Pith. sign in

REVIEW 3 major objections 6 minor 58 references

A language-only semantic source, injected at the model's fusion point with a retrained action head, restores paraphrase robustness without any paraphrase training.

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

2026-08-04 06:05 UTC pith:CFNZ3JVJ

load-bearing objection Useful diagnostic work and a credible intervention that is sold with a causal story the evidence only supports for VLA-Adapter; worth refereeing if the claims are scoped down. the 3 major comments →

arxiv 2608.02497 v1 pith:CFNZ3JVJ submitted 2026-08-03 cs.RO

Grounded Semantic Re-Binding for Robust Instruction Generalization in Vision-Language-Action Models

Zhaokai Yin , Zhipeng Zhang This is my paper
classification cs.RO
keywords vision-language-action modelsparaphrastic invarianceinstruction followingrobotic manipulationsemantic re-bindingaction policy retraininglanguage groundingdecoupled architecture
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.

This paper tries to prove that vision-language-action models fail on reworded instructions not because they misunderstand them, but because their architecture routes task information through a joint image-and-text encoding whose small systematic shifts the action policy cannot absorb. It establishes this by probing three models, showing the correct task identity survives internally, and by a causal swap that fixes execution when canonical features replace paraphrased features at the last attention block. It then proposes Grounded Semantic Re-binding: extract instruction semantics with a frozen pure-language encoder, inject them into the model's native visual computation, and train a fresh action head from scratch using only canonical demonstrations. On a full paraphrase benchmark this raises success from about 47% to 71% on one model and from 4% to 49% on another, and a deliberately decoupled 0.33B-parameter model shows near-zero gap between canonical and paraphrased success. The payoff is that robustness to rewording can be bought with architecture rather than with expensive paraphrase data scaling.

Core claim

The paper's central claim is that the bottleneck is in the joint vision-language encoding and its downstream use, not in semantic understanding. Evidence: probing shows paraphrase outputs are closer to canonical than to wrong-task actions; replacing the paraphrased language feature with its canonical counterpart in the final attention block lifts paired success from 60% to 96%; and removing a learned 32-dimensional wording-shift direction cuts the canonical-paraphrase action gap roughly in half. The proposed fix is GSR: a frozen pure-language encoder (T5) supplies task semantics, the model's own visual pathway provides scene grounding, and a completely re-initialized action expert learns to

What carries the argument

Grounded Semantic Re-binding (GSR): the instruction is encoded by a frozen pure-language transformer independent of the current image; the resulting token features are projected and injected at the point where the model already fuses vision, language, and robot state; and the action expert is re-initialized and trained from scratch on canonical demonstrations. For fragile native language routes, the original wording is replaced by a fixed neutral sentence so the action policy is forced to use the stable semantic source; for robust routes the native instruction is kept as a complementary signal. The paper's diagnostic engine is the paired token swap in the final attention block, which convert

Load-bearing premise

The load-bearing premise is that the systematic shift in joint vision-language features, identified by a single-block swap in one model, is the common failure mechanism across all VLA families; the paper's own layerwise probes show task information flows in a distributed way in other architectures, so if those models fail for a different structural reason, the root-cause diagnosis and the GSR recipe derived from it do not generalize as stated.

What would settle it

Run the same canonical-feature swap at scored layers of a VLA whose paraphrase failures are already strong; if no single-layer or few-layer replacement restores a large fraction of the action gap, or if a model with fully decoupled language and vision encoding still shows a large canonical-paraphrase success gap, the joint-encoding bottleneck is not the universal cause. The paper itself reports best single-layer recovery of 10.5% and 31.3% for two architectures, so testing whether those failures also vanish when all paraphrase-altered directions are removed would settle whether the mechanism i

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

Share X LinkedIn Reddit HN

If this is right

  • Robustness to instruction rewording can be restored without any paraphrase data, so the expensive data-scaling route is not the only corrective.
  • Canonical-only trained lightweight policies can be brought close to much larger models on paraphrase success, changing cost/performance tradeoffs.
  • Even a heavily scaled, already-robust policy improves under GSR, suggesting the architectural fix complements rather than substitutes for scale.
  • The correct injection point and whether to neutralize native text depend on the architecture, so the method is a guideline rather than a single patch.
  • A natively decoupled design, with language and vision separate until the action expert, nearly eliminates the canonical-versus-paraphrase gap at small scale.

Where Pith is reading between the lines

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

  • If the canonical-paraphrase shift is a low-dimensional, language-only direction in feature space, then inference-time interventions (subtracting that direction, or steering with a frozen language encoder) might fix many failed rollouts without retraining.
  • The paper's cross-model layerwise analysis shows no single bottleneck in two of the three models; a cautious extension is that GSR may succeed via different mechanisms per architecture, and 'root cause' should be read as model-specific unless further evidence appears.
  • The scaling bottleneck reported for the decoupled design suggests a possible hybrid: use early decoupling for semantic stability and later re-integration for visual scaling.

