{"id":"ac14dbd7-ea97-4c0c-b0b9-e87e697fb9c6","arxiv_id":"2610.05719","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"RACE predicts manipulation subskill transition timing and conditions action generation on it, enabling reliable execution of longer action chunks in VLA policies.","lead":"The paper analyzes why long action chunks fail in vision-language-action robot policies and finds errors spike at subskill transitions. It introduces RACE, a framework that predicts transition timing and conditions action generation on it, improving success rates with 2x to 4x longer chunks and cutting idle time about 5x on a real robot.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The central mechanism is validated only on demonstration states: the Fig. 2/4 error curves and Tab.","rationale":"The paper has genuine independent support: three-seed statistics on the headline VLABench number (Tab. 18: 33.3±0.5), held-out validation of PELT labels against generator boundaries (Appendix A.5), sensitivity analyses over PELT penalty (C.1), inference-time perturbation tests (Tab. 5, D.2, D.3), and matched-control ablations (Tab. 3). The empirical claim that RACE improves success rates at longer chunks therefore appears solid, and the reader's CONDITIONAL verdict is appropriate. My stress-test refines the reader's weakest assumption rather than replacing it. The PELT-label imperfection is largely mitigated by the paper's own ablations and sensitivity checks, and the authors explicitly acknowledge imperfect transition timing (F.1). What remains unresolved is sharper: the central mechanistic claim, that errors concentrate at subskill transitions and that conditioning on predicted transition timing reduces them, is only measured on demonstration states. The paper itself states in Appendix E.4 that errors are measured on demonstration states because ground truth is unavailable on rollout states, and Tab. 31 uses 500 training demonstrations for the matched-variation test. On training demos, RACE's transition-error reduction is partly by construction because L_timing optimizes the action expert's features to predict PELT positions on exactly those demos. The only held-out mechanistic evidence, Tab. 27's gripper-switch timing, shows a non-significant exact-match gain on held-out demos but a significant ±1-step gain. Since the model is brittle to one-step prior shifts (Tab. 5), the prior's accuracy on rollout states is the condition that would determine whether the mechanism actually explains the success-rate gains. This is a genuine soft spot in the argument: the end-to-end success rates are strong, but the explanation of why the method works is not validated on the distribution where the policy operates. Verdict remains CONDITIONAL: accept that longer chunks become more reliable, but with conditions on releasing code/weights and on verifying the transition-timing mechanism on rollout states.","tokens_in":36778,"tokens_out":11920,"duration_ms":151225,"concrete_test":"Collect rollout-state measurements on VLABench at H_exec=20: run RACE and fine-tuned π0.5 over 300 held-out episodes, recording the actual observation at each policy call together with the predicted prior p̂_t and the subsequently executed chunk. Recompute the error analysis of Fig. 4 and Tab. 31 on these rollout states, defining transition steps from the executed action trace via the same PELT procedure (or the generator reference boundaries), and measure the head's ±3-step localization accuracy against both label sets. If localization accuracy on rollout states falls well below the held-out-demonstration value of 65.1% (Tab.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two parts: (1) long-chunk errors concentrate at subskill transitions, and (2) RACE's transition-timing conditioning reduces those errors and thereby raises the chunk-extension ceiling. Part (1) defines transitions via PELT change points, which agree with the VLABench generator's reference boundaries at only 0.82 precision and 0.70 recall (Appendix A.5), so roughly a third of reference transitions are unlabeled. The paper partially defends label quality (sensitivity to PELT penalty in Appendix C.1, label-source ablation in Table 4a), so label noise alone is not the decisive gap. The load-bearing gap is that every mechanistic measurement of the claimed effect is made on demonstration states, not rollout states (Appendix E.4). Moreover, the error-diagnosis figures (Fig. 2, Fig. 4) and the matched-variation comparison behind them (Tab. 31) use training demonstrations: on training demos, RACE's transition-error reduction is partly by construction, since L_timing back-propagates into the action expert and explicitly reshapes its features to predict PELT positions on exactly these demos. Held-out support is thinner: Tab. 27's exact gripper-switch gain on held-out demos is +3.2 pp with a 95% CI of [-1.2, 7.6]. The transition-timing prior's accuracy is validated only against PELT labels on demonstration states (Tab. 23); on the states the policy actually visits after open-loop drift, neither prior accuracy nor transition-error reduction is measured. This matters because the model is sensitive to prior placement: Tab. 5 shows a one-step shift drops success rate from 33.0 to 30.2, below the 31.8 zero-prior setting. If the one-step auxiliary pass cannot recover reliable transition signal on true rollout observations, the mechanism collapses even if end-to-end success rates remain elevated for other reasons (e.g., the auxiliary L_timing objective alone contributes about +4.9 pp over fine-tuning, per Tab.