Pith. sign in

REVIEW 3 major objections 5 minor 59 references

JoyNexus claims that VLA post-training can be made multi-tenant by keeping a frozen backbone resident and group-batching heterogeneous data, yielding a 28.3% cut in aggregate GPU time.

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

2026-08-01 21:25 UTC pith:27BLP57O

load-bearing objection Useful systems integration for multi-tenant VLA post-training, but the utilization headline is not like-for-like and the efficiency claims need a re-run before being taken as firm. the 3 major comments →

arxiv 2607.16074 v1 pith:27BLP57O submitted 2026-07-17 cs.DC cs.AIcs.SE

JoyNexus: Service-Oriented Multi-Tenant Post-Training for VLA Models

Haoran Sun , Wentao Zhang , Junyang Hua , Hedan Yang , Yongjian Guo , Yifei Zhang , Xiaolong Xiang , Mingxi Luo
show 13 more authors
Jing Long Chen Zhao Chen Zhou Wanting Xu Qiming Yang Hui Zhang Song Wang Xiaodong Bai Shuai Di Xu Chu Xiaotie Deng Yicheng Gong Junwu Xiong
This is my paper
classification cs.DC cs.AIcs.SE
keywords multi-tenant training serviceVLA post-traininggroup batchingfrozen backbone sharingresource utilizationreinforcement learningrobot learning infrastructureGPU scheduling
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.

JoyNexus argues that vision-language-action (VLA) post-training—supervised fine-tuning (SFT), reinforcement learning, and evaluation—does not need dedicated per-tenant GPU clusters. The paper proposes a service where one frozen vision-language backbone stays resident and is shared across tenants, while each tenant keeps private action modules, optimizers, and policy versions. The central mechanism is group batching: heterogeneous tenant batches are normalized to a compatible VLAFeature prefix, concatenated through a single backbone forward pass, then split back to tenant-specific decoders. In a four-tenant simulation this reduces aggregate GPU time by 28.3% (1.39× efficiency) and roughly doubles Training Model Service utilization, and a controlled forward-stage test reports 1.71×–3.24× speedups at eight tenants. A reader would care because it turns bursty, small robot-training jobs from a GPU-idling liability into shareable, schedulable workloads.

Core claim

The paper's central claim is that VLA post-training can be a multi-tenant service in which one resident frozen backbone is shared across tenants. Its discovery is group batching across heterogeneous schemas: tenant batches are canonicalized to a compatible VLAFeature prefix, concatenated through one shared forward pass, then split back to tenant-private action modules and optimizers. In a workload-matched four-tenant simulation (three RL, one SFT), aggregate GPU time falls 28.3% (1.39×), Training Model Service utilization rises from 20.8% to 41.3% (1.99×), and Inference Model Service utilization from 28.1% to 37.5% (1.33×). A controlled forward-stage experiment reports 1.71×–3.24× speedups a

What carries the argument

The load-bearing object is the 'resident base model': a shared, frozen vision-language backbone kept memory-resident in both Training and Inference Model Services, with tenant-specific 'slots' mounting private action modules, optimizer states, and policy versions. The named data object is 'VLAFeature'—a canonical model-facing tensor schema to which heterogeneous tenant batches are normalized, so that compatible requests can be concatenated into one physical batch and processed in a single shared forward pass. The scheduling machinery is the pair of Training Queue and Inference Queue: the Training Queue carries optimization work (RL trajectories or SFT batches) with admission limits and stale

Load-bearing premise

The efficiency numbers rest on the assumption that VLA post-training keeps the vision-language backbone frozen and that heterogeneous tenant data can be canonicalized into a compatible VLAFeature prefix without changing semantics; if a tenant must update the backbone or its schema cannot be aligned (masks, action dimensions, camera slots, state conventions), the shared-forward lever and the measured speedups do not transfer.

What would settle it

Run the same workload-matched comparison with one tenant that updates the backbone, or with two tenants whose schemas cannot be canonicalized without changing action semantics, and measure aggregate GPU time and utilization; if the 28.3% reduction and 1.99× training utilization still appear, the mechanism is not the frozen-prefix grouping the paper claims. A cheaper check: instrument end-to-end training-step wall-clock time in the eight-tenant group-batching test; if grouped full steps are no faster than serial steps, the forward-stage speedup is being swamped by private backward and optimizer

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

If this is right

  • A provider can serve many VLA post-training tenants on one GPU node without replicating the base model for each tenant, shifting the accounting model from exclusive card-hours toward shared service capacity.
  • RL, SFT, and evaluation can interleave on the same resident services, filling otherwise idle actor and inference gaps with offline work.
  • Group-batching benefit grows with tenant count and shrinks with local batch size, so the architecture is best suited to bursty, small-batch workloads.
  • Tenant training semantics remain unchanged by grouping—losses, optimizers, and policy versions stay separate—so the measured efficiency gains do not come at the cost of altered learning curves.
  • The design targets VLA stacks that freeze the backbone and use modular action heads, making the shared forward stage the dominant cost being amortized.