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 / 6 minor

Summary. The paper proposes Grounded Semantic Re-binding (GSR), an architectural intervention for vision-language-action (VLA) models that aims to restore paraphrastic invariance without any paraphrase training. The authors argue that VLAs internally preserve the correct task identity under paraphrased instructions and that failures instead arise from a downstream action policy mishandling systematic feature shifts induced by joint vision-language encoding. GSR extracts task semantics with a frozen T5 encoder, injects them into the model's native multimodal computation, and re-trains a re-initialized action expert on canonical demonstrations only. Experiments on LIBERO-Para report large improvements for VLA-Adapter (46.82% to 70.94% full-Para success) and SmolVLA (4.47% to 49.12%), a smaller gain for π0.5 (73.60% to 75.59%), and a new PRIDE score of 70.4. The paper also presents ParaVLA, a 0.33B natively decoupled model with 92.0% canonical and 91.0% paraphrase success on LIBERO-Goal, and includes real-robot pilot results.

Significance. The empirical core is valuable and, if properly re-scoped, a meaningful contribution to VLA robustness. The paired evaluation protocol (App. A.4) controls episode identity, initial state, seeds, and stochasticity, which makes the large success-rate differences credible. The claim that the method uses canonical demonstrations only is supported by App. A.1 and App. F.2, ruling out paraphrase memorization. The code release and detailed diagnostic definitions are strengths, as is the causal feature-swap experiment for VLA-Adapter, which is a clean demonstration of a model-specific bottleneck. However, the paper's headline causal narrative is broader than the evidence: the universal 'root cause is architectural' claim is directly contradicted by the paper's own layerwise analysis and App. D.3, which explicitly identifies the bottleneck as model-specific. The statistical reporting promised in App. A.5 is also absent. With the causal claims re-scoped and the missing statistics supplied, the paper would be a solid empirical contribution; in its current form the central claims overreach the presented evidence.

major comments (3)
  1. [Abstract; Sec. 2.2; Sec. 4.1; App. D.3] The abstract and Sec. 1 state as a universal finding that 'the root cause is architectural' and that failures 'stem from the joint encoding of dynamic visual observations and text' producing feature shifts that a downstream action policy misroutes. The direct causal evidence is confined to VLA-Adapter: the Sec. 2.2 feature-swap intervention (60% to 96% paired success) targets the final Bridge-Attention block of that model. Sec. 4.1/Fig. 4 reports best single-layer recovery of only 10.5% for SmolVLA and 31.3% for π0.5, and App. D.3 explicitly states that no model satisfied the criteria for a universal semantic breakpoint and that the identified bottleneck is model-specific. The paper's own evidence therefore contradicts the universal causal claim. The empirical GSR gains can stand, but the causal narrative must be re-scoped to VLA-Adapter, with an explicitly architecture-dependent rationa
  2. [App. A.5; Tables 2-4] App. A.5 promises exact two-sided McNemar tests and task-stratified bootstrap 95% confidence intervals for paired outcomes, but no p-values or intervals appear in the main text or appendix. This matters for the headline 'highest reported PRIDE score of 70.4' (vs. Xiaomi's 69.2) and for the π0.5 improvements (Native 73.60 vs GSR 75.59; GSR* 75.76), which are small enough that sampling noise cannot be dismissed, especially with only ten tasks and task-stratified resampling. The large VLA-Adapter and SmolVLA gains are likely robust, but the paper should report the promised paired statistics and, for PRIDE, the uncertainty under task-stratified bootstrap.
  3. [Sec. 3.3; Sec. 4.2; Tab. 3] The 'general guideline' includes a per-model decision—neutralize the native text input for VLA-Adapter and SmolVLA, retain it for π0.5—that is made after observing each model's paraphrase behavior. The route-conflict results in Tab. 3 show that the native pathway can dominate (VLA-Adapter W/C drops to 5.11%) or already be robust (π0.5 Native Drop is 1.6), so the correct choice is essentially the empirical outcome the guideline is supposed to predict. As stated, the neutralize/retain rule is a post-hoc recipe requiring paraphrase labels for the target model, not a cause-derived design. Please either provide a selection criterion that can be evaluated without full paraphrase benchmarks or explicitly position this decision as an architecture-specific design choice rather than part of the general mechanism.
minor comments (6)
  1. [Table 2] The π0.5 Native row reports PRIDE as '–', although PRIDE is computed from paraphrase-weighted success and would be directly comparable to the GSR value of 70.4. Reporting it would strengthen the comparison.
  2. [Sec. 2.2; Sec. 2.3] Retrieval@1 and R in Table 1 are computed from predicted action chunks, not from internal hidden representations. The text later refers to 'internal semantic representation' and 'retained task identity internally.' Align the wording with what is actually measured, or add hidden-state diagnostics.
  3. [Sec. 3.2] The SmolVLA Native + T5 control collapses to 13.49% paraphrase success while canonical success is 76.0%. The paper attributes this to loss of visual grounding, but no analysis is provided for why late injection fails so severely. A brief diagnostic would help readers understand the architecture-specific behavior.
  4. [Fig. 4; Sec. 4.1] The layerwise recovery curves are presented without error bars or per-layer numerical values for SmolVLA and π0.5. Given that the 'no single control point' conclusion is load-bearing for the re-scoped claim, reporting the underlying numbers or a data table would be useful.
  5. [Table 4; App. F] The ParaVLA canonical/paraphrase values (92/91) appear to be based on the Goal-100 diagnostic with a fixed manual paraphrase, while 'Full Para' is reported separately. Clarify this in the table caption or text so readers do not conflate the two evaluation regimes.
  6. [App. D.3] The text says the diagnostic criterion was 'predetermined.' If there is no public preregistration, avoid wording that implies formal preregistration, or state where the protocol was registered.

