composition pattern, plus the transition interfaces between adjacent primitives. A greedy structural exposure objective allocates selection budget across composition-pattern buckets, weighting each prim

Load-bearing premise

The method segments trajectories at end-effector state flips (gripper grasp/release) under the assumption that these physically grounded boundaries correctly divide demonstrations into reusable behavioral primitives. If meaningful behavior transitions occur without gripper state changes, or if gripper flips happen mid-behavior, the entire primitive vocabulary and downstream composition analysis would be built on an invalid structural prior. The paper does not test alternative

What would settle it

A task domain where meaningful behavioral transitions do not align with gripper grasp/release state changes. For example, tasks dominated by continuous contact manipulation (pushing, sliding, wiping) where the gripper state rarely or never flips, or tasks where a single grasp-release cycle contains multiple distinct behavioral phases. In such domains, the segmentation would produce primitives that do not correspond to reusable behavioral units, and the structural exposure allocation would optimize over a degenerate composition-pattern space, likely yielding no improvement over random selection

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

If this is right

  • Robot learning datasets could be pre-filtered for compositional diversity at collection time, reducing storage and annotation costs before training begins.
  • The primitive vocabulary discovered by SIEVE could serve as an interpretable diagnostic tool: composition-pattern distributions reveal which behavioral structures are over- or under-represented in a dataset.
  • If gripper-state segmentation generalizes to dexterous hands and multi-finger manipulators, the same pipeline could apply to increasingly complex embodiments without architectural changes.
  • The structural exposure objective could be inverted for active data collection: instead of selecting from an existing pool, it could guide which new demonstrations to collect to maximize compositional coverage.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

1 major / 6 minor

Summary. The paper proposes SIEVE, a structure-aware data selection method for imitation learning with VLA models. SIEVE segments trajectories at gripper state flips, clusters segment representations (from V-JEPA2) into reusable primitives, allocates a selection budget across composition patterns by maximizing a structural exposure objective with diminishing returns, and selects medoid trajectories within each pattern bucket. The method is evaluated on three datasets (Bridge-V2, Fractal, GR00T-X-Sim) and two VLA models (Qwen3-VL-4B-GR00T, Qwen3-VL-4B-OFT), showing that SIEVE with 50% data and 50% training steps can match or exceed full-data training. The structural exposure objective (Eq. 11) and K-selection criterion (Eq. 6) are derived from internal principles (MDL, reuse-discriminability) without fitting to downstream task performance, which is a strength. The ablation study (Table 4) supports each component. The central concern is that all results appear to be from single-seed runs with no confidence intervals or standard deviations, making it difficult to assess whether the reported margins (1.4–4.5 points) reflect systematic improvement over training noise.

Significance. The paper addresses a practically important problem: selecting compact, informative subsets from large robot demonstration datasets for VLA training. The methodological design is principled: the structural exposure objective (Eq. 11) is derived from MDL considerations without fitting to downstream task performance, the K-selection criterion (Eq. 6) balances reuse and discriminability internally, and the greedy budget allocation is well-specified. The ablation study (Table 4) isolates the contribution of transition exposure, primitive exposure, and medoid selection. The composition-pattern redistribution analysis (Figure 3) provides qualitative insight into how SIEVE reshapes the selected subset. Code is publicly available. These are genuine strengths. However, the significance of the empirical contribution is tempered by the absence of statistical significance testing, which is needed to determine whether the headline claim—that SIEVE surpasses full-data training—holds reliably.

