Pith. sign in

REVIEW 3 major objections 4 minor 10 cited by

MG-Select proposes a verifier-free test-time scaling method for vision-language-action models, using KL divergence from a condition-masking reference distribution to select among sampled actions.

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 · deepseek-v4-flash

2026-08-04 11:15 UTC pith:V6CLMZYC

load-bearing objection A plausible verifier-free test-time scaling idea with real gains, but the headline 'consistent improvement' is weakened by per-dataset tuning on the test set and small real-world samples. the 3 major comments →

arxiv 2510.05681 v2 pith:V6CLMZYC submitted 2025-10-07 cs.RO cs.AIcs.LG

Verifier-free Test-Time Sampling for Vision-Language-Action Models

Suhyeok Jang , Dongyoung Kim , Changyeon Kim , Youngsuk Kim , Jinwoo Shin This is my paper
classification cs.RO cs.AIcs.LG
keywords vision-language-action modelstest-time scalingbest-of-N selectionKL divergenceconfidence metriccondition maskingrobot manipulationimitation learning
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.

The paper argues that the single-greedy-inference paradigm is a key bottleneck for vision-language-action models in precision tasks like grasping, and that this bottleneck can be reduced by test-time scaling that uses only the model's own internal signals. Concretely, it proposes sampling several candidate actions and selecting the one whose predicted distribution diverges most from a reference distribution produced by the same model from randomly masked state and language inputs. The hypothesis is that this KL-divergence confidence identifies the most precise action without needing an externally trained verifier. The authors add a joint training trick that drops conditions during fine-tuning, making the reference distribution more informative. If right, this makes test-time improvement in robot control available at zero extra training cost for any autoregressive VLA.

Core claim

Repeatedly sampling candidate actions and ranking them by KL divergence from a condition-masking reference distribution reliably improves the success rate of vision-language-action models on pick-and-place tasks, in simulation and on a real Franka arm. The reference distribution is produced by the same policy with text, state, or both conditions masked, and is regularized at high temperature; actions that deviate most from this uncertain reference are treated as the most confident and selected. With the proposed joint training that exposes the model to masked conditions, the method reports a 28% relative gain on real in-distribution tasks, 35% on out-of-distribution tasks, and a 168% relativ

What carries the argument

The central object is the condition-masking distributional confidence: a per-token KL divergence C_i = KL(Q_i || P_i) between the VLA's predicted action-token distribution P_i and a reference distribution Q_i obtained from the same VLA with inputs masked (text, state, or both). Token-level confidences are aggregated over a subset of the action sequence (e.g., the first five tokens for the paper's primary tokenizer, average over the sequence for the other architecture). A joint training objective trains the model on all-condition and condition-masked examples so the reference is not an out-of-distribution input. This self-generated reference replaces external value-function verifiers and make

Load-bearing premise

The method assumes that the reference distribution made by masking conditions is still close enough to the task distribution that KL divergence from it is a meaningful uncertainty signal; the paper itself shows that without tuning the regularization temperature (τ=1.0) this reference can be worse than a uniform one.

What would settle it

Take a held-out pick-and-place task, run MG-Select with N=4, and compare the selected action against a randomly chosen candidate. If random selection matches MG-Select's success rate over enough trials, the confidence score carries no selection signal and the claimed mechanism is unsupported. A sharper check: if in a task where the base model already succeeds at a high rate, increasing N does not improve or even hurts, the method's benefit is not driven by confidence.

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

If this is right

  • Any autoregressive VLA, not just ones with value functions, can use sampling plus this self-confidence score to improve precision at test time.
  • The method shrinks the precision gap in low-data regimes: improvement is largest when only tens of demonstrations are available.
  • Because no external verifier is loaded or interacted with, deployment overhead is mostly the cost of parallel sampling, which the single-prefill variant reduces by about 45% at N=4.
  • Condition masking during fine-tuning alone already helps the base policy, even without test-time selection, suggesting the dropout acts as a regularizer.
  • The same KL-vs-reference criterion could be applied to any tokenized action space, making the scoring scheme transferable across tokenizers.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: The method could generalize to other autoregressive policy classes beyond manipulation, such as navigation or dexterous control, wherever actions are tokenized and the model conditions on removable inputs.
  • Editorial inference: A harder test than those reported would be whether the confidence score predicts per-step success in long-horizon tasks; the reported LIBERO-Long gain is small and the first-5-token aggregation suggests the score is most informative about early action tokens.
  • Editorial inference: The dependence on choosing the right masking variant and regularization temperature per dataset (text vs state vs both; τ=4.0 vs 1.0) suggests a possible extension where the reference is ensembled across all masking variants and temperatures instead of hand-picked per benchmark.