Circularity Check

0 steps flagged

No circularity: headline results are external closed-loop benchmarks; the causal diagnosis is an intervention, not a fit.

full rationale

The paper's derivation chain is not circular. The central diagnosis (Sec. 2.2) is a direct causal intervention: replacing the paraphrased Qwen feature with its canonical counterpart inside Bridge-Attention raises paired success from 60% to 96%, and the measured target is external closed-loop success rather than the probed feature itself. The 32-direction wording-subspace experiment (Sec. 2.3, D.5) fits directions on eight tasks and tests on held-out tasks using action-gap and success-rate outcomes, with random-direction controls, so the result is not the fit by construction. GSR and ParaVLA are trained only on canonical LIBERO-Goal demonstrations (App. A.1, F.2) and evaluated on the external LIBERO-Para benchmark and the official PRIDE metric; no paraphrase data is used in training, so the headline improvements are not fitted inputs renamed as predictions. T5 is an external pretrained encoder, not a self-citation, and the paper contains no load-bearing self-citation or imported uniqueness theorem. The main caveat is a scientific one rather than a circularity: App. D.3 explicitly concedes that 'No model satisfied all three conditions for a universal cross-architecture semantic breakpoint,' and Sec. 4.1 reports best single-layer action recovery of only 10.5% for SmolVLA and 31.3% for π0.5, so the universal 'root cause is joint encoding' claim is directly demonstrated only for VLA-Adapter. That is an overgeneralization/correctness risk, not a definitional reduction of a prediction to its inputs.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 3 invented entities

The central method depends on three externally supplied ingredients — T5-large's paraphrase robustness, LIBERO-Para's rewrite validity, and the diagnostic assumption that action-chunk geometry mirrors semantic retention. The paper's own contributions are the per-model routing decisions and the re-initialized action expert; those are tested internally. The main uncharged assumption is that the VLA-Adapter causal intervention (canonical-feature swap fixes execution) marks the universal failure mode, which the paper's layerwise analysis partially contradicts. No physics-style invented entities appear; the listed architectural mechanisms are new to this paper and carry no independent evidence outside its experiments.

free parameters (4)
  • Neutral native prompt text ('perform the task')
    Ad hoc text chosen to strip task identity from the native VLM route in VLA-Adapter and SmolVLA GSR (Sec. 3.1, B.1); the decision to neutralize vs. retain the native input is a per-model hand rule.
  • Number of wording-subspace directions = 32 directions
    Linear wording-shift directions fit from 8 tasks' paired canonical/paraphrase features and removed from 2 held-out tasks (Sec. 2.3, D.5); diagnostic only, requires paraphrase-supervised data, and the count 32 is hand-chosen.
  • Per-model native-route policy (neutralize vs. retain)
    GSR guideline (Sec. 3.3): neutralize native input for VLA-Adapter/SmolVLA, retain for π0.5. A hand-specified rule validated by route-conflict experiments (Tab. 3) rather than derived from the stated cause.
  • π0.5 task-token mask boundaries
    'Begins at the first instruction token and ends before the final State: delimiter' (B.4); a hand-specified detection rule that selects where T5 semantics enter PaliGemma.
axioms (5)
  • domain assumption Normalized action-chunk distance under identical observations (Eqs. 1–4) is a valid probe of task-semantic retention
    Sec. 2.2 uses Retrieval@1 and R as evidence that 'the correct task identity survives'; shared observation/state components contribute to every candidate action chunk, and no control shows the distance metric isolates language-driven variation.
  • domain assumption Frozen T5-large supplies paraphrase-invariant task semantics
    Foundation of GSR (Sec. 3.1, B.1); the paper shows empirically that T5 helps but never isolates which property of T5 (scale, pretraining data) produces the stability, and Sec. 5 admits the decoupled visual route does not scale.
  • domain assumption LIBERO-Para rewrites preserve task identity and validly test paraphrastic invariance
    Taken on faith from cited benchmark [19]; the paper performs no audit of rewrite quality or per-rewrite difficulty.
  • ad hoc to paper The VLA-Adapter causal feature-swap result generalizes to other architectures
    The 60%→96% intervention (Sec. 2.2) is performed only on VLA-Adapter; Sec. 4.1/Fig. 4 shows SmolVLA and π0.5 have no single control point (10.5%, 31.3% best single-layer recovery), so the shared 'root cause' is assumed across architectures.
  • domain assumption The PRIDE composite with α=0.5 is a meaningful metric and 1.2-point margins are interpretable
    A.5 defines PRIDE with keyword/structural deviation weights but no CI is reported anywhere; the headline 'new record 70.4 vs 69.2' is a sub-noise margin on a composite while the competitor has higher raw Full-Para success (76.0 vs 75.59).
invented entities (3)
  • T5 semantic sidecar (forked K/V attention in action blocks) no independent evidence
    purpose: Inject stable task semantics into VLA-Adapter's Bridge-Attention policy
    New mechanism (B.2); demonstrated only inside this paper's own ablations; no external check.
  • Neutral native prompt ('perform the task') no independent evidence
    purpose: Remove task identity from the fragile native language route so T5 alone controls semantics
    Introduced Sec. 3.1; its neutrality is asserted, not externally measured.
  • Grounding attention over projected T5 tokens no independent evidence
    purpose: Bind T5 semantics into SmolVLM/PaliGemma native language slots before multimodal processing
    Eq. 7 and B.3–B.4; efficacy tested only in this paper.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Grounded Semantic Re-Binding for Robust Instruction Generalization in Vision-Language-Action Models." pith.science (2026). https://pith.science/paper/CFNZ3JVJ