major comments (1)
  1. Tables 1–3 and 6–7: All success rates are reported as single numbers with no confidence intervals, standard deviations, or multiple seeds. The headline claim is that SIEVE surpasses full-data training with 50% data and 50% steps, but the margins are small enough to fall within typical VLA training noise: 4.5 points on Bridge-V2 (56.3 vs 51.8), 1.4 points on Fractal (76.4 vs 75.0), and 2.1 points on GR00T-X-Sim (54.8 vs 52.7). On GR00T-X-Sim, Random selection also beats Full-Training (53.5 vs 52.7), and SIEVE's margin over Full-Training is only 1.3 points larger than Random's. Without multiple seeds, it is impossible to determine whether these gains reflect a systematic advantage of SIEVE's structural exposure allocation or run-to-run variance. This is the most load-bearing issue because it directly determines whether the headline claim is true. At minimum, the authors should report 3+种子,
minor comments (6)
  1. Primitive Discovery, Trajectory Segmentation paragraph: the choice of gripper state flips as segmentation boundaries is a key design decision, but no sensitivity analysis or comparison with alternative segmentation strategies is provided. A brief discussion of failure modes (e.g., tasks where behavior transitions do not coincide with gripper state changes) would strengthen the paper. This is a presentation gap, not a correctness issue, as the method is still well-specified.
  2. Eq. (6): the K-selection criterion (1-J)log R combines a Jaccard-based discriminability measure and a reuse measure. The choice of this specific functional form is not justified beyond intuition. A brief note on why this form (vs alternatives) or a reference to prior work using similar criteria would help readers.
  3. Table 1: the 'Training Steps' column header is ambiguous. It mixes absolute step counts (25K, 35K, 50K) with percentage annotations (50%, 70%, 100%). Consider standardizing the notation.
  4. Figure 3: the x-axis labels showing composition patterns as bracketed lists (e.g., [8,8,8]) are difficult to read at normal scale. Consider providing a zoomed inset or a table summarizing the top patterns.
  5. The paper does not report computational overhead of SIEVE's selection pipeline (clustering, greedy allocation, medoid selection) relative to training cost. A brief note on wall-clock time for the selection step would be useful for practitioners.
  6. References include several 2026 arXiv preprints (e.g., Xu et al. 2026, Wu et al. 2026, Lian et al. 2026). Ensure these are properly cited and accessible.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful and constructive review. The referee correctly identifies that our headline claim—SIEVE surpassing full-data training—rests on single-seed results without confidence intervals, and that the reported margins (1.4–4.5 points) are small enough to potentially fall within VLA training noise. We agree this is the most load-bearing issue and will address it by running multiple seeds for the key comparisons. Below we respond point by point.

read point-by-point responses

  1. Referee: Tables 1–3 and 6–7: All success rates are reported as single numbers with no confidence intervals, standard deviations, or multiple seeds. The headline claim is that SIEVE surpasses full-data training with 50% data and 50% steps, but the margins are small enough to fall within typical VLA training noise: 4.5 points on Bridge-V2 (56.3 vs 51.8), 1.4 points on Fractal (76.4 vs 75.0), and 2.1 points on GR00T-X-Sim (54.8 vs 52.7). On GR00T-X-Sim, Random selection also beats Full-Training (53.5 vs 52.7), and SIEVE's margin over Full-Training is only 1.3 points larger than Random's. Without multiple seeds, it is impossible to determine whether these gains reflect a systematic advantage of SIEVE's structural exposure allocation or run-to-run variance. This is the most load-bearing issue because it directly determines whether the headline claim is true. At minimum, the authors should report 3+种子.

    Authors: The referee is correct. Our current results are from single-seed runs, and we agree that the reported margins—particularly on Fractal (1.4 points) and GR00T-X-Sim (2.1 points)—are small enough that they cannot be distinguished from training noise without replication. We also agree that the GR00T-X-Sim setting is especially concerning in this regard, since Random selection already exceeds Full-Training there, making it difficult to attribute SIEVE's advantage to its structural exposure allocation rather than to the general effect of training on a smaller subset in a relatively in-domain simulated setting. We will address this as follows: (1) We will re-run the key comparisons—Full-Training, Random, and SIEVE—at 3 seeds for the 50% selection budget on all three datasets (Bridge-V2, Fractal, GR00T-X-Sim) using Qwen3-VL-4B-GR00T, and report mean ± standard deviation in all main tables. We will also run at least 3 seeds for the Qwen3-VL-4B-OFT comparison on Bridge-V2. (2) We will add a brief statistical analysis (e.g., paired comparisons or bootstrap confidence intervals on the per-task success rates) to assess whether SIEVE's advantage over Full-Training and Random is statistically significant. (3) If, after replication, the margin on a particular dataset (e.g., Fractal or GR00T-X-Sim) is not statistically significant, we will state this honestly and adjust the headline claim accordingly—e.g., 'SIEVE matches or exceeds full-data training' rather than claiming a consistent surpass. We note that the current evidence does provide some reason to expect systematic improvement: SIEVE outperforms Random on every individual task across all three datasets and both models (Tables 1–3, 6–7), and the ablation study (Table 4) shows that each component contributes in the expected方向. This revision: no

Circularity Check

0 steps flagged

No significant circularity: SIEVE's derivation chain is self-contained, with parameters selected by internal heuristics rather than downstream task performance.

