{"id":"c260911e-b280-432e-acc6-4585d67dc3f4","arxiv_id":"2608.00391","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For vision-language-action robots, token skipping stays reliable only when the gate that selects skipped tokens comes from a clean dense pass, not from the accelerated pass itself.","lead":"This paper shows that a common trick to speed up robot AI models fails silently: when the model's own accelerated run chooses which parts of the image to skip, it can stop noticing important things and task success collapses. Adding one full-speed calculation during the robot's natural waiting time repairs the failure and still serves responses faster than without skipping.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central 'gate, not cache' claim overgeneralized: own Fig. 6 shows mechanism dominates provenance at high skip ratios.","rationale":"The paper's core scientific contribution is the claim that gate provenance—not the skipping mechanism—determines closed-loop reliability. This is stated absolutely in the title, abstract, and conclusion. The strongest evidence is the controlled factorial at 0.9 skip ratio, which cleanly separates the two factors. However, the same paper reports a regime (0.97 skip ratio, Fig. 6) where the mechanism completely changes the outcome: deletion (VLA-Pruner) is immune to provenance degradation while reuse (VLA-Cache) collapses. Under the paper's own attribution rule, this is a 'capacity boundary', but that label is not a mechanism; it is an admission that the skipping mechanism can dominate. Therefore, the universal dichotomy is unsupported and the headline overgeneralizes. The actuation-slack premise (the reader's concern) is a boundary of the proposed repair, not of the diagnosis; the diagnosis's validity is what the paper's central claim rests on. A direct test—extending the controlled factorial to high skip ratios—would settle whether provenance is the dominant factor across the operating range or only at moderate ratios.","tokens_in":15054,"tokens_out":12413,"duration_ms":124078,"concrete_test":"Independently rerun the controlled factorial (Sec. 3.3) on LIBERO-Object at skip ratios 0.9 and 0.97, with both reuse and deletion and both clean and self-harvested gates, using the instrument gate. Test the provenance × mechanism interaction via paired bootstrap. If at 0.97 deletion's self-harvested cell is not significantly degraded relative to clean while reuse's is, the universal 'gate, not cache' claim is unsupported and the paper should be revised to a conditional claim (provenance matters for mechanisms without live gate inputs).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline assertion (title, Abstract, Sec. 6) — 'Where the gate signal comes from, not how tokens are skipped, decides closed-loop reliability' — is not supported by its own results. The controlled factorial at 0.9 on LIBERO-Object (Sec. 3.3) cleanly separates provenance from mechanism, and Table 1 shows gate provenance dominates KV freshness. But at 0.97 on the same task (Fig. 6), the skipping mechanism reverses the picture: VLA-Pruner (deletion) maintains 0.98 even with a self-harvested gate, while VLA-Cache (reuse) collapses to 0.06 (and +R holds only 0.76). Thus 'how tokens are skipped' can be the deciding factor at high ratios, irrespective of provenance. The paper's explanation ('capacity boundary', Sec. 5.4) is descriptive and post hoc, not derived from a formal account of when provenance vs. mechanism dominates. Because the abstract and conclusion state the universal dichotomy without this caveat, the central claim overgeneralizes the controlled result. The refresh method may still work where provenance is the bottleneck, but the central scientific claim — that gate provenance, not mechanism, bounds reliability — is conditional on gate family and operating ratio, and needs reframing.