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper studies why extending the action chunk of flow-matching VLAs degrades task success. Using PELT change-point detection on demonstration action traces to define subskill transitions (Appendix A), the authors show that action prediction errors concentrate near transitions and grow with chunk length (Fig. 2, Fig. 4, Tab. 31). They propose RACE, which predicts a per-step transition-timing prior from an auxiliary one-step denoising pass plus cached VLM features (Sec. 3.2) and conditions every denoising step on that prior through token-wise adaptive-RMSNorm modulation with learned per-step gates (Sec. 3.3). Training adds a BCE timing loss whose gradient back-propagates into the action expert. In simulation, RACE outperforms same-chunk fine-tuned pi 0.5 on VLABench, RoboCasa-H50, and LIBERO, exceeds ACoT-VLA and efficient/adaptive baselines at 2x longer chunks, and stays competitive at 4x. On a real robot, 4x-longer chunks reduce idle time about 5x while raising success above same-chunk fine-tuning (Tab. 6).","tokens_in":37042,"tokens_out":11103,"duration_ms":63438,"significance":"If the results hold, RACE provides a cheap post-training route to a higher reliable chunk-extension ceiling (about 1.3x fine-tuning training cost, 4% per-call latency overhead, Tabs. 15-17) with a clean mechanism, which is directly relevant to stop-and-go VLA deployment. The empirical practice is a genuine strength: three-seed statistics (Tab. 18), bootstrap confidence intervals (Tab. 27), PELT-penalty sensitivity (Tab. 19), validation against the generator's reference boundaries (Tab. 8), a matched local-variation control (Tab. 31), and careful latency and training-cost decompositions (Tabs. 15-17, 28). The headline rollout-level result (RACE beats same-chunk fine-tuning on three benchmarks and a real robot) is solid and, because the error-concentration diagnosis is established on fine-tuned pi 0.5 without any timing objective, that half is not circular. The weaker half is the RACE-specific mechanism: its error measurements are demonstration-state only, and the held-out exact-timing gain is not significant, so the causal narrative is supported but not established.","major_comments":[{"comment":"The mechanistic claims are measured only on demonstration states. Appendix E.4 states that ground-truth actions exist only for demonstration states, and Fig. 2, Fig. 4, and Tab. 31 use training demonstrations. On training demonstrations, part of RACE's transition-error reduction is built in by construction: L_timing back-propagates into the action expert (Sec. 3.2) and explicitly reshapes its features to predict PELT positions on exactly those demonstrations. The out-of-sample support is partial: Tab. 27 shows the exact-step gripper gain on held-out demonstrations is +3.2 pp with a 95% CI of [-1.2, 7.6], and Tab. 23 shows the head recalls only 50-58 percent of transitions within ±1 step on held-out demonstrations at H >= 10. To support the narrative 'RACE reduces transition errors, hence raises the chunk ceiling', the paper should either add rollout-state evidence (e.g., flag predicted transition steps in the policy's own rollouts and compare deviation from a replanned or demonstrated reference) or explicitly reframe the Sec. 4.4 claims as demonstration-state measurements that support, but do not establish, the causal story.","section":"Sec. 4.4 and Appendix E.4","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The timing-supervision stack rests on PELT pseudo-labels that agree with the VLABench generator's reference boundaries at 0.82 precision and 0.70 recall (Tab. 8), so roughly a third of reference transitions are unlabeled. The paper defends robustness to the penalty beta (Tab. 19) and to the label source (Tab. 4a), which is fair, but the residual risk is systematic rather than random: if PELT consistently misses a particular transition type, the prior is biased and the head's held-out recall of 50-58 percent at ±1 step for H >= 10 (Tab. 23) will not be uniform across subskill types. A concrete and cheap test is to report per-reference-boundary recall broken down by skill duration or action-change magnitude; this would show whether missed transitions are a random subset or a biased one, and whether the timing-conditioning benefit survives on chunks where the reference boundary is captured.","section":"Appendix A.5 and Tab. 23","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The sentence 'RACE switches within one step of the correct step more often' should be qualified, because on held-out demonstrations the exact-step gain is +3.2 pp with a 95% CI of [-1.2, 7.6] (not significant), while the ±1-step and missed-switch differences are significant; the current wording overstates the exact-timing benefit.","section":"Sec. 4.4 and Tab. 27","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The real-robot success rates (36, 48, and 66 percent) lack confidence intervals and seed information; a binomial 95% CI on 50 trials would align this section with the paper's otherwise high statistical standard.","section":"Tab. 6 and Appendix E.