Where Pith is reading between the lines

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

  • Editorial inference: the same frozen-prefix grouping should transfer to other multi-tenant fine-tuning settings where a large frozen encoder feeds small task-specific heads—visual encoders, speech encoders, or shared embedding towers—provided schemas can be canonicalized.
  • Editorial inference: because the reported speedups isolate the shared forward stage, end-to-end training-time gains will be smaller than 1.71×–3.24× whenever tenant-private backward passes and optimizer updates are non-negligible; a natural next experiment is to report full-iteration wall-clock time.
  • Editorial inference: if a future workload requires updating the backbone, the paper's main lever disappears; a testable extension would be selective layer sharing or gradient-checkpointed grouped backward passes, and the utilization numbers would need to be re-measured.
  • Editorial inference: the service boundary suggests a pricing implication the paper leaves implicit—usage-based or priority-aware pricing becomes feasible once utilization is actually multiplexed, but it also needs per-tenant fairness guarantees to prevent one tenant's bursts from starving another.

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 / 5 minor

Summary. JoyNexus is a service-oriented architecture for multi-tenant post-training of vision-language-action (VLA) models. The paper decomposes the post-training stack into a resident shared VLM backbone with tenant-specific action modules, decoupled Training/Inference/Environment services, dual Training/Inference queues, and group batching that canonicalizes compatible tenant batches into a shared 'VLAFeature' prefix so that the expensive backbone forward is executed once for several tenants. SFT, RL, and evaluation workflows are composed from the same service primitives. Evaluation in §5.1 uses a trace-based simulation with four tenants (three RL, one SFT) on one 8-GPU node and claims a 28.3% aggregate GPU-time reduction (1.39×), 1.99× Training Model Service utilization, and 1.33× Inference Model Service utilization against isolated single-tenant execution; §5.2 reports 1.71×–3.24× shared-forward speedups for group batching across 2–8 tenants with small local batches, with per-tenant loss trajectories preserved. The paper positions itself as the first Tinker-style service for a complete VLA post-training ecosystem.

Significance. If the results hold, JoyNexus targets a genuine gap: VLA post-training is heterogeneous (SFT, online/offline RL, evaluation) and bursty, while standard cloud execution gives each tenant exclusive, often idle, resources. The service decomposition, dual-queue scheduling, tenant-private action/optimizer state over a resident shared backbone, and training-side group batching are reasonable design choices grounded in prior systems (Tinker-style APIs, RLinf-VLA/RL-VLA3, and LoRA-style multi-tenant serving). The empirical section deserves credit for honesty: §5.2 confines its speedup metric to the shared-forward stage and explicitly disclaims end-to-end training-throughput gains, and Figure 12 provides a useful per-tenant loss-equivalence check. Machine-checked or reproducible artifacts are not shipped, and the headline efficiency claims rest on a single unseeded trace (§5.1) plus a utilization baseline that is not workload-matched (Table 1). With those corrected, the paper would be a solid systems contribution; as written, the main numbers need revision.

major comments (3)
  1. [§5.1, Table 1] The 1.99× Training Model Service utilization claim is not workload-matched. The multi-tenant row (41.3%) includes the SFT tenant, which fills idle actor intervals, but the isolated row is defined as 'the mean across the three standalone RL traces' and omits the SFT tenant's standalone training utilization. Since SFT is compute-continuous, its isolated utilization is likely well above the 20.8% RL-only mean; adding it to the baseline would raise the denominator and shrink the improvement. Figure 9, by contrast, appends the same SFT work to the isolated GPU-time baseline, so the two comparisons use inconsistent baseline definitions. Please report an SFT-inclusive baseline (or relabel the row as an RL-only baseline) and re-derive the utilization gain; the current 1.99× figure is a headline claim and cannot stand as reported.
  2. [§5.1] The headline 28.3% GPU-time reduction rests on a single trace-based simulation: no repeated runs, no seeds, and no error bars, and the matched comparison window ('first 90 minutes') is chosen after the fact. It is not established that the reduction is robust to the window length, the round-robin admission order, or variability in simulator interaction times. Please add a sensitivity check (e.g., vary the matched window or report the full-trace comparison), or state explicitly that the result is from one representative trace and should be read as a case study rather than a general performance guarantee.
  3. [§4.3 and §5.2] The efficiency gains depend on two preconditions: the VLM backbone is frozen during post-training and heterogeneous tenant schemas can be canonicalized into a compatible 'VLAFeature' without semantic change. These are asserted in §3.1/§4.3 rather than validated; the group-batching experiments cover two model families but group only within a family over LeRobot-standardized LIBERO/CALVIN datasets, so genuinely heterogeneous schemas (masks, action dimensions, camera slots, state conventions) are not exercised. The Section 6 caveat acknowledges this, but the precondition should be stated as an explicit compatibility requirement in §4.3 and the reported speedups scoped to tenants satisfying it; otherwise the title-level claim of heterogeneous multi-tenant post-training overstates what is demonstrated.
minor comments (5)
  1. [Figure 1] Subfigure labels are incomplete ('(d) StarVLA-') and the figure is an adaptation of StarVLA Figure 2; please fix the label and clarify the caption's attribution.
  2. [Figure 3] Diagram contains a typo: 'Schedular' should read 'Scheduler'.
  3. [Abstract and §1] Group batching is described without noting that the speedups in §5.2 are restricted to the shared forward stage; add a qualifier so readers do not infer end-to-end training-throughput gains.
  4. [§5.1] Define how utilization is computed for Table 1 (per-GPU busy fraction over which interval) so the row definitions match Figure 9's baseline, and justify the 1-1-2 isolated layout against the 2-2-4 multi-tenant layout, since the baseline choice directly affects the utilization comparison.
  5. [§2.1] MARLaaS [50] is a closely related multi-tenant RL-as-a-service system; the two-sentence mention should state the concrete differences (VLA-specific service decomposition, canonicalized training-side group batching) to make the novelty claim precise.

