{"id":"9f5c08f7-27b1-470e-916f-332a2dc6ffe7","arxiv_id":"2608.07434","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A post-hoc conformal wrapper converts any video temporal grounder's single interval into a region that contains the true moment with probability at least 1-alpha.","lead":"This paper introduces Cover, a post-hoc wrapper that widens any video temporal grounder's prediction into a region guaranteed, under exchangeability, to contain the true moment with probability at least 1-alpha. A generalist should read it because it attaches reliability statements to a task where event boundaries are inherently ambiguous and point metrics hide miscalibration.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Per-boundary two-parameter variant lacks a valid joint coverage proof as written: separate 1-alpha quantiles give at best 1-2alpha coverage, so Table A2/A3 efficiency claims are not at matched coverage.","rationale":"The reader's weakest_assumption is exchangeability, which is a legitimate and well-documented limitation, but it is also the standard conformal assumption and the paper tests it extensively (permutation tests, cross-dataset study, Proposition 8). The most load-bearing internal issue is the per-boundary variant's joint coverage claim. The paper uses this variant in a headline efficiency comparison and asserts both members have the same guarantee, yet the text never specifies per-side levels that would make the joint claim true. A simple union-bound calculation shows the stated construction can undercover by a factor of (1-alpha) per side. This is not an attack on the main split-conformal result; it is a concrete, fixable gap in a secondary but promoted contribution. The reader's rationale does mention the per-boundary joint coverage claim as a condition, so there is partial agreement, but the reader did not elevate it to the weakest-assumption slot. A synthetic replication and a check of the released code would settle whether the variant is actually implemented with a correction or whether the paper's efficiency comparisons are invalid as written. If the variant is corrected, the verdict remains CONDITIONAL with the same conditions; if not, the paper's claims for the two-parameter variant should be withdrawn or substantially qualified.","tokens_in":29579,"tokens_out":6598,"duration_ms":69479,"concrete_test":"Re-derive and simulate the per-boundary variant: generate n=1000 calibration and 10000 test pairs with independent standard-normal start and end boundary errors around a fixed predicted interval; calibrate q_s and q_e as the k-th smallest of each side's scores with k=ceil((n+1)(1-alpha)) for alpha=0.1 and 0.2; measure joint coverage of [s_hat-q_s*l, e_hat+q_e*l]. If coverage is approximately (1-alpha)^2, the 'same coverage guarantee' claim is false as written. Then inspect the released code to see whether levels alpha/2 or a joint correction were actually used; if so, Tables A2 and A3 must be re-reported at the corrected levels and the efficiency conclusions may change.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central Proposition 4 is sound for the single-parameter scores. The load-bearing weak spot is the per-boundary variant. The paper calibrates q_s and q_e 'on their own quantiles' (Sec. 4), then states 'Both members carry the same coverage guarantee' (Sec. 7.2). If each side uses the same 1-alpha empirical quantile, the joint event s_start <= q_s and s_end <= q_e has probability at least 1-2alpha by union bound, not 1-alpha; for independent start/end boundary errors it is exactly (1-alpha)^2 (0.81 at alpha=0.10). No Bonferroni, alpha/2, or joint-rank correction is stated in the method or appendix. The claim could be rescued by calibrating each side at 1-alpha/2, or by a joint/union correction, but then the q_s/q_e values in Table A3 and the length savings in Table A2 change. As written, the efficiency comparison between one- and two-parameter variants is not at matched coverage; the two-parameter variant may be shorter precisely because it undercovers. This is an internal correctness issue, independent of exchangeability, and it directly affects a claimed contribution and the paper's headline empirical comparison.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes COVER, a post-hoc, model-agnostic conformal wrapper for video temporal grounding. For a fixed grounder g (trained localizer or black-box VLM), COVER computes a temporal nonconformity score on a held-out calibration set, takes the (1-alpha) empirical quantile qhat of that score, and widens each test prediction by qhat (in a score-dependent way) to produce a temporal region claimed to contain the true interval I* with probability at least 1-alpha under exchangeability. Two score families are introduced: an interval-widening score based on the maximum per-side overshoot normalized by predicted length, and a super-level-set score for grounders exposing a relevance signal. The paper also presents theory on region-length optimality (Theorem 5), length-conditional coverage under a scale model (Proposition 7), and a bound on coverage loss under non-exchangeability (Proposition 8). Experiments on Charades-STA, ActivityNet-Captions, and QVHighlights with five grounders show that calibrated coverage tracks the target for most pairs, while hand-picked margins produce coverage between 0.096 and 1.000, and an evidential grounder (SRAM) misses its nominal level by 14 points.","tokens_in":29822,"tokens_out":4390,"duration_ms":41222,"significance":"If the central claim holds, the paper makes a useful and practical contribution: a single scalar calibration step converts any interval-producing or relevance-signal-producing grounder into one with finite-sample, distribution-free marginal coverage, requiring no retraining or white-box access. The paper is unusually explicit about its assumptions and failure modes, and it reports several honest and informative negative results (e.g., the QVHighlights QD-DETR coverage gap, the positivity limitation in cross-dataset transfer, the SRAM shortfall). The main proof of marginal coverage (Proposition 4) is the standard split-conformal rank argument and is correct. The paper also provides public code, a clear algorithmic description, and experiments across multiple datasets and grounders. The two-parameter per-boundary variant, which is presented as carrying the same guarantee and is used in the headline efficiency comparison, is not justified as written; this is a load-bearing gap that affects the central claim's scope and the validity of Table A2/A3 comparisons.","major_comments":[{"comment":"The per-boundary variant calibrates qhat_s and qhat_e independently at the same (1-alpha) level and then states in Sec. 7.2 that \"Both members carry the same coverage guarantee.