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies training-free token skipping in closed-loop vision-language-action (VLA) policies, distinguishing two skipping mechanisms (KV-cache reuse and token deletion) from the temporal provenance of the gate signal (self-harvested from a sparse forward vs. clean from a dense pass). A controlled paired factorial on LIBERO-Object at a 0.9 skip ratio shows that a self-harvested gate drops success from a dense 1.00 to 0.68 (reuse) and 0.31 (deletion), while a clean gate keeps both mechanisms near dense level. The paper then proposes actuation-slack refresh: one dense forward executed while the robot executes its action chunk, off the critical path, supplying the next step with a clean gate and a fresh KV base. Integrated into VLA-Cache and VLA-Pruner, the refresh repairs the reported collapses and yields 18--22% lower serve latency than dense on OpenVLA-OFT, in simulation and on a physical robot. The paper's central claim is that gate provenance, not the skipping mechanism, decides closed-loop reliability.","tokens_in":15284,"tokens_out":8374,"duration_ms":85464,"significance":"The controlled factorial is a genuine contribution: it isolates gate provenance from KV freshness and mechanism, and the paired protocol with disclosed crash-attrition sensitivity is more careful than is typical in this literature. The refresh idea is simple, practical, and off the critical path, and the paper ships a detailed baseline-fidelity audit. If the central claim is suitably qualified, the paper would be a useful advance for training-free VLA acceleration. However, the paper's own high-ratio results show that the mechanism can dominate provenance, so the headline claim as stated overgeneralizes and needs reframing before the paper can be accepted.","major_comments":[{"comment":"The categorical claim 'Where the gate signal comes from, not how tokens are skipped, decides closed-loop reliability' is contradicted by the paper's own Fig. 6 at a 0.97 skip ratio on LIBERO-Object. There, VLA-Pruner (deletion) maintains 0.98 with a self-harvested gate, while VLA-Cache (reuse) collapses to 0.06 and refresh recovers only 0.76. In that operating regime, the mechanism, not gate provenance, is the deciding factor. Table 2 similarly shows VLA-Pruner receives no consistent benefit from refresh (e.g., CogACT drawer 0.63 to 0.60 at ratio 0.9). The caveat in Sec. 5.4 that 'the precise boundary depends on the task, substrate, and gate family' is too narrow to neutralize the abstract and conclusion. Please restate the central claim as an operating-envelope claim: gate provenance dominates where the self-harvest chain is the limiting factor, while at high ratios and on some substrat","section":"Title, Abstract, Sec. 6 vs. Sec. 5.4 / Fig. 6"},{"comment":"The explanation of the high-ratio crossover is post hoc. The paper attributes VLA-Pruner's survival to its design components (layers 0--3 full-token, union selection, diversity filter) and labels VLA-Cache's collapse as a 'capacity boundary,' but no formal or quantitative model is given for when provenance dominates versus when mechanism dominates. As stated, the account amounts to re-naming whichever cell loses. The contribution list promises an explanation 'beyond the controlled cells'; please either derive a testable boundary condition (e.g., in terms of live full-token layers, mask overlap, or feedback gain) or explicitly mark the account as a hypothesis, and evaluate it against the full Fig. 6/Table 2 pattern.","section":"Sec. 5.4 / contribution 3"},{"comment":"The 'unconditional refresh is necessary' conclusion would benefit from a clearer oracle-trigger comparison. The paper reports that an oracle trigger at a high threshold (γ=0.995) fires on 90.4% of encodes and recovers 1.00 success, yet also states that 'even with the clean-signal cost excluded, oracle-driven triggering reaches 0.66--0.92 across the tested thresholds,' below the unconditional 0.98. The relationship between these numbers and the A.5 trigger-rate table is not transparent. Please specify which thresholds are included in the 0.66--0.92 range, how the clean-signal cost is excluded, and why high-frequency successful oracle triggers are not in that range. Otherwise, the claim that no reactive trigger can replace unconditional refresh overgeneralizes from the specific thresholds tested.","section":"Sec. 5.5 / A.5"}],"minor_comments":[{"comment":"The text says the clean gate is 'statistically indistinguishable from dense,' but no confidence intervals or significance tests are reported. Given the relatively small paired samples, please add error bars or a paired test.","section":"Sec. 3.3 / Fig. 2"},{"comment":"Reporting arithmetic means over four suites/tasks can hide large heterogeneity (e.g., SIMPLER put-in-drawer is much harder). Please report per-cell episode counts and paired confidence intervals, and consider a heterogeneity-aware summary.","section":"Table 2"},{"comment":"Define S_live_{t+1} at first use and clarify that the