Circularity Check

0 steps flagged

No load-bearing circularity: headline quantities are measured on the authors' own prototype, not derived from inputs; self-citations (RL-VLA3/StarVLA) are engineering and design-context dependencies rather than evidence for the claims. The Table 1 utilization baseline asymmetry is a benchmark-validity concern, not a circular step.

full rationale

The paper's headline results are empirical measurements, not derivations, and none reduces to its inputs by construction. The 28.3% GPU-time reduction (1.39x) in §5.1 is a measured difference between the multi-tenant trace and a sequentially replayed isolated baseline; the 1.99x/1.33x utilization ratios in Table 1 are arithmetic ratios of measured utilizations; the 1.71x-3.24x shared-forward speedups in §5.2 are measured ratios of serial vs. grouped pipeline runs. Where a definitional component exists (one shared forward instead of N repeats), the paper acknowledges it as the mechanism being measured rather than a derived prediction, explicitly scoping the claim: 'our efficiency measurements isolate the shared forward stage rather than claiming end-to-end training-throughput improvements' and 'the reported speedup does not include tenant-private backward passes or optimizer updates.' No parameter is fitted and renamed as a prediction, and no equation maps an input into an output claim. Self-citations are present — RL-VLA3 [44] (first author Haoran Sun overlaps) is the integrated 'local VLA RL stack [44, 49, 51]', and StarVLA [43] is the structural basis for the base-model/action-module split — but these are engineering dependencies and design-context citations; the load-bearing frozen-backbone assumption is stated as the paper's own observation (§3.1: 'We also observe that VLA post-training often keeps the base VLM fixed') and is explicitly bounded by the §6 limitation that canonicalization 'can enable sharing, but only when masks, action dimensions, camera slots, and state conventions preserve the semantics required by each tenant.' No uniqueness theorem is imported from prior work. One correctness concern, not circularity: the Table 1 isolated training-utilization baseline 'is the mean across the three standalone RL traces' and excludes the SFT tenant, while the multi-tenant row includes SFT batches filling actor idle intervals; the matched GPU-time comparison (Figure 9) 'append[s] the same SFT work' to the baseline, so the 1.99x training-utilization headline is not like-for-like and likely overstates the scheduling-only gain. This is a benchmark-validity flaw, not a reduction of a claim to its inputs.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

No new physical/mathematical entities are introduced; JoyNexus's services are software components, not invented entities. The claimed efficiency depends on frozen-backbone VLA design, schema canonicalizability, and the favorable small-batch regime.

free parameters (5)
  • staleness bound Δ_max = 128 actor microbatch updates
    Chosen in §5.1 to bound rollout lag; no sensitivity analysis, affects RL update behavior.
  • global actor microbatch / per-GPU microbatch = 256 global, 128 per GPU
    Experiment config in §5.1; directly affects training throughput and overlap.
  • in-flight and rollout limits = ≤2 training jobs per RL tenant, ≤3 active rollouts globally, 16 SFT prefetch
    Admission-control settings in §5.1; chosen without sensitivity analysis.
  • inference batch cap and waiting-time budget = 128 samples; T not specified
    Scheduler knobs in §4.3/§5.1; affect group-batching gains but T is not reported.
  • resource layout = 2 actor GPUs / 2 inference GPUs / 4 environment GPUs; baseline 1/1/2
    Deployment choice in §5.1; the claimed 1.39× is relative to this specific layout.
axioms (4)
  • domain assumption VLA post-training typically freezes the shared vision-language backbone; only action modules are trainable.
    Stated in §3.1 as an observation and used to justify resident shared base and group batching.
  • domain assumption Most VLA models adopt a unified data prototype, so heterogeneous tenant batches can be canonicalized and concatenated at the VLAFeature boundary.
    Assumed in §4.3 and Appendix A.3; limitations admit schema compatibility is not always safe.
  • domain assumption Shared backbone forward dominates tenant-private action-module work.
    Measurements in §5.2 show it for tested models; used to argue group batching gives meaningful speedup.
  • ad hoc to paper Single-tenant execution with 1 actor, 1 inference, and 2 environment GPUs is a representative baseline.
    Used to compute the 1.39× claim; no comparison to alternative layouts or existing multi-tenant systems.

