{"id":"e4be6be3-b23a-4012-a07b-f058e459d636","arxiv_id":"2608.09125","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A trajectory-divergence safety gate that truncates unstable action chunks improves VLA-based surgical manipulation success on real hardware.","lead":"A new test-time filter for surgical robots detects when a vision-language-action model's planned motion is becoming unreliable and replans before errors grow. On a da Vinci-like dual-arm platform it raised end-to-end success on tissue manipulation from 55% to 80% and on needle manipulation from 55% to 60%.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"TDHD's central claim hinges on the untested assumption that single-sample flow-matching divergence δ_i (Eq. 2) tracks per-step execution risk; without validating this proxy, the reported gains may reflect threshold tuning or replanning frequency rather than the divergence signal.","rationale":"The reader's weakest assumption is indeed load-bearing: the entire mechanism depends on δ_i being a trustworthy indicator of when open-loop execution becomes unsafe. I agree with that identification and sharpen it with a concrete technical detail: δ_i is estimated from a single random perturbation η, so even if the proxy is conceptually sound, the estimate is noisy and could trigger truncation at arbitrary steps. The paper reports no calibration of δ_i against execution outcomes, and the unreported thresholds θ_abs and r_thresh make hidden tuning a real risk. The proposed random-truncation control is decisive because it isolates whether the divergence signal, rather than the act of replanning more often, is responsible for the gains. In the needle suite the headline improvement is 55% to 60% (one additional success in 20 trials), so without a control the result is especially fragile. This concern does not invalidate the paper outright; it is empirically testable, and the reader's CONDITIONAL verdict already captures the need for stronger validation. I therefore recommend leaving the verdict unchanged while emphasizing that the proxy validation and threshold disclosure should be prerequisites for acceptance.","tokens_in":11981,"tokens_out":3955,"duration_ms":40561,"concrete_test":"Run a control condition on the needle-pick and tissue-resection tasks in which truncation times are drawn randomly from TDHD's empirical truncation distribution (same s_min, s_max, and average executed step lengths), independent of δ_i, with the same number of trials. If random truncation matches TDHD's success rates (60% and 80% in Table I), then the divergence signal is not doing causal work and the central claim fails. If random truncation performs substantially worse, the proxy is at least informative and the concern is alleviated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim (Section I, Table I) is that TDHD improves VLA success by truncating action chunks at 'unstable' steps. This requires the assertion in Section III-C that δ_i = ||A_i^(1) - A_i^(2)||_2, the L2 distance between two flow-matching trajectories generated from z1 and z1 + εη, is a reliable proxy for prediction instability and execution risk. That assertion is never tested: no analysis correlates δ_i with actual per-step action error, stage-level failure, or execution success. Moreover, δ_i is computed from a single random perturbation η, making it a high-variance one-sample estimate of sensitivity to initial noise rather than a calibrated uncertainty measure. Because θ_abs and r_thresh are not reported and no ablation varies them, the observed gains could come from a carefully chosen truncation schedule that happens to replan during contact transitions, rather than from the divergence signal itself. The Table II ablation compares against fixed horizons but includes no control that truncates at matched random times, so it cannot separate 'replan more often in hard stages' from 'divergence identifies hard steps.' Without such a control, the central claim is underdetermined.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Trajectory Divergence Horizon Decision (TDHD), a test-time mechanism that adaptively decides how many steps of a predicted action chunk to execute before replanning. For each chunk, the policy generates two action trajectories via deterministic flow matching from a primary noise sample and a perturbed copy; the per-step L2 divergence δ_i between the two rollouts (Eq. 2) is used as a proxy for prediction instability. A dual-threshold rule (absolute threshold θ_abs and ratio threshold r_thresh, Eqs. 3-5) truncates execution at the first 'unstable' step, with the executed horizon clamped to [s_min, s_max] (Eq. 6). The method is evaluated on a real da Vinci-like dual-arm platform with 600 teleoperated demonstrations across needle (reach, pick, regrasp) and tissue (reach, lift, resection) suites, comparing against π0.5, π0, RDT, and Octo. The headline results are end-to-end success improvements from 55% to 60% for needle manipulation and from 55% to 80% for tissue manipulation, with