@misc{pith2026260802497,
  author       = {Pith},
  title        = {Pith review of: Grounded Semantic Re-Binding for Robust Instruction Generalization in Vision-Language-Action Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CFNZ3JVJ}},
  note         = {Machine review of arXiv:2608.02497}
}
Share X LinkedIn Reddit HN
read the original abstract

Vision-Language-Action (VLA) models excel in robotic manipulation but suffer catastrophic performance drops when canonical instructions are simply paraphrased. Although this brittleness is typically addressed through costly data scaling, our probing reveals that the root cause is architectural rather than a lack of semantic understanding. Specifically, we demonstrate that current VLAs successfully retain the correct task identity internally. The failure actually stems from the joint encoding of dynamic visual observations and text, which introduces systematic feature shifts. Because the downstream action policy is highly vulnerable to these variations, it fails to translate the preserved semantics into correct control commands. To resolve this structural bottleneck, we propose Grounded Semantic Re-binding (GSR), an elegant intervention that bypasses unstable joint routing by explicitly fusing independently extracted task semantics with native visual features to train a completely re-initialized action expert from scratch. This targeted intervention dramatically restores paraphrastic invariance using only canonical demonstrations. On the LIBERO-Para benchmark, GSR improves success rates by up to 44.6 percent. It enables lightweight models to rival massively scaled baselines and pushes state-of-the-art models to a new record PRIDE score of 70.4, outperforming the recently introduced large-scale pretrained model Xiaomi-Robotics-0 in instruction generation capabilities. Building on these insights, we also introduce ParaVLA, a natively decoupled 0.33B-parameter model exhibiting near-perfect robustness to instruction rewording. Ultimately, our work proves that robust semantic grounding can be achieved through elegant structural design, bypassing the inefficient brute-force data scaling paradigm.

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

58 extracted references · 1 canonical work pages

  1. [1]

    Qwen technical report, 2023

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng X...

    Pith/arXiv arXiv 2023

  2. [2]

    Paligemma: A versatile 3b vlm for transfer, 2024

    Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, Thomas Unterthiner, Daniel Keysers, Skanda Koppula, Fangyu Liu, Adam Grycner, Alexey Gritsenko, Neil Houlsby, Manoj Kumar, Keran Rong, Julian Eisenschlos, Rishabh Kabra, Matthias Bau...

    Pith/arXiv arXiv 2024

  3. [3]

    URLhttps://arxiv.org/abs/2410.24164

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Lucy Xiaoyang Shi, James Tanner, Quan Vuong, Anna Walling, Haohuan Wang, and Ury Zhilinsky.π0: A visio...

    Pith/arXiv arXiv 2026

  4. [4]

    Rt-2: Vision-language-action models transfer web knowledge to robotic control, 2023

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, Pete Florence, Chuyuan Fu, Montse Gonzalez Arenas, Keerthana Gopalakrishnan, Kehang Han, Karol Hausman, Alexander Herzog, Jasmine Hsu, Brian Ichter, Alex Irpan, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov...

    Pith/arXiv arXiv 2023

  5. [5]

    Rt-1: Robotics transformer for real-world control at scale, 2023

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Tomas Jackson, Sally Jesmonth, Nikhil J Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Kuang-Huei Lee, Sergey Levine, Yao Lu, Utsav Malla, ...