pith-pipeline@v1.3.0-alltime-deepseek · 20099 in / 14152 out tokens · 128868 ms · 2026-08-01T21:25:14.554276+00:00 · methodology

0 comments
read the original abstract

The post-training of Vision-Language-Action (VLA) models is essential due to the diversity of simulators, robot embodiments, and task objectives. Existing compute services, whether offered as direct accelerator rental or batch-workload submission, typically allocate an exclusive set of GPU and CPU resources to a single tenant. While this paradigm maximizes client flexibility, it burdens users with infrastructure adaptation, and the fixed card-hour accounting model renders short or bursty workloads both expensive for tenants and inefficient for the service provider. To address these challenges, we present JoyNexus, a unified service for multi-tenant VLA supervised fine-tuning, reinforcement learning, and evaluation. JoyNexus decouples the Training Model Service, Inference Model Service, and Environment Service, each accessed through APIs and backed by resident shared base models with tenant-specific slots. Tenants can directly invoke high-level semantic APIs for training, rollout, and evaluation, or compose custom algorithms using lower-level APIs and their assigned endpoints. Multiple tenants submit workloads concurrently; their action modules, optimizers, rollout records, and policy versions remain isolated, and the service is scheduled by the global Training Queue and Inference Queue. To further improve multi-tenant training efficiency, JoyNexus introduces group batching for heterogeneous VLA data schemas that share a compatible model-facing prefix, enabling a single shared backbone forward pass over grouped samples. Finally, we evaluate JoyNexus through workload simulation and a group-batching pipeline in a realistic embodied scenario. Results show that, compared with isolated single-tenant execution, JoyNexus reduces aggregate GPU time and improves service utilization via cross-tenant scheduling on shared resources.

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

59 extracted references · 25 linked inside Pith

  1. [1]

    Gulavani, and Ramachandran Ramjee

    Amey Agrawal, Ashish Panwar, Jayashree Mohan, Nipun Kwatra, Bhargav S. Gulavani, and Ramachandran Ramjee. SARATHI: Efficient LLM inference by piggybacking decodes with chunked prefills.arXiv preprint arXiv:2308.16369, 2023

    Pith/arXiv arXiv 2023

  2. [2]

    Above the clouds: A berkeley view of cloud computing

    Michael Armbrust, Armando Fox, Rean Griffith, Anthony D Joseph, Randy H Katz, Andrew Konwinski, Gunho Lee, David A Patterson, Ariel Rabkin, Ion Stoica, et al. Above the clouds: A berkeley view of cloud computing. Technical report, Technical Report UCB/EECS-2009-28, EECS Department, University of California ..., 2009

    2009

  3. [3]

    Experiment tracking with weights and biases, 2020.https://www.wandb.com/

    Lukas Biewald. Experiment tracking with weights and biases, 2020.https://www.wandb.com/. Software available from wandb.com

    2020

  4. [4]

    Gr00t n1: An open foundation model for generalist humanoid robots

    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, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    arXiv preprint arXiv:2410.24164, 2024

    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. arXiv preprint arXiv:2410.24164, 2024

    Pith/arXiv arXiv 2024

  6. [6]

    RT-1: Robotics transformer for real-world control at scale.arXiv preprint arXiv:2212.06817, 2022

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. RT-1: Robotics transformer for real-world control at scale.arXiv preprint arXiv:2212.06817, 2022

    Pith/arXiv arXiv 2022

  7. [7]

    Lerobot: State-of-the-art machine learning for real-world robotics in pytorch.https://github.com/huggingface/lerobot, 2024

    Remi Cadene, Simon Alibert, Alexander Soare, Quentin Gallouedec, Adil Zouitine, Steven Palma, Pepijn Kooij- mans, Michel Aractingi, Mustafa Shukor, Dana Aubakirova, Martino Russi, Francesco Capuano, Caroline Pascal, Jade Choghari, Khalil Meftah, Maxime Ellerbach, Jess Moss, and Thomas Wolf. Lerobot: State-of-the-art machine learning for real-world robotic...