dense refresh runs on frame x_t, while the gate is applied to frame x_{t+1}; the one-chunk staleness premise is central and deserves a precise notation.","section":"Eq. (1) / Sec. 4"},{"comment":"The baseline-fidelity deviations are listed clearly, but it would help to state explicitly which deviations apply only to the no-refresh arm and which also affect the +refresh arm; the current text mixes implementation choices with protocol-wide constraints.","section":"A.1 / Sec. 5.1"},{"comment":"The sentence 'Every collapse cell survives worst-case treatment of simulator-crash attrition' is terse. Add a pointer to A.2 and state the number of collapse verdicts (39) in the main text for reproducibility.","section":"Sec. 5.2 / A.2"},{"comment":"The real-robot latency table omits task names and per-task episode counts. Please include them in the table or cite A.4 explicitly.","section":"Sec. 5.7 / Table 5"}],"recommendation":"major_revision","confidential_remarks":"The empirical core is careful: the controlled factorial, paired protocol, and sensitivity analysis are real strengths, and I see no evidence of circular fitting. The main obstacle is the overstatement of the central claim in the title, abstract, and conclusion, which the authors' own Fig. 6 undermines. I believe the required changes are reframing and clarification rather than new experiments, but because the overclaim is load-bearing for the paper's headline, I recommend major revision rather than minor."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know this paper before the next round of VLA efficiency work: it isolates a real failure mode—self-harvested gates in training-free token skipping—and shows a simple off-path dense refresh fixes it. The controlled factorial in Sec. 3.3 is genuinely well done: identical episodes, paired comparison, clean-KV vs accumulated-KV, and Table 1 cleanly separates gate provenance from KV freshness. The result that a clean gate holds 0.98 where a self-harvested gate drops to 0.31–0.68 at 0.9 skip ratio is credible and important. The actuation-slack refresh is a neat, almost obvious idea once you see it: run a dense pass during the robot's execution window, discard the action, and use its attention for the next gate and KV base. That it repairs every collapse in their tested baselines, and cuts serve latency 18–22% below dense, is a solid empirical contribution. The paper also discloses its confounds (e.g., the proprioception renormalization in A.1) and runs sensitivity analysis on simulator crash attrition; that's the right discipline.\n\nThe soft spots are real but not fatal. The headline claim—\"where the gate signal comes from, not how tokens are skipped, decides closed-loop reliability\"—is too strong. Their own Fig. 6 shows that at 0.97 skip ratio, VLA-Pruner (deletion) holds 0.98 with a self-harvested gate while VLA-Cache (reuse) collapses to 0.06. Mechanism matters at the stress point. The paper's Sec. 5.4 explains this, and the text in 3.3 says the result doesn't imply mechanism never matters, but the title and abstract don't carry that caveat. That's an overgeneralization, not a load-bearing error, and it's fixable by reframing.\n\nTwo smaller issues. Success-rate tables have no error bars; the appendix gives n and the paired protocol, but readers need confidence intervals, especially where n drops to 50–58. No code is released, which matters for a paper whose contribution is partly methodological. And the refresh's off-path guarantee depends on the actuation window being longer than a dense forward; on CogACT the speed gain largely disappears, and the paper says so, but it means the headline '18–22% faster' is substrate-specific.\n\nWho should read this: anyone working on VLA acceleration, token pruning, or closed-loop policy reliability. The diagnosis and the repair are worth engaging with seriously. With a reframed title and a code release, this becomes a strong paper rather than a conditional one. I'd send it to a good referee.