full rationale

The paper's core derivation chain — primitive discovery (Eq. 2-6), structural exposure allocation (Eq. 11-15), and learning-friendly trajectory selection (Eq. 17) — is constructed from first principles (MDL, reuse-discriminability, diminishing returns) without fitting any parameter to downstream task success rates. The K-selection criterion (Eq. 6) optimizes an internal clustering score on a 10% subset, not the target metric. The structural exposure objective (Eq. 11) is a greedy optimization over composition patterns with no fitted parameters. Self-citations (Wu et al. 2026 ScalSelect, Lian et al. 2026 IntentVLA) appear in related work and are not load-bearing for the method's derivation. The segmentation boundary choice (gripper state flips) is a design assumption, not a circular definition. No step in the derivation reduces to its own inputs by construction.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

The axiom ledger reveals that SIEVE introduces no postulated entities without independent evidence. The three main constructs (primitives, transitions, composition patterns) are all data-driven and validated through ablations. The primary risk lies in the domain assumptions, particularly the gripper-state-flip segmentation boundary, which is the most consequential and least validated design choice.

free parameters (5)
  • Number of frames sampled per segment = 8
    Chosen by hand for segment representation extraction; no sensitivity analysis provided.
  • PCA dimensionality = 256
    Chosen by hand to reduce segment representation; no justification or sensitivity analysis.
  • State transition persistence threshold = 5
    Number of consecutive frames for accepting a gripper state flip as a boundary; chosen ad hoc.
  • Number of clusters K = Dataset-dependent (auto-selected)
    Selected by maximizing (1-J)logR on a 10% subsample. Not fitted to task performance, but the search range and sampling fraction are hand-chosen.
  • Selection budget = 50% or 70%
    Experimental condition, not a fitted parameter of the method itself.
axioms (5)
  • domain assumption End-effector state flips (gripper grasp/release) are physically grounded interaction boundaries that correctly segment trajectories into reusable primitives.
    Primitive Discovery section, Trajectory Segmentation paragraph. This is the foundational structural prior. No validation or comparison with alternative segmentation strategies is provided.
  • domain assumption V-JEPA2 pretrained video encoder produces representations that capture semantically meaningful behavior content for robot manipulation segments.
    Primitive Discovery section, Segment Representation paragraph. The entire primitive discovery pipeline depends on the quality of these features.
  • domain assumption Trajectories near the center of a composition-pattern bucket provide more stable and consistent supervision for behavior cloning than outliers.
    Learning-Friendly Trajectory Selection section. Used to justify medoid selection as a proxy for action consistency.
  • standard math The MDL principle implies that useful training data is rich in extractable, reusable structure (non-random regularities a bounded learner can internalize).
    Introduction, paragraph 3. The connection from MDL to the structural exposure objective is conceptual rather than formally derived.
  • domain assumption Mini-Batch K-Means is an appropriate clustering algorithm for discovering behavior primitives in segment representation space.
    Primitive Discovery section, Primitive discovery via clustering paragraph. No comparison with alternative clustering methods.
invented entities (3)
  • Visuo-motor primitives independent evidence
    purpose: Reusable atomic behavior units discovered from trajectory segments via clustering.
    The primitives are data-driven clusters, not postulated entities. Their utility is independently validated through downstream task performance and ablation studies (Table 4).
  • Transition interfaces independent evidence
    purpose: Local compositional interfaces between adjacent primitives in a composition pattern, representing how behavior units connect.
    Derived from the primitive sequence, not independently postulated. The ablation removing transition exposure (Table 4, w/o Trans.) shows a 5.5-point drop, providing falsifiable evidence for their importance.
  • Composition patterns independent evidence
    purpose: Ordered sequences of primitives representing the structural form of a trajectory.
    Derived from the primitive vocabulary. Figure 3 shows redistribution of composition patterns after selection, providing observable evidence of the mechanism.

pith-pipeline@v1.1.0-glm · 18331 in / 3079 out tokens · 257424 ms · 2026-07-08T05:35:13.882886+00:00 · methodology

0 comments
read the original abstract

Vision-Language-Action (VLA) models are typically trained by imitation learning on large-scale robot demonstration datasets, but more data does not necessarily yield better policies due to redundancy, noise, and uneven coverage. Existing data selection methods often assess demonstrations at either the trajectory or state-action level, missing the reusable structures that compose long-horizon behaviors. In this paper, we propose SIEVE, a structure-aware data selection method for VLA imitation learning. SIEVE views demonstrations as compositions of reusable primitives and transition interfaces. It first discovers visuo-motor primitives from segmented trajectories, then allocates selection budgets to composition patterns by maximizing reuse-aware structural exposure under diminishing returns. Finally, it selects medoid trajectories within each composition-pattern bucket to retain central, stable, and imitation-friendly demonstrations. Experiments across multiple datasets, benchmarks, and VLA models show that SIEVE consistently outperforms competitive data selection baselines. Notably, SIEVE can surpass full-data training while using only 50% of demonstrations and 50% of training steps, suggesting that reusable structure, captured through primitives and transitions, is an important signal for efficient VLA imitation learning.