    2024

  8. [8]

    Punica: Multi-tenant LoRA serving.arXiv preprint arXiv:2310.18547, 2023

    Lequn Chen, Zihao Ye, Yongji Wu, Danyang Zhuo, Luis Ceze, and Arvind Krishnamurthy. Punica: Multi-tenant LoRA serving.arXiv preprint arXiv:2310.18547, 2023

    Pith/arXiv arXiv 2023

  9. [9]

    Diffusion policy: Visuomotor policy learning via action diffusion.The International Journal of Robotics Research, 44(10-11):1684–1704, 2025

    Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion.The International Journal of Robotics Research, 44(10-11):1684–1704, 2025

    2025

  10. [10]

    Clearml - your entire mlops stack in one open-source tool, 2024.https://clear.ml/

    ClearML. Clearml - your entire mlops stack in one open-source tool, 2024.https://clear.ml/. Software available from http://github.com/clearml/clearml

    2024

  11. [11]

    Relax: An asynchronous reinforcement learning engine for omni-modal post-training at scale, 2026.https://arxiv.org/abs/2604.11554

    Relax Contributors. Relax: An asynchronous reinforcement learning engine for omni-modal post-training at scale, 2026.https://arxiv.org/abs/2604.11554

    Pith/arXiv arXiv 2026

  12. [12]

    Queueing theory

    Robert B Cooper. Queueing theory. InProceedings of the ACM’81 conference, pages 119–122, 1981

    1981

  13. [13]

    Franklin, Joseph E

    Daniel Crankshaw, Xin Wang, Guilio Zhou, Michael J. Franklin, Joseph E. Gonzalez, and Ion Stoica. Clipper: A low-latency online prediction serving system. InUSENIX Symposium on Networked Systems Design and Implementation, pages 613–627, 2017

    2017

  14. [14]

    Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. PaLM-E: An embodied multimodal language model. In International Conference on Machine Learning, 2023

    2023

  15. [15]

    AReaL: A large-scale asynchronous reinforcement learning system for language reasoning

    Wei Fu, Jiaxuan Gao, Xujie Shen, Chen Zhu, Zhiyu Mei, Chuyi He, Shusheng Xu, Guo Wei, Jun Mei, Jiashu Wang, Tongkai Yang, Binhang Yuan, and Yi Wu. AReaL: A large-scale asynchronous reinforcement learning system for language reasoning. InAdvances in Neural Information Processing Systems, volume 38, 2025

    2025

  16. [16]

    Maniskill2: A unified benchmark for generalizable manipulation skills.arXiv preprint arXiv:2302.04659, 2023

    Jiayuan Gu, Fanbo Xiang, Xuanlin Li, Zhan Ling, Xiqiang Liu, Tongzhou Mu, Yihe Tang, Stone Tao, Xinyue Wei, Yunchao Yao, et al. Maniskill2: A unified benchmark for generalizable manipulation skills.arXiv preprint arXiv:2302.04659, 2023

    Pith/arXiv arXiv 2023

  17. [17]

    OpenRLHF: An easy-to-use, scalable and high-performance RLHF framework.arXiv preprint arXiv:2405.11143, 2024

    Jian Hu, Xibin Wu, Zilin Zhu, Xianyu, Weixun Wang, Dehao Zhang, and Yu Cao. OpenRLHF: An easy-to-use, scalable and high-performance RLHF framework.arXiv preprint arXiv:2405.11143, 2024. 16

    Pith/arXiv arXiv 2024

  18. [18]

    Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Es- mail, Michael Equi, Chelsea Finn, Niccolo Fusai, et al.π0.5: a vision-language-action model with open-world generalization.arXiv preprint arXiv:2504.16054, 2025

    Pith/arXiv arXiv 2025

  19. [19]

    Rlbench: Therobotlearningbenchmark & learning environment.IEEE Robotics and Automation Letters, 5(2):3019–3026, 2020

    StephenJames, ZicongMa, DavidRovickArrojo, andAndrewJDavison. Rlbench: Therobotlearningbenchmark & learning environment.IEEE Robotics and Automation Letters, 5(2):3019–3026, 2020

    2020

  20. [20]

    Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024

    Pith/arXiv arXiv 2024

  21. [21]

    Fine-tuning vision-language-action models: Optimizing speed and success.arXiv preprint arXiv:2502.19645, 2025

    Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-tuning vision-language-action models: Optimizing speed and success.arXiv preprint arXiv:2502.19645, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    Mira murati’s stealth AI lab launches its first product

    Will Knight. Mira murati’s stealth AI lab launches its first product. WIRED, October 2025.https://www.wired. com/story/thinking-machines-lab-first-product-fine-tune

    2025

  23. [23]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with PagedAttention. InACM Symposium on Operating Systems Principles, pages 611–626, 2023

    2023

  24. [24]

    Mint: Managed infrastructure for training and serving millions of llms.arXiv preprint arXiv:2605.13779, 2026

    Mind Lab, Song Cao, Vic Cao, Andrew Chen, Kaijie Chen, Cleon Cheng, Steven Chiang, Kaixuan Fan, Hera Feng, Huan Feng, et al. Mint: Managed infrastructure for training and serving millions of llms.arXiv preprint arXiv:2605.13779, 2026

    Pith/arXiv arXiv 2026

  25. [25]