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

3 major / 4 minor

Summary. The paper proposes MG-Select, a verifier-free test-time scaling framework for vision-language-action models (VLAs). At inference time it samples N candidate action sequences from an autoregressive VLA and selects the best one using a confidence score defined as the KL divergence between the model's conditional action-token distribution and a 'condition-masking' reference distribution obtained by randomly masking text and/or state inputs. A joint training stage that drops conditions during fine-tuning is introduced to improve the quality of this reference. Experiments on RoboCasa, SIMPLER-WidowX, LIBERO, and a real Franka platform report consistent gains over π0-FAST and OpenVLA baselines, including a 28%/35% improvement on real-world in-distribution/out-of-distribution tasks and a 168% relative gain on RoboCasa pick-and-place with 30 demonstrations.

Significance. If the reported gains are robust, the paper makes a useful contribution: it removes the need for externally trained value functions or verifiers, and the single-prefill deployment strategy reduces latency by 45% at N=4, making test-time scaling more practical for robot pipelines. The ablation study is thorough, covering sampling temperature, number of candidates, masking variants, regularization temperature, aggregation, and the joint-training component. The paper also clearly discloses implementation details and the hyperparameter search, which is commendable for reproducibility. However, the central claim of 'consistent improvement' is weakened by the fact that the hyperparameters were selected on the evaluation datasets themselves, and by several per-suite results that show negligible or negative gains.

major comments (3)
  1. [Appendix A.3, Table 5] The paper states: 'We search for the optimal configuration on each dataset within the following ranges ... and report the best result for each policy.' This search includes the masking variant, regularization temperature, sampling temperature, and N. Table 5(c) shows that the best masking variant differs between task families (text vs state), and Table 5(e) shows that the regularization temperature is load-bearing: at τ=1.0, the condition-masking reference is worse than a uniform-KL baseline (44.3 vs 46.5) and only τ=4.0 produces the headline 48.1 result. Because no held-out validation split is used to select these hyperparameters, the reported 'consistent' improvement over baselines is at least partially attributable to test-set selection. The authors should either use a validation split, freeze hyperparameters a priori, or demonstrate that the main results are insensitive to these choi
  2. [Table 6] The abstract and Section 4 claim that MG-Select 'consistently improves base models.' Table 6 does not fully support this. For OpenVLA, MG-Select* decreases LIBERO-Spatial from 85.2 to 81.7 and LIBERO-Goal from 75.5 to 73.6; for π0-FAST, LIBERO-Goal drops from 95.6 to 94.5 and LIBERO-Spatial is essentially flat (97.4 to 97.2). Only the average scores improve. This directly contradicts the 'consistent improvement' wording and should be addressed, either by softening the claim or by explaining the conditions under which gains are expected.
  3. [Tables 3 and 4] The out-of-distribution real-world claim of a 35% improvement rests on only 16 trials per task (32 total). For 'Pick up Tape,' the improvement is 56.3% to 68.8% (9/16 vs 11/16); for 'Take Cup out of Bowl,' 50% to 75% (8/16 vs 12/16). At these sample sizes the differences are within normal sampling variability, so the 35% gain is not statistically reliable. The in-distribution results use 24 trials per task, which is also modest. We recommend more trials or reporting confidence intervals, and tempering the generalization claim accordingly.
minor comments (4)
  1. [Throughout (especially Sec. 3.1 and Table 5e)] The symbol τ is used for both the sampling temperature and the regularization temperature for the condition-masking distribution. These are distinct hyperparameters with very different roles; using separate symbols (e.g., τ_s and τ_r) would prevent confusion.
  2. [Abstract] The abstract says the reference distribution 'ensur[es] maximum uncertainty,' while Section 3.2 more cautiously describes it as 'uncertain yet not too distant' from the target. The stronger phrasing is not justified by the paper's analysis.
  3. [Table 1] In the 300-demo regime, the all-task success rate with MG-Select (without joint training) is 61.3 vs 61.2 for the base model, i.e., essentially no gain. The authors should acknowledge that the benefit is concentrated in low-data and high-precision settings, and adjust the 'consistently improves' language accordingly.
  4. [Table 5(f)] The aggregation strategy is a significant factor (First-5 works best, while Sum and Avg are worse). The paper offers only a hypothesis about the FAST tokenizer. A brief analysis of how this choice interacts with action chunk horizon/tokenizer would strengthen the paper.