    Pith/arXiv arXiv 2023

  6. [6]

    Observing and controlling features in vision- language-action models, 2026

    Hugo Buurmeijer, Carmen Amo Alonso, Aiden Swann, and Marco Pavone. Observing and controlling features in vision- language-action models, 2026. URLhttps://arxiv.org/abs/2603.05487

    arXiv 2026

  7. [7]

    Xiaomi-robotics-0: An open-sourced vision-language-action model with real-time execution, 2026

    Rui Cai, Jun Guo, Xinze He, Piaopiao Jin, Jie Li, Bingxuan Lin, Futeng Liu, Wei Liu, Fei Ma, Kun Ma, Feng Qiu, Heng Qu, Yifei Su, Qiao Sun, Dong Wang, Donghao Wang, Yunhong Wang, Rujie Wu, Diyun Xiang, Yu Yang, Hangjun Ye, Yuan Zhang, and Quanyun Zhou. Xiaomi-robotics-0: An open-sourced vision-language-action model with real-time execution, 2026. URL http...

    arXiv 2026

  8. [8]

    Semantically controllable augmentations for generalizable robot learning, 2024

    Zoey Chen, Zhao Mandi, Homanga Bharadhwaj, Mohit Sharma, Shuran Song, Abhishek Gupta, and Vikash Kumar. Semantically controllable augmentations for generalizable robot learning, 2024. URLhttps://arxiv.org/abs/2409.00951

    Pith/arXiv arXiv 2024

  9. [9]

    Embodiment Collaboration, Abby O’Neill, Abdul Rehman, Abhinav Gupta, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, Albert Tung, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anchit Gupta, Andrew Wang, Andrey Kolobov, Anikait Singh, Animesh Garg, Aniruddha...

  10. [10]

    When does language matter? multilingual instructions reveal step-wise language sensitivity in vision-language-action models, 2026

    Xuan Dong, Zhe Han, Tianhao Niu, Qingfu Zhu, and Wanxiang Che. When does language matter? multilingual instructions reveal step-wise language sensitivity in vision-language-action models, 2026. URL https://arxiv.org/abs/2606. 11906

    2026

  11. [11]

    Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint, Klaus Greff, Andy Zeng, Igor Mordatch, and Pete Florence. Palm-e: An embodied ...

    Pith/arXiv arXiv 2023

  12. [12]

    Libero-plus: In-depth robustness analysis of vision-language-action models, 2025

    Senyu Fei, Siyin Wang, Junhao Shi, Zihao Dai, Jikun Cai, Pengfang Qian, Li Ji, Xinzhe He, Shiduo Zhang, Zhaoye Fei, Jinlan Fu, Jingjing Gong, and Xipeng Qiu. Libero-plus: In-depth robustness analysis of vision-language-action models, 2025. URL https://arxiv.org/abs/2510.13626

    Pith/arXiv arXiv 2025

  13. [13]

    Cast: Counterfactual labels improve instruction following in vision-language-action models, 2026

    Catherine Glossop, William Chen, Arjun Bhorkar, Dhruv Shah, and Sergey Levine. Cast: Counterfactual labels improve instruction following in vision-language-action models, 2026. URLhttps://arxiv.org/abs/2508.13446

    Pith/arXiv arXiv 2026

  14. [14]

    Not all features are created equal: A mechanistic study of vision-language-action models, 2026

    Bryce Grant, Xijia Zhao, and Peng Wang. Not all features are created equal: A mechanistic study of vision-language-action models, 2026. URLhttps://arxiv.org/abs/2603.19233

    arXiv 2026

  15. [15]

    Ocean: Object-aware anchor-free tracking with matching-relation learning.Machine Intelligence Research, 23(3):565–592, 2026

    Weiming Hu, Zhipeng Zhang, Bing Li, Houwen Peng, and Stephen Maybank. Ocean: Object-aware anchor-free tracking with matching-relation learning.Machine Intelligence Research, 23(3):565–592, 2026. doi: 10.1007/s11633-026-1634-0. URL https://doi.org/10.1007/s11633-026-1634-0

    work page doi:10.1007/s11633-026-1634-0 2026

  16. [16]

    Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Manuel Y . Galliker, Dibya Ghosh, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Devin LeBlanc, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsc...

    Pith/arXiv arXiv 2025

  17. [17]

    Bc-z: Zero-shot task generalization with robotic imitation learning, 2022

    Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning, 2022. URLhttps://arxiv.org/abs/2202.02005

    Pith/arXiv arXiv 2022

  18. [18]

    Learning what to say to your vla: Mostly harmless vision language action model steering, 2026

    Hyun Joe Jeong, Gokul Swamy, and Andrea Bajcsy. Learning what to say to your vla: Mostly harmless vision language action model steering, 2026. URLhttps://arxiv.org/abs/2606.12299. 20

    Pith/arXiv arXiv 2026

  19. [19]

    Libero-para: A diagnostic benchmark and metrics for paraphrase robustness in vla models, 2026