an ablation comparing TDHD against fixed-horizon execution strategies.","tokens_in":12269,"tokens_out":3494,"duration_ms":40147,"significance":"If the reported gains are robust, TDHD would offer a lightweight, retraining-free intervention that improves the reliability of vision-language-action policies in contact-rich surgical manipulation, which is a genuine and timely contribution. The strengths of the paper include real-hardware deployment, a new dual-arm surgical benchmark with synchronized multi-view perception, a relatively large teleoperation dataset, hierarchical success metrics, and comparison against several recent VLA baselines. The central idea of using divergence between two flow-matching rollouts as a stability signal is interesting and clearly motivated by local Lyapunov exponents. However, the significance is currently limited by evidentiary issues: the proxy is not validated, the free thresholds are unreported, the main success differences are within the noise of 20-trial evaluations, and the ablation lacks a matched random-truncation control. These issues do not invalidate the concept, but they do mean the quantitative claims are not yet established.","major_comments":[{"comment":"The central claim rests on the assertion that the step-wise L2 divergence δ_i between two flow-matching trajectories generated from z1 and z1 + εη is a reliable proxy for prediction instability and execution risk. This assertion is never tested: no analysis correlates δ_i with measured per-step action error, stage-level failure, or end-to-end success. Moreover, δ_i is computed from a single random perturbation η, making it a high-variance one-sample estimate of sensitivity to initial noise rather than a calibrated uncertainty measure. A validation experiment is needed, e.g., correlating δ_i with downstream position error or with success under different truncation rules, to show that the divergence signal rather than the act of replanning drives the gains.","section":"III-C, Eq. (2)"},{"comment":"The method's free parameters are never reported: θ_abs, r_thresh, ε, s_min, s_max, and the action chunk length H are all absent, and no sensitivity analysis is provided. Because the dual-threshold rule is defined by these parameters, the reader cannot assess how robust the reported gains are to threshold choice. Additionally, the 'Fixed s_best' baseline in Table II is selected by scanning s ∈ {5,10,15,20,25} on what appears to be the evaluation set, which is selection on the test data and invalidates it as a fair comparison. Finally, the ablation text states there were 'minor setup changes (e.g., changes in camera viewpoints),' so Table II is not directly comparable to Table I (for example, fixed s_max tissue reach drops from 80% in Table I to 70% in Table II). These issues leave room for the observed gains to reflect hidden calibration rather than the divergence mechanism.","section":"IV-E, Table II; IV-D, Table I"},{"comment":"No confidence intervals or significance tests are reported for any success rate. With 20 trials per task setting, the headline needle improvement from 55% to 60% is a difference of one successful trial (11/20 vs 12/20), and the pick conditional improvement from 85% to 90% is also one trial (17/20 vs 18/20). Under an exact binomial test, the needle total difference is far from significant. The claim in the abstract and Section IV-D that TDHD 'consistently improves performance' is not supported by the needle result without additional statistical evidence or a larger number of trials. At minimum, the paper should report exact binomial confidence intervals and pairwise tests, and ideally increase the trial count for the headline conditions.","section":"IV-D, Table I"},{"comment":"The ablation cannot separate the mechanism's specific benefit from a general benefit of more frequent replanning during contact-sensitive stages. The comparison is only against fixed-horizon execution; there is no control that truncates at matched random time steps or at fixed intervals calibrated to TDHD's average executed horizon (e.g., 17-23 steps). Since the paper reports that TDHD tends to truncate during contact transitions and the largest gains appear in exactly those stages, the observed improvement could be explained by 'replan more often in hard stages' rather than by 'divergence correctly identifies hard steps.' Adding a random-truncation control with matched replanning frequency would directly address this underdetermination.","section":"IV-E"}],"minor_comments":[{"comment":"The dataset description is ambiguous: it says 'we collected exactly 100 successful trials' and then 'we allocated 80 trials for the training set and 20 for testing.' It is unclear whether the 20 test trials are successful demonstrations used for evaluation or whether the 20 trials per policy in Table I are independent re-executions that may include failures. Please clarify the relationship between the demonstration split and the evaluation protocol.","section":"IV-B"},{"comment":"The