Circularity Check

1 steps flagged

Reported 'consistent improvements' are test-set-selected maxima over masking variant and regularization temperature; the confidence metric itself is self-referential but not circular.

specific steps
  1. fitted input called prediction [Appendix A.3 (Deployment); Table 5(c) and Table 5(e) in Section 4.3]
    "We search for the optimal configuration on each dataset within the following ranges: ... and report the best result for each policy. ... We empirically find that naively using the condition-masking distribution (τ=1.0) as a reference does not work well ... we apply an appropriate high temperature (e.g., τ=4.0) to the condition-masking distribution, which regularizes its concentration and results in superior performance."

    The paper reports that MG-Select 'consistently improves' the base model, but the supporting tables are produced after searching the masking variant and the regularization temperature directly on each evaluation benchmark and taking the best result (Appendix A.3). The ablations show these knobs are decisive: text masking gives 48.1 vs 46.7/46.3 for state/both, and τ=1.0 gives 44.3 (below the 46.5 uniform-KL baseline) while τ=4.0 gives the reported 48.1. Therefore the reported improvement is, by construction, the maximum of the searched configurations on the test set; it is a selected optimum rather than an out-of-sample prediction. With no held-out validation split, the central empirical claim reduces to 'there exists a test-set-selected configuration that improves,' not to a property of th

full rationale

The core derivation of MG-Select is not circular in the self-definitional sense: the confidence score is a KL divergence between two conditionings of the same VLA, but the selection is validated by external task success, not by the score itself. No equation equates the selection criterion with the ground-truth label, and no load-bearing uniqueness theorem or ansatz is imported from the authors' own prior work (the method is evaluated against external baselines and benchmarks). The one substantial circular element is in the empirical claim: Appendix A.3 states that the masking variant and regularization temperature—the knobs that the ablations show determine whether MG-Select beats the uniform-KL baseline—are searched on each dataset and the best result is reported. The headline 'consistently improves' is therefore a test-set-selected optimum, not an out-of-sample prediction. This is a fitted-input-called-prediction pattern, so the score is moderate rather than zero.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claim rests on several domain assumptions about the usefulness of the model's own uncertainty, plus a set of hyperparameters tuned on each benchmark. No new physical entities are introduced.

free parameters (6)
  • Sampling temperature τ = tuned per dataset in {0.1,0.3,0.5,0.7,1.0}; ablation uses 0.5
    Controls sample diversity; selected based on test performance.
  • Number of candidates N = tuned in {4,8}; main results use 4
    Best-of-N sample count; more samples did not always help.
  • Masking variant = per environment: text / state / text+state
    Which condition to mask for the reference distribution; chosen per benchmark.
  • Regularization temperature for reference distribution = tuned in {4,6,8,10,12,14,16}; best often 4.0
    Sharpens/softens the reference distribution; critical for effectiveness.
  • Aggregation strategy for token-level scores = first 5 tokens sum for π0-FAST, average for OpenVLA
    Token-range choice based on tokenizer structure and performance.
  • Dropout rates for joint training = 10% per masking variant
    Fraction of data with each condition-masking; fixed but chosen without sensitivity analysis.
axioms (5)
  • domain assumption Autoregressive VLA action distribution factorizes token-wise
    Standard for the VLA models used, stated in Section 2.
  • domain assumption KL divergence to a condition-masked reference is a meaningful confidence signal
    Core hypothesis of the method; empirically validated but not proved.
  • domain assumption The condition-masking distribution remains aligned with the target task distribution
    Stated in Section 3.2; the paper shows sensitivity to temperature, so this alignment is not automatic.
  • domain assumption Model uncertainty as measured by KL correlates with action correctness
    Underpins the selection criterion; supported only indirectly by experiments.
  • domain assumption Joint training with condition dropout improves reference distribution quality without degrading policy performance
    Shown in ablation (Table 5d) but not analyzed theoretically.

pith-pipeline@v1.3.0-alltime-deepseek · 13649 in / 6829 out tokens · 58488 ms · 2026-08-04T11:15:41.705933+00:00 · methodology