    Chanyoung Kim, Minwoo Kim, Minseok Kang, Hyunwoo Kim, and Dahuin Jung. Libero-para: A diagnostic benchmark and metrics for paraphrase robustness in vla models, 2026. URLhttps://arxiv.org/abs/2603.28301

    arXiv 2026

  20. [20]

    Roboalign: Learning test-time reasoning for language-action alignment in vision-language-action models, 2026

    Dongyoung Kim, Sumin Park, Woomin Song, Seungku Kim, Taeyoung Kim, Huiwon Jang, Jinwoo Shin, Jaehyung Kim, and Younggyo Seo. Roboalign: Learning test-time reasoning for language-action alignment in vision-language-action models, 2026. URLhttps://arxiv.org/abs/2603.21341

    arXiv 2026

  21. [21]

    Openvla: An open-source vision-language-action model, 2024

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model, 2024. URL https: //arxiv.org/abs/...

    Pith/arXiv arXiv 2024

  22. [22]

    Fine-tuning vision-language-action models: Optimizing speed and success, 2025

    Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-tuning vision-language-action models: Optimizing speed and success, 2025. URLhttps://arxiv.org/abs/2502.19645

    Pith/arXiv arXiv 2025

  23. [23]

    Scaling verification can be more effective than scaling policy learning for vision-language-action alignment, 2026

    Jacky Kwok, Xilun Zhang, Mengdi Xu, Yuejiang Liu, Azalia Mirhoseini, Chelsea Finn, and Marco Pavone. Scaling verification can be more effective than scaling policy learning for vision-language-action alignment, 2026. URL https://arxiv.org/ abs/2602.12281

    arXiv 2026

  24. [24]

    Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation, 2024

    Qixiu Li, Yaobo Liang, Zeyu Wang, Lin Luo, Xi Chen, Mozheng Liao, Fangyun Wei, Yu Deng, Sicheng Xu, Yizhong Zhang, Xiaofan Wang, Bei Liu, Jianlong Fu, Jianmin Bao, Dong Chen, Yuanchun Shi, Jiaolong Yang, and Baining Guo. Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation, 2024. URL https://arxi...

    Pith/arXiv arXiv 2024

  25. [25]

    Vision-language foundation models as effective robot imitators, 2024

    Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, Hang Li, and Tao Kong. Vision-language foundation models as effective robot imitators, 2024. URL https: //arxiv.org/abs/2311.01378

    Pith/arXiv arXiv 2024

  26. [26]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling,

  27. [27]

    Libero: Benchmarking knowledge transfer for lifelong robot learning, 2023

    Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning, 2023. URLhttps://arxiv.org/abs/2306.03310

    Pith/arXiv arXiv 2023

  28. [28]

    Rdt-1b: a diffusion foundation model for bimanual manipulation, 2025

    Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation, 2025. URLhttps://arxiv.org/abs/2410.07864

    Pith/arXiv arXiv 2025

  29. [29]

    Rovla: Multi-consistency constraints for robust vision-language-action models, 2026

    Jingzhou Luo, Yifan Wen, Yongjie Bai, Xinshuai Song, Yang Liu, and Liang Lin. Rovla: Multi-consistency constraints for robust vision-language-action models, 2026. URLhttps://arxiv.org/abs/2605.19678

    Pith/arXiv arXiv 2026

  30. [30]

    Interactive language: Talking to robots in real time, 2022

    Corey Lynch, Ayzaan Wahid, Jonathan Tompson, Tianli Ding, James Betker, Robert Baruch, Travis Armstrong, and Pete Florence. Interactive language: Talking to robots in real time, 2022. URLhttps://arxiv.org/abs/2210.06407

    Pith/arXiv arXiv 2022

  31. [31]

    Smolvlm: Redefining small and efficient multimodal models, 2025

    Andrés Marafioti, Orr Zohar, Miquel Farré, Merve Noyan, Elie Bakouch, Pedro Cuenca, Cyril Zakka, Loubna Ben Allal, Anton Lozhkov, Nouamane Tazi, Vaibhav Srivastav, Joshua Lochner, Hugo Larcher, Mathieu Morlon, Lewis Tunstall, Leandro von Werra, and Thomas Wolf. Smolvlm: Redefining small and efficient multimodal models, 2025. URL https://arxiv.org/ abs/2504.05299

    Pith/arXiv arXiv 2025

  32. [32]

    Mechanistic finetuning of vision-language-action models via few-shot demonstrations, 2025

    Chancharik Mitra, Yusen Luo, Raj Saravanan, Dantong Niu, Anirudh Pai, Jesse Thomason, Trevor Darrell, Abrar Anwar, Deva Ramanan, and Roei Herzig. Mechanistic finetuning of vision-language-action models via few-shot demonstrations, 2025. URL https://arxiv.org/abs/2511.22697

    arXiv 2025

  33. [33]

    Colosseum v2: Benchmarking generalization for vision language action models, 2026

