{"id":"31fbc251-80bb-40ab-8538-67296efe6966","arxiv_id":"2610.25961","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"PaceVLA lets a robot policy choose per phase how coarse and how long its action chunks are, speeding execution up to 1.8x while keeping task success roughly constant.","lead":"PaceVLA adds a learned judge and router to vision-language-action robot policies so the robot can switch between fine and coarse action steps depending on what the current moment requires. The paper reports up to 1.8x faster task completion with roughly unchanged success rates across three simulation benchmarks and six real-world tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The quantizability-judge supervision pipeline is the load-bearing link: the undisclosed 'computed facts' may leak privileged information into training labels, and the LVLM ratings are never validated, yet Table 5a shows the judge is the entire source of the efficiency gain.","rationale":"The reader's weakest assumption is the same one I would isolate. The central claim is empirical: substantial execution speedup while maintaining task success. The mechanism that produces adaptive speed is the judge and router, and Table 5a quantifies that the judge is load-bearing: removing it lowers TPR from 2.75 to 2.01 and ES from 1.14 to 1.01. The judge's training signal consists of LVLM-generated per-chunk quantizability scores; if that signal does not track phase-level precision requirements, the judge cannot select the correct group and the method degrades toward a router that favors easy-to-quantize targets. Appendix C.5 and E make the prompt structure transparent but leave the computed-facts input undefined and provide no evidence that the Gemini ratings are reliable. This is not an objection to using an LVLM; it is a missing control on the pipeline's load-bearing input. I also considered the controller-gain asymmetry (Sec. 3.5, C.3) as an alternative primary concern, since the base model uses default gains while accelerated methods get tuned gains. That is a real secondary issue for the real-world success comparison, but the simulation speedups are step-count based and the adaptive-selection mechanism is still governed by the judge, so the label-validation gap is more central. The proposed check directly settles whether the supervision is grounded in observations: if labels change when computed facts are removed, or downstream ES drops, the published pipeline is not reproducible and the claimed mechanism is unsupported. Because the concern has not yet been shown to falsify the measured results, the reader's CONDITIONAL verdict remains appropriate, and I would not change it.","tokens_in":25983,"tokens_out":10581,"duration_ms":74450,"concrete_test":"Disclose the computed-facts template and run a controlled annotation study on a fixed sample of about 100 chunks per benchmark: generate Gemini labels with the current prompt including the facts, and again with the facts block removed. Measure per-check agreement between the two label sets (e.g., Cohen's kappa or mean absolute difference). Then retrain the judge using only the no-facts labels and rerun the RoboCasa-Kitchen evaluation from Table 5a. If ES stays near 1.14, the supervision is observationally grounded; if agreement is low or ES drops materially, the current judge is trained on a target that depends on undisclosed or privileged information, and the real-robot gains cannot be attributed to the published adaptive-selection mechanism. In the same pass, check whether any computed fact is derived from future frames, which would settle the leakage question directly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing link in the central claim is the quantizability-judge supervision pipeline (Sec. 3.3, App. C.5, App. E). Table 5a shows that removing the judge drops ES from 1.14 to 1.01, so the efficiency gain above the router-only variant rests entirely on the judge's ability to identify phases that tolerate coarse execution. That ability is only as good as the Gemini 3.8 Flash labels, and the paper provides no validation of them: no human agreement study, no noise analysis, and no definition of the 'computed facts' injected into every prompt in Appendix E. Because those facts are described as measurements derived from the demonstration that the model must treat as true, the labels may encode information unavailable to the judge at inference (for example, ground-truth or future quantities), making the supervision target partly unachievable from the conditioning context Phi_t. If the labels are noisy or depend on privileged facts, the judge selects the wrong granularity group in deployment and the claimed success-maintaining speedup collapses. The authors acknowledge perception-related limitations (b351-b355) but not this validation gap, and releasing prompts alone does not pin down the supervision signal that the whole adaptive-selection mechanism depends on.