\n\nBest","headline":"Solid controlled study showing self-harvested gates cause silent closed-loop failure in VLA token skipping, and a simple off-path dense refresh fixes it; the only real problem is the title overclaims, since their own Fig. 6 shows mechanism still matters at high ratios.","tokens_in":15826,"tokens_out":3221,"would_cite":true,"duration_ms":28712,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that for closed-loop VLA token skipping, gate provenance—which forward pass supplies the skip decision—bounds reliability: a self-harvested gate collapses success from dense 1.00 to 0.68 (reuse) and 0.31 (deletion), while a","keywords":["vision-language-action models","token skipping","gate provenance","actuation-slack refresh","closed-loop reliability","KV cache reuse","visual token pruning","robotic manipulation"],"falsifier":"Take a closed-loop VLA task at a 0.9 skip ratio where the action-head or vision encoder makes a dense forward pass exceed the action-chunk execution window (as the paper reports for CogACT), and compare success under a clean dense gate versus a self-harvested gate. If the clean gate no longer keeps dense-level success, or the refresh misses its deadline, the gate-provenance claim is substrate-limited rather than universal.","tokens_in":14925,"feed_emoji":"🤖","tokens_out":9273,"duration_ms":86536,"temperature":0.7,"pith_summary":"Token skipping accelerates vision-language-action robots by computing only a few visual tokens per step, selected by a gate. This paper isolates what actually breaks such systems in closed loop: not whether skipped tokens are cached or deleted, but whether the gate is computed by the model's own degraded forward pass ('self-harvested') or by a dense pass that skipped nothing. At a 0.9 skip ratio on LIBERO-Object, the self-harvested gate drops success from dense 1.00 to 0.68 with reuse and 0.31 with deletion, while a clean gate keeps both mechanisms at 0.98; the collapse is invisible to the action-level detectors the authors instrumented. The proposed repair, actuation-slack refresh, hides one dense pass inside the robot's ~400ms action-execution window, handing the next step a clean gate and fresh KV base, and restores 0.98 success while serving 18–22% faster than dense.","feed_headline":"Hidden dense pass restores skip-gated VLA success to 0.98","feed_subtitle":"Gate provenance, not reuse-versus-deletion, decides reliability; the refresh keeps speed while repairing collapse.","key_machinery":"The load-bearing object is the gate, a per-patch signal that decides which visual tokens a VLA forward computes. A self-harvested gate is read from the very sparse forward that served the current action chunk, so tokens skipped at one step receive no fresh signal at the next, allowing blind spots to compound; a clean gate comes from a dense forward that skipped nothing. A KV base stores per-layer keys and values for skippable visual tokens. The paper's mechanism is the actuation-slack refresh: during the roughly 400ms the robot spends executing the current eight-action chunk, the policy runs one dense pass over the same frame, discards its action output, and uses the full-token attention to","core_discovery":"The central claim is that for closed-loop VLA policies, gate provenance—which forward pass supplies the per-patch skip decision—bounds reliability, while the skipping mechanism (cache reuse vs deletion) does not. In a four-cell factorial on LIBERO-Object at a 0.9 skip ratio, a gate computed by a dense forward that skipped nothing keeps both mechanisms at 0.98–1.00 success, statistically indistinguishable from dense. A self-harvested gate computed by the served sparse forward collapses them to 0.68 under reuse and 0.31 under deletion. The failure is silent: the action-level detectors the authors instrumented score near chance, so the paper argues triggering is unreliable and the repair must b","pith_inferences":["Editorial inference: the same self-harvest feedback should appear in any closed-loop policy where the camera moves as a consequence of the policy's own sparse inference, including asynchronous and streaming VLA serving, so those methods deserve the same provenance audit.","Editorial inference: a concrete testable rule follows for practitioners—when validating a skip ratio, compare self-harvested vs clean gate on identical episodes; a gap means the deployment gate must come from a dense pass, and any trigger must be shown to see the pruned tokens.","Editorial inference: the paper's staleness data (every-2-chunk refresh 1.00, every-4-chunk 0.90 on Object) suggest the margin is tunable: one could schedule the refresh less often than every chunk on tasks with slow visual change, though this is not claimed by the authors.","Editorial inference: since detectors were near chance on the tested signals, a promising extension is to detect the failure through the gate's own blind