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Spd is defined over episodes that both the model and the pi 0.5 baseline complete successfully and excludes simulator execution time (Appendix E.2); this conditioning is not flagged in the main text, and Tab. 29 shows non-trivial task-level variation, so a footnote would prevent misreading.","section":"Tables 1 and 2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The 'Modulation offsets' columns report the learned RMSNorm offset magnitudes but no units or norm are specified; please state that these are, e.g., mean L2 norms over the offset tensors.","section":"Table 4a","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The prior-accuracy numbers (Tab. 23) are PELT-agreement scores on demonstration states; the main text should state this at the point of use in Sec. 4.4, since the current phrasing could be read as rollout-level transition accuracy.","section":"Sec. 4.4 and Appendix D.1","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Colleague,\n\nWhat's actually new here: the diagnosis that long-chunk errors concentrate at subskill transitions and grow with chunk length, plus a mechanism (RACE) that predicts transition timing from an auxiliary one-step denoising pass and conditions every denoising step on it. That conditioning mechanism is not in the cited prior work, which either shortens execution around transitions (PACE), extends chunks via RL (PolicyTrim), or injects subskill structure as goals or auxiliary outputs. The paper clearly identifies a real failure mode of chunk extension and attacks it directly.\n\nThe empirical scaffolding is unusually solid for a robotics paper: PELT penalty sensitivity, validation of the labels against the VLABench generator's reference boundaries (0.82 precision / 0.70 recall), a matched local-action-variation control, bootstrap CIs on gripper-switch timing, and three-seed statistics on the headline number. They also show the prior matters: shifting it one step drops success below the zero-prior setting, and it moves the generated gripper switch. That is good, honest work.\n\nSoft spots, in proportion. The biggest is that every mechanistic measurement of the transition-error reduction is made on demonstration states, not rollout states (Appendix E.4). The authors acknowledge this directly, but it means the central claim—\"RACE reduces transition errors\"—is demonstrated on states the policy never visits. They partially compensate with the inference-time shift experiments, which show the prior affects rollouts, but the transition-error reduction itself on true rollout states is unmeasured. Second, the labels are PELT pseudo-labels with no human annotation; against the generator they miss about a third of reference transitions. The authors do sensitivity analysis, and label noise alone isn't decisive, but the whole timing-supervision stack rests on them. Third, the real-robot result is a single pick-and-place task, 50 trials, no CIs, and only successful-trial means. Fourth, code and weights are promised, not shipped.\n\nIs the paper partly circular? The model is trained to minimize errors exactly where it claims to reduce them, so the reduction is partly by construction. But the held-out gripper-switch gain (+4.6 pp within ±1 step, CI excluding zero) and the success-rate gains at 2x and 4x chunks are not circular, and they hold up.\n\nIn short: a solid, deployable improvement for action-chunking and efficient-VLA researchers. The paper deserves serious refreeing. I'd send it to peer review with a request for conditional acceptance: address the demo-state measurement limitation (at least by reporting rollout-state transition errors on a subset), strengthen the real-robot statistics, and ship the code.\n\nRegards.","headline":"Careful empirical work on a real failure mode; the mechanism is plausible but the transition-error claim is measured on demonstration states, so accept conditionally.","tokens_in":37704,"tokens_out":1039,"duration_ms":8393},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"Long action chunks fail where manipulation subskills switch, and the paper shows that predicting the switch timing makes longer chunks reliable.","keywords":["vision-language-action models","action chunking","subskill transitions","transition timing","change-point detection","flow matching","robotic manipulation","efficient inference"],"falsifier":"Measure the transition error on the states the policy actually visits during rollouts rather than on demonstration states: the paper's error analysis (Appendix E.4) is computed on demonstration states only, so a rollout-state measurement showing RACE's transition-error reduction disappears would undercut the central claim. A second, direct check: use a label source systematically wrong for one subskill type (e.g., gripper-only switches) and confirm that the success-rate gain vanishes on tasks dominated by that subskill type.","tokens_in":40065,"feed_emoji":"🤖","tokens_out":2556,"duration_ms":42921,"temperature":0.7,"pith_summary":"The paper asks why longer action chunks in vision-language-action (VLA) robot policies become unreliable, and identifies a specific culprit: action errors concentrate at the steps where one manipulation subskill switches to the next (approach to grasp to lift, say), and these