Figures

Figures reproduced from arXiv: 2607.06442 by Bin Yu, Changti Wu, Cong Huang, Kai Chen, Lei Zhang, Shijie Lian, Xiaopeng Lin, Zhaolong Shen, Zhirui Zhang.

Figure 1
Figure 1. Figure 1: Motivation of SIEVE. Training demonstrations [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of SIEVE. 1) Primitive Discovery: trajectories are segmented into primitives through representation learning [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Distribution of top-50 composition patterns be [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Selection of the number of primitive clusters [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages · 17 internal anchors

  1. [1]

    Conference on Robot Learning , pages=

    Rt-2: Vision-language-action models transfer web knowledge to robotic control , author=. Conference on Robot Learning , pages=. 2023 , organization=

    work page 2023

  2. [2]

    2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0 , author=. 2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2024 , organization=

    work page 2024

  3. [3]

    OpenVLA: An Open-Source Vision-Language-Action Model

    Openvla: An open-source vision-language-action model , author=. arXiv preprint arXiv:2406.09246 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    The International Journal of Robotics Research , volume=

    Diffusion policy: Visuomotor policy learning via action diffusion , author=. The International Journal of Robotics Research , volume=. 2025 , publisher=

    work page 2025

  5. [5]

    $\pi_0$: A Vision-Language-Action Flow Model for General Robot Control

    _0 : A Vision-Language-Action Flow Model for General Robot Control , author=. arXiv preprint arXiv:2410.24164 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    Intelligence, Physical and Black, Kevin and Brown, Noah and Darpinian, James and Dhabalia, Karan and Driess, Danny and Esmail, Adnan and Equi, Michael and Finn, Chelsea and Fusai, Niccolo and others , journal=. _

    work page

  7. [7]

    Intelligence, Physical and Ai, Bo and Amin, Ali and Aniceto, Raichelle and Balakrishna, Ashwin and Balke, Greg and Black, Kevin and Bokinsky, George and Cao, Shihao and Charbonnier, Thomas and others , journal=. _

    work page

  8. [8]

    Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success

    Fine-tuning vision-language-action models: Optimizing speed and success , author=. arXiv preprint arXiv:2502.19645 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  9. [9]

    GR00T N1: An Open Foundation Model for Generalist Humanoid Robots

    Gr00t n1: An open foundation model for generalist humanoid robots , author=. arXiv preprint arXiv:2503.14734 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    Shortcut Learning in Generalist Robot Policies: The Role of Dataset Diversity and Fragmentation

    Shortcut learning in generalist robot policies: The role of dataset diversity and fragmentation , author=. arXiv preprint arXiv:2508.06426 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    IntentVLA: Short-Horizon Intent Modeling for Aliased Robot Manipulation

    IntentVLA: Short-Horizon Intent Modeling for Aliased Robot Manipulation , author=. arXiv preprint arXiv:2605.14712 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    ACM Computing Surveys (CSUR) , volume=

    Imitation learning: A survey of learning methods , author=. ACM Computing Surveys (CSUR) , volume=. 2017 , publisher=

    work page 2017

  13. [13]

    Proceedings of the thirteenth international conference on artificial intelligence and statistics , pages=

    Efficient reductions for imitation learning , author=. Proceedings of the thirteenth international conference on artificial intelligence and statistics , pages=. 2010 , organization=

    work page 2010

  14. [14]

    arXiv preprint arXiv:2510.18137 , year=

    Quality Over Quantity: Curating Contact-Based Robot Datasets Improves Learning , author=. arXiv preprint arXiv:2510.18137 , year=

    work page arXiv

  15. [15]

    Advances in neural information processing systems , volume=

    Data quality in imitation learning , author=. Advances in neural information processing systems , volume=

    work page

  16. [16]