    Tinker, 2026.https://thinkingmachines.ai/tinker/

    Thinking Machines Lab. Tinker, 2026.https://thinkingmachines.ai/tinker/

    2026

  26. [26]

    Simplevla-rl: Scaling vla training via reinforcement learning.arXiv preprint arXiv:2509.09674, 2025

    Haozhan Li, Yuxin Zuo, Jiale Yu, Yuhao Zhang, Zhaohui Yang, Kaiyan Zhang, Xuekai Zhu, Yuchen Zhang, Tianxing Chen, Ganqu Cui, et al. Simplevla-rl: Scaling vla training via reinforcement learning.arXiv preprint arXiv:2509.09674, 2025

    Pith/arXiv arXiv 2025

  27. [27]

    VLA-RFT: Vision-language-action reinforcement fine-tuning with verified rewards in world simulators.arXiv preprint arXiv:2510.00406, 2025

    Hengtao Li, Pengxiang Ding, Runze Suo, Yihao Wang, Zirui Ge, Dongyuan Zang, Kexian Yu, Mingyang Sun, Hongyin Zhang, Donglin Wang, and Weihua Su. VLA-RFT: Vision-language-action reinforcement fine-tuning with verified rewards in world simulators.arXiv preprint arXiv:2510.00406, 2025

    arXiv 2025

  28. [28]

    Colossal-AI:Aunifieddeeplearningsystemforlarge-scaleparalleltraining.arXiv preprint arXiv:2110.14883, 2021

    Shenggui Li, Hongxin Liu, Zhengda Bian, Jiarui Fang, Haichen Huang, Yuliang Liu, Boxiang Wang, and Yang You. Colossal-AI:Aunifieddeeplearningsystemforlarge-scaleparalleltraining.arXiv preprint arXiv:2110.14883, 2021

    Pith/arXiv arXiv 2021

  29. [29]

    Gonzalez, and Ion Stoica

    Zhuohan Li, Lianmin Zheng, Yinmin Zhong, Vincent Liu, Ying Sheng, Xin Jin, Joseph E. Gonzalez, and Ion Stoica. AlpaServe: Statistical multiplexing with model parallelism for deep learning serving.arXiv preprint arXiv:2302.11665, 2023

    Pith/arXiv arXiv 2023

  30. [30]

    Gonzalez, Michael I

    Eric Liang, Richard Liaw, Philipp Moritz, Robert Nishihara, Roy Fox, Ken Goldberg, Joseph E. Gonzalez, Michael I. Jordan, and Ion Stoica. RLlib: Abstractions for distributed reinforcement learning. InInternational Conference on Machine Learning, pages 3053–3062. PMLR, 2018

    2018

  31. [31]

    LobRA: Multi-tenant fine-tuning over heterogeneous data.Proceedings of the VLDB Endowment, 18(8):2616–2625, 2025

    Sheng Lin, Fangcheng Fu, Haoyang Li, Hao Ge, Xuanyu Wang, Jiawen Niu, Yaofeng Tu, and Bin Cui. LobRA: Multi-tenant fine-tuning over heterogeneous data.Proceedings of the VLDB Endowment, 18(8):2616–2625, 2025. doi: 10.14778/3742728.3742752

    arXiv 2025

  32. [32]

    Libero: Benchmarking knowledge transfer for lifelong robot learning.Advances in Neural Information Processing Systems, 36:44776– 44791, 2023

    Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning.Advances in Neural Information Processing Systems, 36:44776– 44791, 2023

    2023

  33. [33]

    Calvin: A benchmark for language- conditioned policy learning for long-horizon robot manipulation tasks.IEEE Robotics and Automation Letters, 7(3):7327–7334, 2022

    Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wolfram Burgard. Calvin: A benchmark for language- conditioned policy learning for long-horizon robot manipulation tasks.IEEE Robotics and Automation Letters, 7(3):7327–7334, 2022

    2022

  34. [34]

    Twinkle: Trainingworkbenchtomakeyourmodelglow.https://github.com/modelscope/twinkle,

    ModelScope. Twinkle: Trainingworkbenchtomakeyourmodelglow.https://github.com/modelscope/twinkle,

  35. [35]

    Ray: A distributed framework for emerging AI applications

    Philipp Moritz, Robert Nishihara, Stephanie Wang, Alexey Tumanov, Richard Liaw, Eric Liang, Melih Elibol, Zongheng Yang, William Paul, Michael I Jordan, and Ion Stoica. Ray: A distributed framework for emerging AI applications. In13th USENIX Symposium on Operating Systems Design and Implementation, pages 561–577, 2018. 17

    2018

  36. [36]

    Octo: An open-source generalist robot policy.arXiv preprint arXiv:2405.12213, 2024

    Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, Jianlan Luo, You Liang Tan, Lawrence Yunliang Chen, Pannag Sanketi, Quan Vuong, Ted Xiao, Dorsa Sadigh, Chelsea Finn, and Sergey Levine. Octo: An open-source generalist robot policy.arXiv preprint arXiv:2405.12213, 2024