transition errors grow sharply with chunk length. It introduces RACE, which predicts the timing of these subskill switches from a cheap auxiliary one-step denoising pass and conditions every denoising step of the action expert on that prediction. If the claim is right, the reliability ceiling on chunk extension is substantially raised: RACE outperforms ordinary fine-tuning at every tested chunk length, beats recent state-of-the-art and efficient VLAs in success rate at twice the chunk length, stays competitive at four times, and on a real robot cuts stop-and-go idle time by about fivefold while exceeding the same-chunk fine-tuning success rate.","feed_headline":"Predicting skill-switch timing makes long robot chunks reliable","feed_subtitle":"RACE predicts when a robot task changes phase, beats fine-tuning at every chunk length, and cuts idle time ~5x.","key_machinery":"The central object is the transition-timing prior $\\hat{p}_t$, a score in $[0,1]^H$ for each action in a chunk indicating how close it is to a subskill switch. It is produced by (1) an auxiliary one-step denoising pass, which yields the action expert's final-layer hidden states $F_t$ at the cost of about 4 percent of a policy call; (2) a prediction head combining those features with cached VLM features via cross-attention, self-attention, and a sigmoid; and (3) a supervision signal from PELT change-point detection on demonstration action traces, converted into Gaussian soft labels so each transition affects roughly seven steps. Conditioning injects the prior into every denoising step by adding token-wise offsets to the scale, shift, and gate parameters of the adaptive RMSNorm layers, gated per step by learnable $\\alpha_k$, which after training decrease monotonically so the prior modulates the early denoising steps most strongly. The mechanism is that the prior tells the denoiser where boundaries will appear, and the learned gates align that injection with the steps where the action expert's representation of adjacent actions is most separable.","core_discovery":"Action errors in long chunks do not merely grow smoothly with prediction horizon; they spike at subskill transitions, and the spikes grow sharply with chunk length. RACE treats 'when to switch' as a learnable quantity: an auxiliary one-step denoising pass extracts the action expert's final-layer hidden states, a prediction head converts these together with cached VLM features into a per-position transition-timing prior $\\hat{p}_t \\in [0,1]^H$, and that prior modulates every denoising step through token-wise offsets into the action expert's adaptive RMSNorm layers, with per-step learnable gates $\\alpha_k$. Timing supervision also back-propagates into the action expert through the head, making its representations transition-aware even before conditioning is applied. On VLABench, RoboCasa-H50, and LIBERO, RACE beats fine-tuning at the same chunk length; at $2\\times$ longer chunks it surpasses recent state-of-the-art and efficient VLAs in success rate; at $4\\times$ it stays competitive; and on a real robot a $4\\times$-longer chunk reduces idle time about $5\\times$ while exceeding same-chunk fine-tuning success.","pith_inferences":["The observation that errors concentrate at subskill boundaries, and that boundary timing degrades with horizon, plausibly transfers beyond flow-matching VLAs to any sequence predictor over structured skills - autoregressive or diffusion models in video generation or long-horizon planning - where boundary prediction is cheap relative to the sequence. That is an editorial extrapolation, not the pape","A testable extension the authors leave implicit: combine RACE's predicted transition timing with adaptive chunking, truncating execution exactly at the predicted next transition, which could recover both reliability and latency that pure extension loses at the longest chunks.","The transition head misses roughly one to 1.6 transitions per episode on held-out demonstrations (Appendix D.1); recovering the missed ones - via a second pass or by exploiting gripper and velocity signals - may close part of the remaining gap at $H_{exec}=40$, where RACE falls below the baseline.","The learned gates and the adjacent-action-token similarity dip suggest that a sparsity or schedule prior concentrating the timing signal on the earliest denoising steps could sharpen transition separation further; this is measurable with the paper's own Fig. 6 setup."],"forward_implications":["RACE beats fine-tuning at every tested chunk length on all three benchmarks, so the transition-timing signal helps not only very long chunks but every chunk length.","At $2\\times$ longer chunks it exceeds recent state-of-the-art and efficient VLAs in success rate, and at $4\\times$ it remains competitive, establishing a higher reliability ceiling for chunk extension.","The gain comes with light post-training - a moderate overhead over fine-tuning (roughly 17-36 hours across benchmarks on four GPUs), far below RL-based chunk extension.","On real hardware with asynchronous execution, a $4\\times$-longer chunk removes nearly all idle time after startup while improving success over same-length fine-tuning.","Gains grow with chunk length and with the number of subskill transitions per task, consistent with the proposed mechanism that transition timing is what degrades