    Jeremy Morgan, Prajwal Vijay, Hyeonho Oh, Jincen Song, Ashvin Arora, Alina Du, Gaurav Sukhatme, Jesse Thomason, and Ishika Singh. Colosseum v2: Benchmarking generalization for vision language action models, 2026. URL https: //arxiv.org/abs/2605.27759

    Pith/arXiv arXiv 2026

  34. [34]

    Goal representations for instruction following: A semi-supervised language interface to control, 2023

    Vivek Myers, Andre He, Kuan Fang, Homer Walke, Philippe Hansen-Estruch, Ching-An Cheng, Mihai Jalobeanu, Andrey Kolobov, Anca Dragan, and Sergey Levine. Goal representations for instruction following: A semi-supervised language interface to control, 2023. URLhttps://arxiv.org/abs/2307.00117

    Pith/arXiv arXiv 2023

  35. [35]

    Temporal representation alignment: Successor features enable emergent compositionality in robot instruction following, 2025

    Vivek Myers, Bill Chunyuan Zheng, Anca Dragan, Kuan Fang, and Sergey Levine. Temporal representation alignment: Successor features enable emergent compositionality in robot instruction following, 2025. URL https://arxiv.org/ abs/2502.05454

    Pith/arXiv arXiv 2025

  36. [36]

    Gr00t n1: An open foundation model for generalist humanoid robots, 2025

    NVIDIA, :, Johan Bjorck, Fernando Castañeda, Nikita Cherniadev, Xingye Da, Runyu Ding, Linxi "Jim" Fan, Yu Fang, Dieter Fox, Fengyuan Hu, Spencer Huang, Joel Jang, Zhenyu Jiang, Jan Kautz, Kaushil Kundalia, Lawrence Lao, Zhiqi Li, Zongyu 21 Lin, Kevin Lin, Guilin Liu, Edith Llontop, Loic Magne, Ajay Mandlekar, Avnish Narayan, Soroush Nasiriany, Scott Reed...

    Pith/arXiv arXiv 2025

  37. [37]

    Dinov2: Learning robust visual features without supervision, 2024

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick La...

    Pith/arXiv arXiv 2024

  38. [38]

    Bring the apple, not the sofa: Impact of irrelevant context in embodied ai commands on vla models, 2025

    Daria Pugacheva, Andrey Moskalenko, Denis Shepelev, Andrey Kuznetsov, Vlad Shakhuro, and Elena Tutubalina. Bring the apple, not the sofa: Impact of irrelevant context in embodied ai commands on vla models, 2025. URL https://arxiv. org/abs/2510.07067

    arXiv 2025

  39. [39]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer, 2023. URL https://arxiv.org/ abs/1910.10683

    Pith/arXiv arXiv 2023

  40. [40]

    Disc: Decoupling instruction from state-conditioned control via policy generation, 2026

    Hanxiang Ren, Pei Zhou, Xunzhe Zhou, and Yanchao Yang. Disc: Decoupling instruction from state-conditioned control via policy generation, 2026. URLhttps://arxiv.org/abs/2605.20856

    Pith/arXiv arXiv 2026

  41. [41]

    Cliport: What and where pathways for robotic manipulation, 2021

    Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Cliport: What and where pathways for robotic manipulation, 2021. URL https://arxiv.org/abs/2109.12098

    Pith/arXiv arXiv 2021

  42. [42]

    Smolvla: A vision-language-action model for affordable and efficient robotics, 2025

    Mustafa Shukor, Dana Aubakirova, Francesco Capuano, Pepijn Kooijmans, Steven Palma, Adil Zouitine, Michel Aractingi, Caroline Pascal, Martino Russi, Andres Marafioti, Simon Alibert, Matthieu Cord, Thomas Wolf, and Remi Cadene. Smolvla: A vision-language-action model for affordable and efficient robotics, 2025. URL https://arxiv.org/abs/2506.01844

    Pith/arXiv arXiv 2025

  43. [43]

    Open-world object manipulation using pre-trained vision- language models, 2023

    Austin Stone, Ted Xiao, Yao Lu, Keerthana Gopalakrishnan, Kuang-Huei Lee, Quan Vuong, Paul Wohlhart, Sean Kirmani, Brianna Zitkovich, Fei Xia, Chelsea Finn, and Karol Hausman. Open-world object manipulation using pre-trained vision- language models, 2023. URLhttps://arxiv.org/abs/2303.00905

    Pith/arXiv arXiv 2023

  44. [44]

    Sparse autoencoders reveal interpretable and steerable features in vla models, 2026

    Aiden Swann, Lachlain McGranahan, Hugo Buurmeijer, Monroe Kennedy III, and Mac Schwager. Sparse autoencoders reveal interpretable and steerable features in vla models, 2026. URLhttps://arxiv.org/abs/2603.19183

    Pith/arXiv arXiv 2026

  45. [45]

    Programmatically grounded, compositionally generalizable robotic manipulation, 2023