spots—e.g., comparing sparse and dense attention on the dropped tokens—rather than through action signals."],"forward_implications":["Skipping itself is not the enemy: with a clean gate, reuse and deletion both match dense success at 0.9 skip ratio, so acceleration can in principle be reliability-preserving.","Methods that report near-lossless success at a given operating point can hide a failure cliff: at 0.97 skip ratio VLA-Cache falls to 0.06 without refresh.","Triggered repair is structurally risky: SAFE-Pruner's self-harvested consistency trigger fires 3 times in 3,222 encodes on LIBERO-Object while success falls to 0.19, so an unconditional refresh is the reliable default.","The refresh's benefit is not only reliability: on OpenVLA-OFT it serves 18–22% faster than dense, measured in simulation and on a physical robot.","The provenance account grades existing methods: VLA-Cache collapses when the gate is self-harvested, while VLA-Pruner retains enough clean signal to avoid the failure until its substrate boundary is reached."],"supporting_citations":[{"why":"Supplies the reuse mechanism (VLA-Cache) whose self-harvested gate collapses at high skip ratios and which the refresh repairs.","marker":"(Xu et al. 2025)"},{"why":"Supplies the deletion mechanism (VLA-Pruner) and the comparison that shows mechanism does not determine reliability.","marker":"(Liu et al. 2025)"},{"why":"The closest triggered-refresh baseline; its self-harvested consistency trigger fails to fire during provenance collapse.","marker":"(Ma et al. 2026)"},{"why":"The OpenVLA-OFT policy, the primary substrate for the controlled factorial, latency, and real-robot measurements.","marker":"(Kim, Finn, and Liang 2025)"},{"why":"The CogACT policy used to test the refresh on SIMPLER and to show the substrate boundary where no speed gain remains.","marker":"(Li et al. 2024)"},{"why":"The LIBERO benchmark that provides the task suites and the 91 initial states for the paired factorial.","marker":"(Liu et al. 2023)"},{"why":"The SIMPLER benchmark that supplies the Google Robot tasks for cross-policy generalization tests.","marker":"(Li et al. 2025)"},{"why":"The theoretical basis for compounding error under closed-loop execution, supporting why self-harvested blindness accumulates across chunks.","marker":"(Ross, Gordon, and Bagnell 2011)"}],"fun_headline_variants":["Gate source, not skip method, decides VLA closed-loop reliability","Self-harvested gates silently collapse skip-accelerated VLAs","One dense pass off critical path fixes skipped-gate VLA collapse","Gate provenance: why token skipping fails on closed-loop VLA","Actuation-slack refresh: clean gate rescues skip-gated VLAs"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The repair's off-path guarantee rests on the robot's action-execution window being reliably longer than one full dense forward pass on the deployed hardware, and on a dense gate computed at chunk t remaining valid when used at chunk t+1.","fun_headline_variants_meta":{"raw":{"variants":["Gate source, not skip method, decides VLA closed-loop reliability","Self-harvested gates silently collapse skip-accelerated VLAs","One dense pass off critical path fixes skipped-gate VLA collapse","Gate provenance: why token skipping fails on closed-loop VLA","Actuation-slack refresh: clean gate rescues skip-gated VLAs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000254,"raw_usage":{"total_tokens":1481,"prompt_tokens":899,"completion_tokens":582,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":643,"completion_tokens_details":{"reasoning_tokens":489}},"tokens_in":643,"tokens_out":582,"duration_ms":5950,"temperature":1.0,"reasoning_tokens":489,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T04:12:58.857141+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a closed-loop VLA task at a 0.9 skip ratio where the action-head or vision encoder makes a dense forward pass exceed the action-chunk execution window (as the paper reports for CogACT), and compare success under a clean dense gate versus a self-harvested gate. If the clean gate no longer keeps dense-level success, or the refresh misses its deadline, the gate-provenance claim is substrate-limited rather than universal.","supporting_citations":[{"cited_title":"2025 , address =","cited_arxiv_id":null,"evidence_quote":"The OpenVLA-OFT policy, the primary substrate for the controlled factorial, latency, and real-robot measurements."}],"review_version":1}