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"PaceVLA proposes to accelerate VLA policies by making temporal action quantization adaptive. The method augments a pretrained flow-matching VLA with K=4 mode-specific action decoders at two fine-grained and two quantized granularities/horizons, a learned “quantizability judge” that selects whether to use coarse or fine actions at each step, and a context-conditioned router that picks a specific mode within the chosen group. Training combines the standard flow-matching loss (Eq. 1) with a router objective based on per-mode prediction errors (Eq. 2) and a regression target for the judge obtained from LVLM-generated per-step quantizability scores (Eq. 3). Experiments cover RoboCasa-Kitchen, LIBERO, and DexJoCo, plus six real-world tasks on Franka, OpenArm, and ALLEX; the reported results show effective speedups above 1x in all but one simulation setting, 37–45% wall-clock reductions on Franka tasks, and an increase on the ALLEX conveyor task from 3.6 to 6.3 packages/minute while roughly preserving success rates. The central claim is that adaptive, phase-dependent quantization preserves task success while substantially shortening execution, unlike fixed quantization or rule-based acceleration.","tokens_in":26218,"tokens_out":6063,"duration_ms":41626,"significance":"If the mechanism is sound, this is a practical and conceptually clean contribution: it treats temporal granularity and prediction horizon as policy decisions, which is a natural extension of chunked VLA inference, and it validates the idea across two action representations (delta and absolute) and three embodiments. The paper has real strengths: the SO(3)-aware aggregation rules in Appendix A are mathematically correct for delta-pose actions; the flow-matching objectives are conventional and clearly specified; the TPR and ES metrics are sensible because ES charges failures with the full episode budget; and the ablations in Table 5 isolate the contributions of the decoders, judge, and loss terms. The main risk is not in the architecture but in the supervision pipeline that makes the judge work: Table 5a shows that removing the judge reduces ES from 1.14 to 1.01, so the efficiency gain above the router-only variant rests entirely on the unvalidated LVLM-generated labels. Because the authors do not define the “computed facts” injected into every annotation prompt and report no human-agreement or noise analysis, the central mechanism currently rests on an unexamined component.","major_comments":[{"comment":"The quantizability judge is load-bearing: Table 5a shows that removing it drops ES from 1.14 to 1.01, so the entire efficiency gain over the router-only variant depends on the quality of the LVLM-generated labels. However, Appendix C.5 only says that the labels come from Gemini 3.8 Flash annotations, that Claude Opus 5 helped develop the criteria, and that per-task success rates under fixed quantization were used as “task-level cues.” Moreover, every prompt in Appendix E instructs the model to treat “computed facts derived from the demonstration” as true, without ever defining what those facts contain. This creates two concrete risks: (i) the facts may encode information that is not available to the judge at inference from Φ_t, making the regression target in Eq. (3) partially unachievable; and (ii) the labels may be noisy or systematically biased in ways that no experiment in the paper detects. I ask the authors to specify the computed facts, verify that they depend only on quantities available at inference time, and provide label-quality evidence (e.g., human agreement on a held-out set, noise analysis, or an ablation that removes the computed facts from the prompts). This is a correctness-risk concern, not a minor reproducibility issue, because the central speed-success claim depends on this component.","section":"3.3; C.5; E","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The threshold δ that decides when the judge’s predicted score selects the quantized group is never reported. This threshold directly controls the speed-accuracy trade-off and is a free parameter of the method. Without giving its value, or the procedure used to choose it (fixed across benchmarks? tuned per task? selected on a validation set?), the main tables cannot be reproduced, and the reader cannot exclude the possibility that δ was tuned to maximize success rate or speedup on the evaluation tasks. Please report δ for each benchmark and any tuning procedure.","section":"Appendix C.3","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The real-world “maintaining task success” claim is supported by differences on the order of 1–3 percentage points on 32-trial evaluations, e.g., Pick-and-Place 81.3% vs. 84.4% and Table Sweeping 62.5% vs. 59.4%. With 32 Bernoulli trials, these differences are within sampling noise, yet the paper presents them as evidence that success is preserved. The paper should either report confidence intervals or per-task statistical tests, or soften the success-preservation claim to “no clear degradation” given the trial counts. The wall-clock speed gains are robust, but the success-rate component of the central claim needs appropriate uncertainty quantification.","section":"Table 