ablation reports only aggregate success rates over 10 runs per condition; no variance or per-run breakdown is given. Given the small sample, reporting exact counts or confidence intervals would help interpret the differences (e.g., 90% vs 80% in needle pick could be 9/10 vs 8/10).","section":"IV-E"},{"comment":"The sentence listing average executed step lengths for needle manipulation (21.15, 22.08, 20.40) appears twice, once in the paragraph preceding the analysis and again in the analysis itself. Please remove the duplicate.","section":"IV-E, final paragraph"},{"comment":"The ratio threshold uses ξ in the denominator but its value is never specified; even a small positive constant should be reported, since the ratio rule can be sensitive to ξ when δ_{i-1} is near zero.","section":"III-C, Eq. (4)"},{"comment":"The conclusion states that future work will investigate 'threshold calibration, the impact of trajectory ensemble size in TDHD, and its consistency across alternative VLA models.' These are appropriate future directions, but the first two items directly bear on the validity of the current results and should be addressed at least partially in this paper, not deferred entirely, given that thresholds are unreported.","section":"V"}],"recommendation":"major_revision","confidential_remarks":"The paper introduces a timely and practically motivated idea, and the real-hardware benchmark is a potentially valuable asset for the community. My main concern is evidentiary rather than conceptual: the central proxy is unvalidated, the free parameters are unreported, the sample sizes are too small for the headline effect sizes, and the ablation design cannot rule out a generic replanning-frequency benefit. I recommend major revision with a request for (1) direct validation of the divergence proxy against execution outcomes, (2) full reporting of thresholds and a sensitivity analysis, (3) statistical tests or larger trials, and (4) a random-truncation control in the ablation. If the authors can supply these, the paper could become a solid contribution. I do not see a citation or scope problem that would justify rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this paper proposes a cheap test-time wrapper for any flow-matching VLA policy, and it shows real-hardware gains, especially in contact-heavy tissue resection (55% to 80%). That is a useful result if it holds. The novelty is real: truncating the execution horizon based on step-wise divergence between two noise-perturbed trajectories is not in the cited baselines. The benchmark itself is a contribution — a da Vinci-like dual-arm setup with 600 teleoperated demonstrations across needle and tissue subtasks, multi-view perception, and language instructions. That is careful work and deserves credit.\n\nWhat the paper does well: the mechanism is simple, computationally light (one extra decoder pass), and grounded in the determinism of the flow-matching ODE. The ablation against fixed horizons shows the gains come from adaptive truncation, and the pattern — longer horizons in stable phases, shorter in contact transitions — matches intuition. The paper is also honest that this is a controlled phantom setting and that threshold calibration is future work.\n\nWhere it is soft. First, the statistics: 20 trials per setting, no confidence intervals, no significance tests. The needle gain is 60% versus 55%, i.e., one additional success. The tissue gain is bigger, but still 20 trials. Second, the thresholds θ_abs, r_thresh, ε, s_min/s_max are never reported and never ablated. That leaves room for hidden calibration on the evaluation set, especially since the ablation selects Fixed s_best by scanning s values on what looks like the test data. Third, the central proxy — that δ_i, the L2 distance between trajectories from z1 and z1+εη, tracks per-step execution risk — is asserted, not tested. There is no correlation between δ_i and actual action error or stage failure. And it is a single-sample estimate, which is high-variance. The stress-test is right that a control with random truncation times would separate \"replan more in hard stages\" from \"divergence identifies hard steps.\" I do not think the paper is circular — success is not defined as low divergence — but it is underdetermined.\n\nWho it is for: people working on surgical VLA deployment or test-time reliability layers. It is an engineering wrapper, not a new capability class, but the benchmark and the idea are worth engaging.