in long chunks."],"supporting_citations":[{"why":"Supplies the PELT change-point algorithm that produces the unsupervised subskill-transition labels on which all of RACE's timing supervision rests.","marker":"Killick et al., 2012"},{"why":"The base flow-matching VLA ($\\pi0.5$) that RACE post-trains, providing the action expert whose hidden states and adaptive RMSNorm layers are modulated.","marker":"Physical Intelligence et al., 2025"},{"why":"The adaptive-normalization pattern (adaptive RMSNorm) that the transition-conditioned token-wise modulation extends.","marker":"Peebles & Xie, 2023"},{"why":"Flow matching, the training objective for both the auxiliary one-step pass and the full conditioned denoising pass.","marker":"Lipman et al., 2023"},{"why":"Prior work using change-point detection for subskill discovery in demonstrations; RACE's label pipeline follows its approach.","marker":"Jia et al., 2024"},{"why":"Real-time chunking and asynchronous execution, the deployment protocol used in the real-robot experiments and the idle-time analysis.","marker":"Black et al., 2025"},{"why":"VLABench, the main simulation benchmark whose scripted generator provides the reference subskill boundaries used to validate the PELT labels.","marker":"Zhang et al., 2025"},{"why":"ACoT-VLA, the recent state-of-the-art VLA baseline that RACE must beat at longer chunks and compares against in latency and training cost.","marker":"Zhong et al., 2026"},{"why":"PACE, the adaptive-chunking baseline whose speed-minima transition labels are compared against PELT in the label-validity ablation.","marker":"Nie et al., 2026"}],"fun_headline_variants":["Timing skill switches makes robot action chunks longer and reliable","Predict subskill switch times to extend robot action chunks reliably","RACE learns when to switch making long robot action chunks reliable","Switch timing is the key to longer reliable robot action chunks","Cut robot idle time 5x by predicting subskill switch timing"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The whole method rests on PELT change-point detection on demonstration action traces as a stand-in for true subskill boundaries, and those pseudo-labels carry no human annotation, reaching only 0.82 precision and 0.70 recall against the VLABench generator's reference boundaries.","fun_headline_variants_meta":{"raw":{"variants":["Timing skill switches makes robot action chunks longer and reliable","Predict subskill switch times to extend robot action chunks reliably","RACE learns when to switch making long robot action chunks reliable","Switch timing is the key to longer reliable robot action chunks","Cut robot idle time 5x by predicting subskill switch timing"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000721,"raw_usage":{"cost":0.000720748,"is_byok":false,"cost_details":{"upstream_inference_cost":0.000720748,"upstream_inference_prompt_cost":0.000122388,"upstream_inference_completions_cost":0.00059836},"total_tokens":3216,"prompt_tokens":1079,"completion_tokens":2137,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":2052}},"tokens_in":1079,"tokens_out":2137,"cost_source":"openrouter_usage","duration_ms":10882,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":2052,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-10-07T00:23:09.793767+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":"Measure the transition error on the states the policy actually visits during rollouts rather than on demonstration states: the paper's error analysis (Appendix E.4) is computed on demonstration states only, so a rollout-state measurement showing RACE's transition-error reduction disappears would undercut the central claim. A second, direct check: use a label source systematically wrong for one subskill type (e.g., gripper-only switches) and confirm that the success-rate gain vanishes on tasks dominated by that subskill type.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the PELT change-point algorithm that produces the unsupervised subskill-transition labels on which all of RACE's timing supervision rests."},{"cited_title":"IndoorUAV: Benchmarking Vision-Language UAV Navigation in Continuous Indoor Environments","cited_arxiv_id":"2504.16054","evidence_quote":"The base flow-matching VLA ($\\pi0.5$) that RACE post-trains, providing the action expert whose hidden states and adaptive RMSNorm layers are modulated."},{"cited_title":"Scalable diffusion models with transformers","cited_arxiv_id":null,"evidence_quote":"The adaptive-normalization pattern (adaptive RMSNorm) that the transition-conditioned token-wise modulation extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Flow matching, the training objective for both the auxiliary one-step pass and the full conditioned denoising pass."},{"cited_title":"ACoT-VLA : Action chain-of-thought for vision-language-action models","cited_arxiv_id":null,"evidence_quote":"ACoT-VLA, the recent state-of-the-art VLA baseline that RACE must beat at longer chunks and compares against in latency and training cost."},{"cited_title":"PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking","cited_arxiv_id":"2606.00537","evidence_quote":"PACE, the adaptive-chunking baseline whose speed-minima transition labels are compared against PELT in the label-validity ablation."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}