2; Section 4.2","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The prompt heading “Real-World: Allex Conveyer Package Handling Task” misspells “Conveyor” and uses inconsistent capitalization for ALLEX; the main text and Table 4 use “Conveyor.” Please unify the terminology.","section":"Appendix E","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The ES and TPR metrics count action steps, implicitly assuming wall-clock time is proportional to the number of executed steps. Since quantized commands are larger per step, the paper should state explicitly that the simulation controller runs at a fixed control frequency so that step counts and wall-clock time are equivalent; otherwise the reported speedups may overstate wall-clock gains.","section":"Table 1; Appendix B","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The qualitative analysis in Figure 4 is based on a single demonstration. A quantitative summary of judge scores across phases and tasks (e.g., mean score by phase label) would strengthen the claim that the annotations capture phase-level precision requirements.","section":"Figure 4","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Two things to know. First, this is a real contribution: treating temporal granularity and prediction horizon as a routable policy decision, with mode-specific decoders on a shared action expert plus a judge and router, is genuinely not in the cited prior work. DemoSpeedUp and SAIL derive acceleration from action statistics; Mixture of Horizons gates multi-horizon heads without changing granularity. Second, the load-bearing link is the LVLM-generated quantizability supervision, and the paper never validates those labels. That gap, not the architecture, is what needs fixing.\n\nThe paper does a lot well. The SO(3)-aware aggregation in Appendix A is correct and careful — composing rotational increments via Exp/Log instead of summing them — and the delta/absolute distinction is handled properly. The ES metric with the full-episode failure penalty is defensible, and the numbers are internally consistent: ES above 1.0 in all but one setting, up to 1.31, 37–45% wall-clock cuts on the Franka tasks, and 3.6 to 6.3 packages per minute on ALLEX. The ablations are informative: Table 5a shows both the decoders and the judge matter, and removing the judge drops TPR from 2.75 to 2.01. The failure analysis in D.3 is honest, attributing most failures to perception, and they disclose the one ES-below-1 setting rather than hiding it. Prompts and hyperparameters are released.\n\nThe soft spots are proportionate. The judge supervision is the one I would push on. Appendix C.5 says each annotation prompt includes 'computed facts derived from the demonstration' that the LVLM is told to treat as true, but the paper never says what these facts are. If they encode future or ground-truth quantities, the judge's training target is partly unachievable from its inference-time context, and its deployment accuracy is simply untested. There is no human agreement study, no noise analysis, and no ablation over label quality. The bootstrap from per-task success rates is disclosed and is moderate circularity, not fatal — the headline numbers come from genuine rollouts — but the judge's whole job is to transfer the label structure to deployment, and that transfer is unmeasured. Smaller issues: no error bars anywhere, with real-world conclusions resting on 32 trials per task; and controller gains are tuned for the accelerated methods while the base keeps defaults, which slightly mixes policy and controller effects. These are standard for the subfield and neither falsifies the central claim.\n\nThis is for people working on VLA efficiency, action chunking, and faster-than-demonstration execution. It deserves a serious referee: the idea is sound, the effects are large, and the label-validation gap is addressable in revision — define the computed facts, add a human agreement study or label-noise ablation, and report error bars. I would engage with it.","headline":"PaceVLA is a genuine contribution — treating temporal granularity as a routable policy decision — whose large, consistent gains rest on an unvalidated LVLM-judge supervision pipeline; fixable in revision, worth refereeing.","tokens_in":26903,"tokens_out":4652,"duration_ms":28930},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"A robot policy that picks its own action granularity per phase runs up to 1.8x faster without losing task success.","keywords":["vision-language-action models","temporal action quantization","adaptive execution","flow matching","manipulation","dexterous hands","humanoid robot","imitation learning"],"falsifier":"Have several human annotators rate the same first-frame observations used to train the judge for whether coarse execution is safe in that moment, and measure agreement between the LVLM's per-step scores and the human ratings; if agreement is near chance on a held-out set of demonstrations, or if a judge trained on shuffled labels reproduces the reported effective speedup on RoboCasa-Kitchen, then the claimed mechanism is not