0 comments
read the original abstract

Vision-Language-Action models (VLAs) have demonstrated remarkable performance in robot control. However, they remain fundamentally limited in tasks that require high precision due to their single-inference paradigm. While test-time scaling approaches using external verifiers have shown promise, they require additional training and fail to generalize to unseen conditions. We propose Masking Distribution Guided Selection (MG-Select), a novel test-time scaling framework for VLAs that leverages the model's internal properties without requiring additional training or external modules. Our approach utilizes KL divergence from a reference action token distribution as a confidence metric for selecting the optimal action from multiple candidates. We introduce a reference distribution generated by the same VLA but with randomly masked states and language conditions as inputs, providing action uncertainty while remaining aligned with the target task distribution. Additionally, we propose a joint training strategy that enables the model to learn both conditional and unconditional distributions by applying dropout to state and language conditions, thereby further improving the quality of the reference distribution. Our experiments demonstrate that MG-Select provides a reliable reference for action selection through task-relevant condition masking and consistently improves base models across diverse simulation and real-world benchmarks.

Figures

Figures reproduced from arXiv: 2510.05681 by Changyeon Kim, Dongyoung Kim, Jinwoo Shin, Suhyeok Jang, YoungSuk Kim.

Figure 1
Figure 1. Figure 1: Overview of MG-Select. (1) Autoregressive VLA πθ samples action tokens in parallel from the predicted distribution, while simultaneously computing token-wise KL divergence from the condition-masking distribution to the predicted distribution. (2) Best-of-N selection is then performed using an action confidence score Ca˜ obtained by aggregating these token-wise scores. Our approach. To tackle this problem, … view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative results of MG-Select in real-world pick-and-place tasks. We visualize one of our real-world experiments in the “Box to Bowl” task: (a) grasping an object from the box and (b) releasing it into the bowl. The rollout shows that MG-Select can generate high-precision actions at critical moments for task success, whereas the base policy (π0-FAST-DROID) often struggles at these steps [PITH_FULL_IMAG… view at source ↗
Figure 3
Figure 3. Figure 3: Inference latency comparison on LIBERO-Object (Liu et al., 2023). We com￾pare vanilla MG-Select with its efficient de￾ployment variant using single prefill, based on π0-FAST (Pertsch et al., 2025). Regularization temperature. We empirically find that naively using the condition-masking distribu￾tion (τ = 1.0) as a reference does not work well, as shown in [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 10 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. BOKBO (Best of K Bad Options): Calibrated Abstention for VLA Policies

    cs.LG 2026-05 unverdicted novelty 7.0

    BOKBO is the first conformal abstention method for K-sample VLA policies that supplies finite-sample distribution-free guarantees on executed violation rates, with global and Mondrian per-task variants.

  2. E-TTS: A New Embodied Test-Time Scaling Framework for Robotic Manipulation

    cs.RO 2026-06 unverdicted novelty 6.0

    E-TTS introduces a plug-and-play test-time scaling method for embodied tasks that unifies reasoning-action sampling with history buffers and closed-loop refinement to improve performance on manipulation benchmarks.

  3. Robot Critics that Sweat the Small Stuff

    cs.RO 2026-06 unverdicted novelty 6.0

    Fine-tuning VLMs with pairwise progress supervision from policy rollouts improves fine-grained failure detection and boosts robot manipulation success by 11% real-world and 5.9% in simulation.

  4. VeriSpace: Spatially Grounded Action Verification for Vision-Language-Action Models

    cs.RO 2026-06 unverdicted novelty 6.0

    VeriSpace is a 3D-aware action verifier that improves test-time action selection in VLA models by encoding scenes with visual and geometric information and reasoning over spatial relations and goal progress.

  5. Retrieve-then-Steer: Online Success Memory for Test-Time Adaptation of Generative VLAs

    cs.RO 2026-05 unverdicted novelty 6.0

    Retrieve-then-steer stores successful observation-action segments in memory, retrieves relevant chunks, filters them, and uses an elite prior with confidence-adaptive guidance to steer a flow-matching action sampler f...

  6. Retrieve-then-Steer: Online Success Memory for Test-Time Adaptation of Generative VLAs

    cs.RO 2026-05 unverdicted novelty 6.0

    A retrieve-then-steer method stores successful robot actions in memory and uses them to steer a frozen VLA's flow-matching sampler for better test-time reliability without parameter updates.