    International Conference on Learning Representations , volume=

    Data scaling laws in imitation learning for robotic manipulation , author=. International Conference on Learning Representations , volume=

    work page

  17. [17]

    arXiv preprint arXiv:2511.09119 , year=

    Data Assessment for Embodied Intelligence , author=. arXiv preprint arXiv:2511.09119 , year=

    work page arXiv

  18. [18]

    Re-Mix: Optimizing Data Mixtures for Large Scale Imitation Learning

    Re-mix: Optimizing data mixtures for large scale imitation learning , author=. arXiv preprint arXiv:2408.14037 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  19. [19]

    SCIZOR: A Self-Supervised Approach to Data Curation for Large-Scale Imitation Learning

    Scizor: A self-supervised approach to data curation for large-scale imitation learning , author=. arXiv preprint arXiv:2505.22626 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  20. [20]

    Robot Data Curation with Mutual Information Estimators

    Robot data curation with mutual information estimators , author=. arXiv preprint arXiv:2502.08623 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  21. [21]

    DataMIL: Selecting Data for Robot Imitation Learning with Datamodels

    Datamil: Selecting data for robot imitation learning with datamodels , author=. arXiv preprint arXiv:2505.09603 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  22. [22]

    Curating Demonstrations using Online Experience

    Curating demonstrations using online experience , author=. arXiv preprint arXiv:2503.03707 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  23. [23]

    FrameSkip: Learning from Fewer but More Informative Frames in VLA Training

    FrameSkip: Learning from Fewer but More Informative Frames in VLA Training , author=. arXiv preprint arXiv:2605.13757 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  24. [24]

    arXiv preprint arXiv:2606.16208 , year=

    ATHENA: Accelerated Multi-Task Heterogeneous Influence Functions for Robot Data Curation , author=. arXiv preprint arXiv:2606.16208 , year=

    work page arXiv

  25. [25]

    arXiv preprint arXiv:2602.11636 , year=

    ScalSelect: Scalable Training-Free Multimodal Data Selection for Efficient Visual Instruction Tuning , author=. arXiv preprint arXiv:2602.11636 , year=

    work page arXiv

  26. [26]

    2026 , publisher=

    Synthetic Data for Multimodal Large Language Models: A Lifecycle-Oriented Survey , author=. 2026 , publisher=

    work page 2026

  27. [27]

    Encyclopedia of statistical sciences , volume=

    Minimum description length principle , author=. Encyclopedia of statistical sciences , volume=. 2004 , publisher=

    work page 2004

  28. [28]

    2007 , publisher=

    The minimum description length principle , author=. 2007 , publisher=

    work page 2007

  29. [29]

    IEEE transactions on information theory , volume=

    The minimum description length principle in coding and modeling , author=. IEEE transactions on information theory , volume=. 1998 , publisher=

    work page 1998

  30. [30]

    arXiv preprint arXiv:2601.03220 , year=

    From Entropy to Epiplexity: Rethinking Information for Computationally Bounded Intelligence , author=. arXiv preprint arXiv:2601.03220 , year=

    work page arXiv

  31. [31]

    V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning

    V-jepa2: Self-supervised video models enable understanding, prediction and planning , author=. arXiv preprint arXiv:2506.09985 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  32. [32]

    Proceedings of the 19th international conference on World wide web , pages=

    Web-scale k-means clustering , author=. Proceedings of the 19th international conference on World wide web , pages=

    work page

  33. [33]

    Conference on Robot Learning , pages=

    Bridgedata v2: A dataset for robot learning at scale , author=. Conference on Robot Learning , pages=. 2023 , organization=

    work page 2023

  34. [34]

    Evaluating Real-World Robot Manipulation Policies in Simulation

    Evaluating real-world robot manipulation policies in simulation , author=. arXiv preprint arXiv:2405.05941 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  35. [35]

    RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots

    Robocasa: Large-scale simulation of everyday tasks for generalist robots , author=. arXiv preprint arXiv:2406.02523 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  36. [36]

    Qwen3-VL Technical Report

    Qwen3-vl technical report , author=. arXiv preprint arXiv:2511.21631 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  37. [37]

    arXiv preprint arXiv:2512.16793 , year=

    Physbrain: Human egocentric data as a bridge from vision language models to physical intelligence , author=. arXiv preprint arXiv:2512.16793 , year=

    work page arXiv

  38. [38]

    Large VLM-based Vision-Language-Action Models for Robotic Manipulation: A Survey

    Large vlm-based vision-language-action models for robotic manipulation: A survey , author=. arXiv preprint arXiv:2508.13073 , year=

    work page internal anchor Pith review Pith/arXiv arXiv