    Pith/arXiv arXiv 2024

  37. [37]

    Open X-embodiment: Robotic learning datasets and RT-X models.arXiv preprint arXiv:2310.08864, 2023

    Open X-Embodiment Collaboration, Abby O’Neill, Abdul Rehman, Abhinav Gupta, Abhiram Maddukuri, Ab- hishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, et al. Open X-embodiment: Robotic learning datasets and RT-X models.arXiv preprint arXiv:2310.08864, 2023

    Pith/arXiv arXiv 2023

  38. [38]

    ZeRO: Memory optimizations toward train- ing trillion parameter models

    Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He. ZeRO: Memory optimizations toward train- ing trillion parameter models. InSC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1–16. IEEE, 2020

    2020

  39. [39]

    Rlds: an ecosystem to generate, share and use datasets in reinforcement learning.arXiv preprint arXiv:2111.02767, 2021

    Sabela Ramos, Sertan Girgin, Léonard Hussenot, Damien Vincent, Hanna Yakubovich, Daniel Toyama, Anita Gergely, Piotr Stanczyk, Raphael Marinier, Jeremiah Harmsen, et al. Rlds: an ecosystem to generate, share and use datasets in reinforcement learning.arXiv preprint arXiv:2111.02767, 2021

    Pith/arXiv arXiv 2021

  40. [40]

    Hybridflow: A flexible and efficient rlhf framework.arXiv preprint arXiv:2409.19256, 2024

    Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. Hybridflow: A flexible and efficient rlhf framework.arXiv preprint arXiv:2409.19256, 2024

    Pith/arXiv arXiv 2024

  41. [41]

    Gonzalez, and Ion Stoica

    Ying Sheng, Shiyi Cao, Dacheng Li, Coleman Hooper, Nicholas Lee, Shuo Yang, Christopher Chou, Banghua Zhu, Lianmin Zheng, Kurt Keutzer, Joseph E. Gonzalez, and Ion Stoica. S-LoRA: Serving thousands of concurrent LoRA adapters.arXiv preprint arXiv:2311.03285, 2023

    Pith/arXiv arXiv 2023

  42. [42]

    Megatron-LM: Training multi-billion parameter language models using model parallelism.arXiv preprint arXiv:1909.08053, 2019

    Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan Catanzaro. Megatron-LM: Training multi-billion parameter language models using model parallelism.arXiv preprint arXiv:1909.08053, 2019

    Pith/arXiv arXiv 1909

  43. [43]

    StarVLA: A lego-like codebase for vision-language-action model developing.arXiv preprint arXiv:2604.05014, 2026

    StarVLA Community. StarVLA: A lego-like codebase for vision-language-action model developing.arXiv preprint arXiv:2604.05014, 2026

    Pith/arXiv arXiv 2026

  44. [44]

    Rl-vla3: A flexible and asynchronous reinforcement learning framework for vla training

    Haoran Sun, Yongjian Guo, Zhong Guan, Shuai Di, Xiaodong Bai, Jing Long, Tianyun Zhao, Mingxi Luo, Hongke Zhao, Likang Wu, et al. Rl-vla3: A flexible and asynchronous reinforcement learning framework for vla training. arXiv e-prints, pages arXiv–2602, 2026

    2026

  45. [45]

    Large- scale cluster management at google with borg

    Abhishek Verma, Luis Pedrosa, Madhukar Korupolu, David Oppenheimer, Eric Tune, and John Wilkes. Large- scale cluster management at google with borg. InProceedings of the tenth european conference on computer systems, pages 1–17, 2015

    2015

  46. [46]

    Scaling proprioceptive-visual learning with heteroge- neous pre-trained transformers

    Lirui Wang, Xinlei Chen, Jialiang Zhao, and Kaiming He. Scaling proprioceptive-visual learning with heteroge- neous pre-trained transformers. InAdvances in Neural Information Processing Systems, volume 37, 2024. doi: 10.52202/079017-3952

    doi:10.52202/079017-3952 2024

  47. [47]

    In18th USENIX Symposium on Operating Systems Design and Implementation (OSDI 24), pages 911–927, 2024

    Bingyang Wu, Ruidong Zhu, Zili Zhang, Peng Sun, Xuanzhe Liu, and Xin Jin.{dLoRA}: Dynamically orches- trating requests and adapters for{LoRA}{LLM}serving. In18th USENIX Symposium on Operating Systems Design and Implementation (OSDI 24), pages 911–927, 2024

    2024

  48. [48]

    Deepspeed-chat: Easy, fast and affordable rlhf training of chatgpt-like models at all scales.arXiv preprint arXiv:2308.01320, 2023

    Zhewei Yao, Reza Yazdani Aminabadi, Olatunji Ruwase, Samyam Rajbhandari, Xiaoxia Wu, Ammar Ahmad Awan, Jeff Rasley, Minjia Zhang, Conglong Li, Connor Holmes, et al. Deepspeed-chat: Easy, fast and affordable rlhf training of chatgpt-like models at all scales.arXiv preprint arXiv:2308.01320, 2023