  7. Test-Time Perturbation Learning with Delayed Feedback for Vision-Language-Action Models

    cs.CV 2026-04 unverdicted novelty 6.0

    PDF improves VLA success rates on LIBERO and Atari by applying test-time perturbation learning with delayed feedback to correct trajectory overfitting and overconfidence.

  8. Real-Time Execution with Autoregressive Policies

    cs.RO 2026-06 unverdicted novelty 5.0

    Autoregressive VLA policies achieve real-time execution via tokenization horizon adjustment and constrained decoding, outperforming flow-matching policies in speed and performance across simulated and real environments.

  9. Test-Time Scaling in Multimodal Foundation Models: A Comprehensive Survey of Generation and Reasoning

    cs.CV 2026-06 unverdicted novelty 5.0

    A survey of test-time scaling for multimodal foundation models that introduces a three-way taxonomy of sampling, feedback, and search approaches along with applications and benchmarks.

  10. Is the Future Compatible? Diagnosing Dynamic Consistency in World Action Models

    cs.RO 2026-05 unverdicted novelty 5.0

    Action-state consistency in World Action Models distinguishes successful from failed imagined futures and supports value-free selection of better rollouts via consensus among predictions.

Reference graph

Works this paper leans on

19 extracted references · 10 linked inside Pith · cited by 9 Pith papers

  1. [1]

    Paligemma: A versatile 3b vlm for transfer.arXiv preprint arXiv:2407.07726,

    Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer.arXiv preprint arXiv:2407.07726,

    Pith/arXiv arXiv

  2. [4]

    Agibot world colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems.arXiv preprint arXiv:2503.06669,

    Qingwen Bu, Jisong Cai, Li Chen, Xiuqi Cui, Yan Ding, Siyuan Feng, Shenyuan Gao, Xindong He, Xuan Hu, Xu Huang, et al. Agibot world colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems.arXiv preprint arXiv:2503.06669,

    Pith/arXiv arXiv

  3. [6]

    Safe: Multitask failure detection for vision-language-action models.arXiv preprint arXiv:2506.09937,

    Qiao Gu, Yuanliang Ju, Shengxiang Sun, Igor Gilitschenski, Haruki Nishimura, Masha Itkina, and Florian Shkurti. Safe: Multitask failure detection for vision-language-action models.arXiv preprint arXiv:2506.09937,

    arXiv

  4. [7]

    Scalable best-of-n selection for large language models via self-certainty.arXiv preprint arXiv:2502.18581,

    Zhewei Kang, Xuandong Zhao, and Dawn Song. Scalable best-of-n selection for large language models via self-certainty.arXiv preprint arXiv:2502.18581,

    arXiv

  5. [8]

    Droid: A large-scale in-the-wild robot manipulation dataset.arXiv preprint arXiv:2403.12945,

    Alexander Khazatsky, Karl Pertsch, Suraj Nair, Ashwin Balakrishna, Sudeep Dasari, Siddharth Karamcheti, Soroush Nasiriany, Mohan Kumar Srirama, Lawrence Yunliang Chen, Kirsty Ellis, et al. Droid: A large-scale in-the-wild robot manipulation dataset.arXiv preprint arXiv:2403.12945,

    Pith/arXiv arXiv

  6. [9]

    Robomonkey: Scaling test-time sampling and verification for vision-language- action models.arXiv preprint arXiv:2506.17811,

    10 Preprint Jacky Kwok, Christopher Agia, Rohan Sinha, Matt Foutter, Shulu Li, Ion Stoica, Azalia Mirhoseini, and Marco Pavone. Robomonkey: Scaling test-time sampling and verification for vision-language- action models.arXiv preprint arXiv:2506.17811,

    Pith/arXiv arXiv

  7. [10]

    Libero: Benchmarking knowledge transfer for lifelong robot learning.arXiv preprint arXiv:2306.03310,

    Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning.arXiv preprint arXiv:2306.03310,

    Pith/arXiv arXiv

  8. [11]

    Fast: Efficient action tokenization for vision-language-action models.arXiv preprint arXiv:2501.09747,

    Karl Pertsch, Kyle Stachowicz, Brian Ichter, Danny Driess, Suraj Nair, Quan Vuong, Oier Mees, Chelsea Finn, and Sergey Levine. Fast: Efficient action tokenization for vision-language-action models.arXiv preprint arXiv:2501.09747,

    Pith/arXiv arXiv

  9. [12]

    The good, the bad, and the greedy: Evaluation of llms should not ignore non-determinism.arXiv preprint arXiv:2407.10457,