\n\nRecommendation: send it to review, but with heavy revision requests: more trials, confidence intervals, full threshold reporting, a random-truncation control, and ideally code and data. The core idea is plausible and the hardware work is real; the evidence is just not at the level the abstract implies.","headline":"Plausible test-time safety gate for surgical VLA, with a real benchmark and real gains in contact stages, but the evidence is statistically thin and the divergence proxy is never validated.","tokens_in":12790,"tokens_out":1906,"would_cite":false,"duration_ms":28623,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A divergence gate lifts surgical VLA success to 80% on tissue tasks","keywords":["surgical robotics","Vision-Language-Action models","flow matching","adaptive execution horizon","trajectory divergence","dual-arm manipulation","test-time safety gate","open-loop execution"],"falsifier":"One decisive check is to record, for the same action chunks, the step at which the true end-effector tracking error first exceeds a tolerance and compare it with TDHD's chosen truncation step; if truncation routinely happens after the error has already accumulated, the divergence proxy is not measuring execution risk. Another is to rerun the tissue trials with different phantom stiffness or target locations and see whether the reported tissue success rate holds; if the gain vanishes when the deformable material changes, the divergence signal is tuned to this benchmark's specific failure modes.","tokens_in":1889,"feed_emoji":"🤖","tokens_out":2392,"duration_ms":129336,"temperature":0.7,"pith_summary":"Surgical Vision-Language-Action (VLA) policies typically commit to a fixed number of open-loop action steps, and under millimeter-precision contact with deformable tissue those blind executions accumulate errors. This paper argues that the right remedy is not a better policy but an adaptive test-time gate: decide, after each action chunk is generated, how many of its steps can be trusted. The proposed Trajectory Divergence Horizon Decision (TDHD) generates a second trajectory from a slightly perturbed noise sample, measures the per-step divergence between the two, and truncates execution at the first unstable step, forcing replanning from fresh observations. On a real dual-arm surgical robot with 600 teleoperated demonstrations, the paper reports that TDHD raises end-to-end success from 55% to 60% on needle manipulation and from 55% to 80% on tissue manipulation, with the largest gains in the contact-heavy final stages. The claim is that a lightweight divergence-based truncation rule can make fixed-horizon VLA policies reliable enough for surgical subtasks without retraining.","feed_headline":"A divergence gate lifts surgical VLA success to 80% on tissue tasks","feed_subtitle":"The gate spots unreliable action steps by comparing two flow-matched trajectories, then replans before errors grow.","key_machinery":"The central object is the per-step trajectory divergence $\\delta_i = \\|A^{(1)}_i - A^{(2)}_i\\|_2$, computed between two action trajectories generated by the same flow-matching velocity field from two noise samples $z_1$ and $z_2 = z_1 + \\epsilon\\eta$. Because flow-matching generation is deterministic given the noise and conditioning, this divergence isolates the model's sensitivity to tiny input perturbations, and the paper treats it as a proxy for prediction instability and execution risk, motivated by the classical idea that nearby trajectories diverging rapidly indicate instability. The dual-threshold stopping rule combines an absolute threshold $\\theta_{\\mathrm{abs}}$ on $\\delta_i$ with a ratio threshold $r_{\\mathrm{thresh}}$ on $\\delta_i/(\\delta_{i-1}+\\xi)$, taking the earliest step where either is violated, and then constrains the final horizon to $[s_{\\min}, s_{\\max}]$ before dispatching actions to the robot.","core_discovery":"The paper's central discovery is that the reliability of a flow-matching VLA action chunk can be assessed at test time by comparing it with a twin trajectory generated from a slightly perturbed initial noise vector. The step-wise $\\ell^2$-distance $\\delta_i$ between the two trajectories serves as an instability score: small divergence means the action at step $i$ is safe to execute open-loop, large or sharply rising divergence means the model's prediction is unreliable and the robot should stop and replan. A dual-threshold rule selects the earliest unsafe step, and the executed horizon is clipped to a physical range $[s_{\\min}, s_{\\max}]$. On real hardware with 20 trials per setting, the paper shows this adaptive horizon control consistently beats fixed-horizon execution, with the clearest gains in needle pick/regrasp and tissue resection, where contact transitions cause the largest prediction drift.","pith_inferences":["The same divergence gate should transfer to any flow-matching or diffusion VLA policy whose action generation is deterministic given the initial noise, since the gate only needs two forward passes of the action decoder; the paper only reports results on one fine-tuned policy family, so this transfer is an extrapolation.","The divergence signal could also serve as an operator-facing risk monitor during semi-autonomous surgery: a rising $\\delta_i$ marks the precise moments when a human should be ready to intervene, a use the paper does not discuss.","A natural extension the paper leaves implicit is per-stage threshold calibration; its own ablation shows task-dependent benefits, so tuning $\\theta_{\\mathrm{abs}}$ and $r_{\\mathrm{thresh}}$ separately for reach, pick/lift, and regrasp/resection could yield further gains.","The gate's estimate uses only one perturbed trajectory; averaging divergence over several perturbations would reduce the variance of the instability estimate, and the paper's own future-work list mentions ensemble size as an open question."],"forward_implications":["If TDHD works as reported, fixed-horizon open-loop execution is identified as a major, addressable source of surgical VLA failure: the same policy with adaptive truncation outperforms any single fixed execution length in the ablation study.","The gains are stage-dependent: TDHD is most active during contact transitions such as needle pick and tissue resection, while stable phases like reach and lift execute close to the full horizon, so the gate behaves like a targeted safety intervention rather than a blanket slowdown.","Because TDHD only modifies how an already generated action chunk is executed, it can be applied on top of existing fine-tuned VLA policies to improve reliability without retraining or changing the policy weights.","The dual-threshold design means the gate responds both to sustained divergence and to sudden divergence jumps, allowing it to catch instability even when absolute divergence values remain modest."],"supporting_citations":[{"why":"Supplies the flow-matching VLA policy that TDHD is applied to and fine-tuned on the paper's own dual-arm demonstration dataset; it is the strongest fixed-horizon baseline the method must beat.","marker":"[28]"},{"why":"Supplies an earlier flow-matching VLA baseline that the paper fine-tunes and compares against; its fixed-horizon failures on contact stages motivate the divergence gate.","marker":"[27]"},{"why":"Supplies a diffusion-based VLA baseline that generates full action trajectories, providing the comparison point for deterministic flow matching and for whole-chunk generation.","marker":"[30]"},{"why":"Supplies an open-source generalist VLA baseline whose low reach-stage success on the surgical tasks emphasizes the need for task-specific fine-tuning and adaptive control.","marker":"[25]"},{"why":"Supplies the dynamical-systems concept that rapidly diverging nearby trajectories indicate instability, which motivates the perturbation-based divergence proxy.","marker":"[33]"},{"why":"Supplies the recent VLA practice of filtering sampled actions via runtime verification, which TDHD adapts from explicit verification to a continuous divergence-based safety gate.","marker":"[34]"},{"why":"Grounds the interpretation of divergence as a mixture of epistemic uncertainty from under-represented states and aleatoric uncertainty from deformable tissue contact.","marker":"[35]"}],"fun_headline_variants":["Divergence gate lifts surgical VLA success to 80%","Adaptive horizon replanning boosts surgical robot success","Surgical VLA gains from trajectory divergence gating","Test-time trajectory check lifts surgical success 55 to 80%"],"cache_read_input_tokens":14848,"weakest_assumption_plain":"The load-bearing premise is that the divergence between two nearly identical generated action predictions tracks how likely those actions are to actually fail on the robot; if that proxy breaks down, the gate will truncate at the wrong times and the reported success-rate gains will not transfer beyond the paper's specific setup.","fun_headline_variants_meta":{"raw":{"variants":["Divergence gate lifts surgical VLA success to 80%","Adaptive horizon replanning boosts surgical robot success","Surgical VLA gains from trajectory divergence gating","Test-time trajectory check lifts surgical success 55 to 80%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00055,"raw_usage":{"total_tokens":2666,"prompt_tokens":1030,"completion_tokens":1636,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":646,"completion_tokens_details":{"reasoning_tokens":1567}},"tokens_in":646,"tokens_out":1636,"duration_ms":13007,"temperature":1.0,"reasoning_tokens":1567,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:59:57.048010+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One decisive check is to record, for the same action chunks, the step at which the true end-effector tracking error first exceeds a tolerance and compare it with TDHD's chosen truncation step; if truncation routinely happens after the error has already accumulated, the divergence proxy is not measuring execution risk. Another is to rerun the tissue trials with different phantom stiffness or target locations and see whether the reported tissue success rate holds; if the gain vanishes when the deformable material changes, the divergence signal is tuned to this benchmark's specific failure modes.","supporting_citations":[{"cited_title":"Local lyapunov exponents in chaotic systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the dynamical-systems concept that rapidly diverging nearby trajectories indicate instability, which motivates the perturbation-based divergence proxy."}],"review_version":1}