    Pith/arXiv arXiv 2023

  49. [49]

    Rlinf: Flexible and efficient large-scale reinforcement learning via macro-to-micro flow transformation.arXiv preprint arXiv:2509.15965, 2025

    Chao Yu, Yuanqing Wang, Zhen Guo, Hao Lin, Si Xu, Hongzhi Zang, Quanlu Zhang, Yongji Wu, Chunyang Zhu, Junhao Hu, et al. Rlinf: Flexible and efficient large-scale reinforcement learning via macro-to-micro flow transformation.arXiv preprint arXiv:2509.15965, 2025

    arXiv 2025

  50. [50]

    MARLaaS: Multi-tenant asynchronous reinforcement learning as a service.arXiv preprint arXiv:2605.08527, 2026

    Timothy Tin Long Yu, Gursimran Singh, Ge Shi, Hanieh Sadri, Yong Zhang, and Zhenan Fan. MARLaaS: Multi-tenant asynchronous reinforcement learning as a service.arXiv preprint arXiv:2605.08527, 2026

    Pith/arXiv arXiv 2026

  51. [51]

    Rlinf-vla: A unified and efficient framework for vla+ rl training.arXiv preprint arXiv:2510.06710, 2025

    Hongzhi Zang, Mingjie Wei, Si Xu, Yongji Wu, Zhen Guo, Yuanqing Wang, Hao Lin, Liangzhi Shi, Yuqing Xie, Zhexuan Xu, et al. Rlinf-vla: A unified and efficient framework for vla+ rl training.arXiv preprint arXiv:2510.06710, 2025

    arXiv 2025

  52. [52]

    A vision-language-action-critic model for robotic real-world reinforcement learning.arXiv preprint arXiv:2509.15937, 2025

    Shaopeng Zhai, Qi Zhang, Tianyi Zhang, Fuxian Huang, Haoran Zhang, Ming Zhou, Shengzhe Zhang, Litao Liu, Sixu Lin, and Jiangmiao Pang. A vision-language-action-critic model for robotic real-world reinforcement learning.arXiv preprint arXiv:2509.15937, 2025. 18

    arXiv 2025

  53. [53]

    Prorl agent: Rollout-as-a-service for rl training of multi-turn llm agents.arXiv preprint arXiv:2603.18815, 2026

    Hao Zhang, Mingjie Liu, Shaokun Zhang, Songyang Han, Jian Hu, Zhenghui Jin, Yuchi Zhang, Shizhe Diao, Ximing Lu, Binfeng Xu, et al. Prorl agent: Rollout-as-a-service for rl training of multi-turn llm agents.arXiv preprint arXiv:2603.18815, 2026

    arXiv 2026

  54. [54]

    X-VLA: Soft- prompted transformer as scalable cross-embodiment vision-language-action model

    Jinliang Zheng, Jianxiong Li, Zhihao Wang, Dongxiu Liu, Xirui Kang, Yuchun Feng, Yinan Zheng, Jiayin Zou, Yilun Chen, Jia Zeng, Ya-Qin Zhang, Jiangmiao Pang, Jingjing Liu, Tai Wang, and Xianyuan Zhan. X-VLA: Soft- prompted transformer as scalable cross-embodiment vision-language-action model. InInternational Conference on Learning Representations, 2026

    2026

  55. [55]

    Sglang: Efficient execution of structured language model programs.Advances in neural information processing systems, 37:62557–62583, 2024

    Lianmin Zheng, Liangsheng Yin, Zhiqiang Xie, Chuyue Sun, Jeff Huang, Cody H Yu, Shiyi Cao, Christos Kozyrakis, Ion Stoica, Joseph E Gonzalez, et al. Sglang: Efficient execution of structured language model programs.Advances in neural information processing systems, 37:62557–62583, 2024

    2024

  56. [56]

    Opentinker: Separating concerns in agentic reinforcement learning, 2026.https: //arxiv.org/abs/2601.07376

    Siqi Zhu and Jiaxuan You. Opentinker: Separating concerns in agentic reinforcement learning, 2026.https: //arxiv.org/abs/2601.07376

    Pith/arXiv arXiv 2026

  57. [57]

    slime: An llm post-training framework for rl scaling

    Zilin Zhu, Chengxing Xie, Xin Lv, and slime Contributors. slime: An llm post-training framework for rl scaling. https://github.com/THUDM/slime, 2025. GitHub repository. Corresponding author: Xin Lv

    2025

  58. [58]

    rl"), domain=

    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. In Conference on Robot Learning, pages 2165–2183. PMLR, 2023. A Composition of VLA Workloads This appendix provides an implementation-oriented expa...

    2023

  59. [2026]

    GitHub repository; accessed 2026-06-26

    2026