    Yifan Song, Guoyin Wang, Sujian Li, and Bill Yuchen Lin. The good, the bad, and the greedy: Evaluation of llms should not ignore non-determinism.arXiv preprint arXiv:2407.10457,

    Pith/arXiv arXiv

  10. [13]

    Fpc-vla: A vision-language-action framework with a supervisor for failure prediction and correction.arXiv preprint arXiv:2509.04018,

    Yifan Yang, Zhixiang Duan, Tianshi Xie, Fuyu Cao, Pinxi Shen, Peili Song, Piaopiao Jin, Guokang Sun, Shaoqing Xu, Yangwei You, et al. Fpc-vla: A vision-language-action framework with a supervisor for failure prediction and correction.arXiv preprint arXiv:2509.04018,

    arXiv

  11. [14]

    Trustscore: Reference-free evaluation of llm response trustworthiness

    11 Preprint Danna Zheng, Danyang Liu, Mirella Lapata, and Jeff Z Pan. Trustscore: Reference-free evaluation of llm response trustworthiness. InICLR 2024 Workshop on Secure and Trustworthy Large Language Models,

    2024

  12. [15]

    12 Preprint A IMPLEMENTATIONDETAILS A.1 TRAINING ONSIMULATIONDATA Imitation learning.We use two representative autoregressive VLA policies as base models: •π 0-FAST (Pertsch et al., 2025): It uses Paligemma-3B VLM (Beyer et al.,

    2025

  13. [16]

    Table 7: Training setups ofπ 0-FAST for different simulation benchmarks. Configuration RoboCasa SIMPLER-WidowXLIBERO30 demos 100 demos 300 demos Training steps 3k 5k 20k 10k 10k Global batch size 64 64 64 64 32 Action chunk horizon 16 16 16 5 10 • OpenVLA (Kim et al., 2024): It uses Prismatic-7B VLM (Karamcheti et al.,

    2024

  14. [17]

    We use a global batch size of 32 for LIBERO (Liu et al., 2023), while other training configurations follow the official OpenVLA implementation

    as the backbone and is trained on 2 NVIDIA A100 GPUs with LoRA fine-tuning (r= 32 ) from the pre-trained checkpoint. We use a global batch size of 32 for LIBERO (Liu et al., 2023), while other training configurations follow the official OpenVLA implementation. Note that, consistent with the OpenVLA configuration, we train the model separately on each LIBE...

    2023

  15. [18]

    Other configurations are fixed with the AdamW optimizer and a cosine decay schedule, with warmup_steps = 300, peak_lr = 1e-5, decay_lr = 1e-6, and decay_steps= 30,000. Joint imitation learning.We follow the same training configuration as the above imitation learning setup, while additionally applying random dropout of 10%/10%/10% (text / state / both text...

    2024

  16. [19]

    † indicates reproduced performance, and ∗ indicates results with additional joint training before applying our test-time scaling framework

    and the respective original papers (Liu et al., 2025; Qu et al., 2025). † indicates reproduced performance, and ∗ indicates results with additional joint training before applying our test-time scaling framework. Model Spoon on Towel Carrot on Plate Stack Cubes Eggplant in BasketAverage Grasp Success Grasp Success Grasp Success Grasp Success Grasp Success ...

    2025

  17. [2023]

    Codemonkeys: Scaling test-time compute for software engineering.arXiv preprint arXiv:2501.14723,

    Ryan Ehrlich, Bradley Brown, Jordan Juravsky, Ronald Clark, Christopher Ré, and Azalia Mirho- seini. Codemonkeys: Scaling test-time compute for software engineering.arXiv preprint arXiv:2501.14723,

    Pith/arXiv arXiv

  18. [2024]

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

    Johan Bjorck, Fernando Castañeda, Nikita Cherniadev, Xingye Da, Runyu Ding, Linxi Fan, Yu Fang, Dieter Fox, Fengyuan Hu, Spencer Huang, et al. Gr00t n1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734,

    Pith/arXiv arXiv

  19. [2025]

    Large language monkeys: Scaling inference compute with repeated sampling

    Bradley Brown, Jordan Juravsky, Ryan Ehrlich, Ronald Clark, Quoc V Le, Christopher Ré, and Azalia Mirhoseini. Large language monkeys: Scaling inference compute with repeated sampling. arXiv preprint arXiv:2407.21787,

    Pith/arXiv arXiv