    Renhao Wang, Jiayuan Mao, Joy Hsu, Hang Zhao, Jiajun Wu, and Yang Gao. Programmatically grounded, compositionally generalizable robotic manipulation, 2023. URLhttps://arxiv.org/abs/2304.13826

    Pith/arXiv arXiv 2023

  46. [46]

    Vla-adapter: An effective paradigm for tiny-scale vision-language-action model, 2025

    Yihao Wang, Pengxiang Ding, Lingxiao Li, Can Cui, Zirui Ge, Xinyang Tong, Wenxuan Song, Han Zhao, Wei Zhao, Pengxu Hou, Siteng Huang, Yifan Tang, Wenhui Wang, Ru Zhang, Jianyi Liu, and Donglin Wang. Vla-adapter: An effective paradigm for tiny-scale vision-language-action model, 2025. URLhttps://arxiv.org/abs/2509.09372

    arXiv 2025

  47. [47]

    Dexvla: Vision-language model with plug-in diffusion expert for general robot control, 2025

    Junjie Wen, Yichen Zhu, Jinming Li, Zhibin Tang, Chaomin Shen, and Feifei Feng. Dexvla: Vision-language model with plug-in diffusion expert for general robot control, 2025. URLhttps://arxiv.org/abs/2502.05855

    Pith/arXiv arXiv 2025

  48. [48]

    Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation, 2025

    Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Kun Wu, Zhiyuan Xu, Ning Liu, Ran Cheng, Chaomin Shen, Yaxin Peng, Feifei Feng, and Jian Tang. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation, 2025. URLhttps://arxiv.org/abs/2409.12514

    Pith/arXiv arXiv 2025

  49. [49]

    Rosa: Harnessing robot states for vision-language and action alignment, 2025

    Yuqing Wen, Kefan Gu, Haoxuan Liu, Yucheng Zhao, Tiancai Wang, Haoqiang Fan, and Xiaoyan Sun. Rosa: Harnessing robot states for vision-language and action alignment, 2025. URLhttps://arxiv.org/abs/2506.13679

    Pith/arXiv arXiv 2025

  50. [50]

    Do what you say: Steering vision-language-action models via runtime reasoning-action alignment verification, 2026

    Yilin Wu, Anqi Li, Tucker Hermans, Fabio Ramos, Andrea Bajcsy, and Claudia Pérez-D’Arpino. Do what you say: Steering vision-language-action models via runtime reasoning-action alignment verification, 2026. URL https://arxiv.org/ abs/2510.16281

    arXiv 2026

  51. [51]

    Stable language guidance for vision-language-action models, 2026

    Zhihao Zhan, Yuhao Chen, Jiaying Zhou, Qinhan Lyu, Hao Liu, Keze Wang, Liang Lin, and Guangrun Wang. Stable language guidance for vision-language-action models, 2026. URLhttps://arxiv.org/abs/2601.04052

    Pith/arXiv arXiv 2026

  52. [52]

    Restoring linguistic grounding in vla models via train-free attention recalibration, 2026

    Ninghao Zhang, Bin Zhu, Shijie Zhou, and Jingjing Chen. Restoring linguistic grounding in vla models via train-free attention recalibration, 2026. URLhttps://arxiv.org/abs/2603.06001

    Pith/arXiv arXiv 2026

  53. [53]

    Ba-net: Bridge attention in deep neural networks, 2024

    Ronghui Zhang, Runzong Zou, Yue Zhao, Zirui Zhang, Junzhou Chen, Yue Cao, Chuan Hu, and Houbing Song. Ba-net: Bridge attention in deep neural networks, 2024. URLhttps://arxiv.org/abs/2410.07860. 22

    Pith/arXiv arXiv 2024

  54. [54]

    X-vla: Soft-prompted transformer as scalable cross-embodiment vision-language-action model, 2025

    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, 2025. URLhttps://arxiv.org/abs/2510.10274

    Pith/arXiv arXiv 2025

  55. [55]

    Vlmbench: A compositional benchmark for vision-and-language manipulation, 2022

    Kaizhi Zheng, Xiaotong Chen, Odest Chadwicke Jenkins, and Xin Eric Wang. Vlmbench: A compositional benchmark for vision-and-language manipulation, 2022. URLhttps://arxiv.org/abs/2206.08522

    Pith/arXiv arXiv 2022

  56. [56]

    Libero-pro: Towards robust and fair evaluation of vision-language-action models beyond memorization, 2026

    Xueyang Zhou, Yangming Xu, Guiyao Tie, Yongchao Chen, Guowen Zhang, Duanfeng Chu, Pan Zhou, and Lichao Sun. Libero-pro: Towards robust and fair evaluation of vision-language-action models beyond memorization, 2026. URL https: //arxiv.org/abs/2510.03827. 23

    Pith/arXiv arXiv 2026

  57. [2023]

    URLhttps://arxiv.org/abs/2210.02747

    Pith/arXiv arXiv

  58. [2025]

    URLhttps://arxiv.org/abs/2310.08864

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.