carrying the result.","tokens_in":27338,"feed_emoji":"🤖","tokens_out":15237,"duration_ms":91707,"temperature":0.7,"pith_summary":"Robot policies trained by imitation learn to copy the slow, cautious tempo of teleoperated demonstrations, so they carry out many redundant, fine-grained action steps even where faster motion is safe. This paper argues that the fix is to let the policy itself decide, at each inference step, how coarsely to act: merge several low-level actions into one larger command when the current phase of the task tolerates it, and fall back to fine-grained steps during precision-sensitive moments such as grasping and placement. Its proposed framework, PaceVLA, couples a shared action expert with four mode-specific decoders (two fine-grained, two quantized, at different prediction horizons), a quantization judge that decides whether coarse execution is appropriate, and a context-conditioned router that picks the decoder to use. The paper reports that this adaptive scheme keeps base-model success rates while achieving effective speedups above 1x in all but one simulation setting, 37 to 45 percent wall-clock reductions on real Franka robot tasks, and faster package throughput on a humanoid robot. If correct, the result is a lightweight way to speed up already-trained VLA policies without retraining on accelerated demonstrations or reinforcement-learning fine-tuning, which matters because execution time is a central obstacle to practical robot deployment.","feed_headline":"Robots that judge when to speed up run 1.8× faster, same success","feed_subtitle":"Each step, a judge-and-router pair picks coarse or fine actions, keeping success flat across three robot forms.","key_machinery":"The framework stands on three coupled objects. First, temporal action quantization with representation-aware aggregation: each source segment of $H_k$ demonstrated actions is collapsed into $T_k$ commands, retaining the final command of each block for absolute actions, while delta translations are summed in a fixed Cartesian frame and rotational increments are composed on SO(3) as $\\varphi^{-1}(\\mathrm{Log}(R_{s_i} \\cdots R_{s_{i-1}+1}))$ so the net commanded rotation survives merging. Second, a shared flow-matching action expert (a diffusion transformer backbone) with four lightweight mode-specific decoders: two fine-grained modes with $q_k = 1$ and two quantized modes with $q_k > 1$, each offered at a longer and a shorter source horizon. Third, the mode-selection layer: a quantization judge $J_\\psi$ regresses an LVLM-annotated per-step quantizability score $y \\in [0,1]$ and gates between the fine and quantized groups at threshold $\\delta$, while an MLP router outputs mode probabilities $\\alpha_t$ trained under a KL-tempered objective that favors whichever decoder has lower prediction error, plus a balance term that prevents collapse onto a single mode. The judge supplies the \"when to accelerate\" decision that fixed quantization lacks, and the router supplies the \"how coarsely\" decision; together they encode the paper's thesis that precision demand is a property of the robot's interaction with the scene, not of the action statistics.","core_discovery":"PaceVLA's central claim is that both temporal granularity and prediction horizon should be treated as policy decisions, adapted online from visual and language context, rather than fixed by the data-collection frequency or by a uniform quantization ratio. A pretrained VLA is fine-tuned so that a quantization judge predicts a per-step quantizability score -- whether the upcoming stretch of motion tolerates coarser commands -- and a router selects among decoders trained at different granularities $q_k = H_k/T_k$ and source horizons. The experiments assert that this adaptive scheme preserves the base policy's success rate where fixed ratio-2 quantization and entropy- or kinematic-rule-based accelerators lose 10 to 30 percentage points: on the Air Fryer task, execution drops from 24.6 s to 14.6 s at unchanged 81.3 percent success, and on the ALLEX conveyor task throughput rises from 3.6 to 6.3 packages per minute while success rises from 86 to 94 percent. The ablations support the two-component design, showing that removing the judge lowers effective speedup from 1.14 to 1.01 and removing the dedicated decoders lowers it to 1.00, so the claimed gains depend on both the \"when to accelerate\" signal and the separately learned coarse decoders.","pith_inferences":["The judge's label source is the untested link: nothing reported shows that the LVLM ratings track true precision demand, so a direct test would be training the same judge on cheap geometric proxies (distance to the contact surface, gripper aperture relative to object size) and checking whether the effective speedup survives.","The threshold $\\delta$ on the judge's score acts as a deployment-time safety knob: operators could tune this single scalar to trade speed against conservatism per environment, which fixed quantization cannot offer.","Because quantized targets are defined by representation-specific aggregation rules, the method transfers most cleanly to policies sharing the same controller conventions; a different action parameterization would require re-deriving the aggregation.","The paper's own failure analysis attributes most residual failures to perception rather than temporal decisions, which suggests that adding depth or tactile sensing could widen the speedup obtainable without sacrificing success."],"forward_implications":["Effective speedup lands above 1x in all but one simulation setting across two VLA backbones, meaning throughput per unit wall-clock time rises even after failed trials are charged their full episode budget.","Real Franka gripper tasks finish 37 to 45 percent faster at comparable or higher success, with the long-horizon Air Fryer task cut from 24.6 s to 14.6 s.","The mechanism transfers across embodiments: on the ALLEX humanoid conveyor task, throughput rises from 3.6 to 6.3 packages per minute while success rises from 86 to 94 percent.","Both core components are needed: removing the quantizability judge drops effective speedup from 1.14 to 1.01, and removing the dedicated coarse decoders drops it to 1.00.","The added per-step inference cost is roughly 5 percent (42.2 ms versus 40.1 ms on the benchmark GPU), so the wall-clock gains come from executing fewer, larger actions rather than from cheaper inference."],"supporting_citations":[{"why":"Supplies the pretrained flow-matching VLA backbone that PaceVLA fine-tunes and that every baseline shares for the main experiments.","marker":"(Bjorck et al., 2025)"},{"why":"Second VLA backbone used on the simulation benchmarks to show the adaptive scheme transfers across base models.","marker":"(Kim et al., 2026)"},{"why":"Defines the fixed ratio-2 temporal action quantization baseline whose success-for-speed trade-off PaceVLA is designed to beat.","marker":"(Tang et al., 2025)"},{"why":"DemoSpeedup baseline that accelerates demonstrations by action entropy; PaceVLA argues this signal cannot see scene-determined precision demand.","marker":"(Guo et al., 2025)"},{"why":"SAIL baseline with preset speed switching, and the source of the throughput-with-regret metric used to score the simulations.","marker":"(Arachchige et al., 2025)"},{"why":"Flow matching objective that trains the shared action expert and all four mode decoders.","marker":"(Lipman et al., 2022)"},{"why":"Diffusion transformer backbone of the action expert onto which the mode-specific decoders are attached.","marker":"(Peebles and Xie, 2023)"},{"why":"Source of the tool used to develop the quantizability criteria and annotation prompts that structure every per-step label.","marker":"(Anthropic, 2026)"},{"why":"The large vision-language model whose API annotations produce the per-step quantizability scores that supervise the judge.","marker":"(Google DeepMind, 2026)"}],"fun_headline_variants":["Robots pick when to speed up, keep success rate flat","Adaptive action steps let robots run faster, still precise","VLA model learns when coarse actions are safe to use","Per-step mode choice boosts robot speed without losing skill","Action quantizer adapts granularity, robots faster and reliable"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The entire efficiency gain rests on the quantization judge, and the judge is only as trustworthy as the per-step quantizability labels generated by a large vision-language model from prompt criteria that the paper never validates against human judgment, noise analysis, or any independent measurement of phase-level precision demand.","fun_headline_variants_meta":{"raw":{"variants":["Robots pick when to speed up, keep success rate flat","Adaptive action steps let robots run faster, still precise","VLA model learns when coarse actions are safe to use","Per-step mode choice boosts robot speed without losing skill","Action quantizer adapts granularity, robots faster and reliable"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000192,"raw_usage":{"cost":0.0001925,"is_byok":false,"cost_details":{"upstream_inference_cost":0.0001925,"upstream_inference_prompt_cost":0.00014798,"upstream_inference_completions_cost":4.452e-05},"total_tokens":1216,"prompt_tokens":1057,"completion_tokens":159,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":0,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":76}},"tokens_in":1057,"tokens_out":159,"cost_source":"openrouter_usage","duration_ms":2797,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":76,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-10-22T01:50:08.206519+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Have several human annotators rate the same first-frame observations used to train the judge for whether coarse execution is safe in that moment, and measure agreement between the LVLM's per-step scores and the human ratings; if agreement is near chance on a held-out set of demonstrations, or if a judge trained on shuffled labels reproduces the reported effective speedup on RoboCasa-Kitchen, then the claimed mechanism is not carrying the result.","supporting_citations":[],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":2}