\" This is not justified: the events that the start side is covered and that the end side is covered each hold with probability at least 1-alpha, but their intersection (joint containment of the whole interval I*) holds with probability at least 1-2alpha by the union bound, and exactly (1-alpha)^2 in the independent case. No Bonferroni correction, alpha/2 split, or joint-rank correction appears in Sec. 4 or in the Appendix. Consequently the efficiency comparisons in Tables A2–A3 are not at matched coverage: the two-parameter variant can appear shorter precisely because it undercovers. The claim can be repaired by calibrating each side at 1-alpha/2 (or by a joint correction), but then qhat_s, qhat_e, and the length savings in Tables A2–A3 would change, so the empirical comparison must be rerun and the text revised.","section":"Sec. 4 (Per-boundary variant) and Sec. 7.2, Tables A2–A3"},{"comment":"The paper's headline claim is a coverage guarantee for \"any video temporal grounder,\" but the experiments document two real violations of Assumption 1: QD-DETR on QVHighlights realizes 0.879 at target 0.90 (Sec. 7.1), attributed to within-dataset subgroup non-exchangeability, and cross-dataset transfer loses up to 18.8 percentage points of coverage with only partial repair by reweighting (Sec. 7.3, Table A13). These are not proof errors, and the paper discusses them in Sec. 8, but they mean the guarantee holds only under the stated exchangeability assumption, which the empirical findings show fails in realistic settings. The abstract and contributions should be qualified accordingly, and the QVHighlights result should be presented as a documented boundary case of the method's validity rather than as a one-off deviation, so readers do not infer unconditional coverage for arbitrary deployment shifts.","section":"Sec. 7.1, Sec. 7.3, Sec. 8"}],"minor_comments":[{"comment":"The sentence \"Tables A2 and A2 in the Appendix compare the single-parameter score against the per-boundary two-parameter variant\" contains a duplicated table number; it should refer to Tables A2 and A3.","section":"Sec. 7.2"},{"comment":"The caption contains the typo \"Fhe same grounder\" and should read \"The same grounder.\"","section":"Figure A1 caption"},{"comment":"The phrase \"under the k/(n+1) floor Proposition 12 predicts\" is ambiguous; it should say \"below the k/(n+1) floor predicted by Proposition 12,\" since the observed coverage 0.879 is lower than the nominal target.","section":"Sec. 7.1"},{"comment":"The paper motivates the task by annotator disagreement and states that ground truth is a distribution over intervals, yet the formal setup and all calibration use a single ground-truth interval I* per query–video pair. The coverage guarantee is therefore coverage of that reference label, not of the distribution of human annotations; the paper should state this explicitly to avoid over-interpreting the guarantee in the presence of annotation variability.","section":"Sec. 3 and Sec. 8"},{"comment":"The statement that Lighthouse caps usable video length and removes 34.5% of ActivityNet-Captions' original pool should be acknowledged as a potentially significant distributional alteration, since the resulting benchmark may not be representative of the full dataset population.","section":"Sec. 6"}],"recommendation":"major_revision","confidential_remarks":"The per-boundary variant is the main technical gap; the central Proposition 4 is correct and the experiments are generally careful, so the issue is fixable within the manuscript's scope. I would not reject, but the paper should either prove a valid joint coverage result for the two-parameter variant (e.g., by calibrating each side at 1-alpha/2) or drop the claim that both members carry the same guarantee and update the efficiency comparisons accordingly. The exchangeability failures are disclosed honestly, but the title and abstract should be tempered to avoid overclaiming unconditional coverage for any grounder in every deployment condition."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core result is real and correctly proved. Proposition 4 is standard split conformal with a temporal nonconformity score, and the proof is the usual exchangeability rank argument; I checked it and it holds. The wrapper itself is useful: it is model-agnostic, needs only the base grounder's outputs, and the experiments are unusually honest. The fixed-margin sweep, the SRAM failure (0.660 vs. its nominal 0.80), and the Qwen start/end asymmetry are all genuinely informative. They also do not hide the QD-DETR QVHighlights shortfall or the cross-dataset positivity limit. Those are real strengths.\n\nThe soft spot is exactly where your stress-test note lands. The per-boundary variant calibrates q_s and q_e separately at 1-alpha and then claims both members carry the same coverage guarantee. That is not justified. Two separate 1-alpha quantiles give joint coverage at best 1-2alpha by union bound, and the efficiency numbers in Table A2 compare the two-parameter variant against the single-parameter one at a lower true coverage level. So the empirical claim “two parameters are shorter for Qwen” may just be measuring undercoverage. This is a paper-internal correctness issue, not a matter of exchangeability. It should be fixed by calibrating each side at 1-alpha/2 and re-running the comparison, or by dropping the variant's coverage claim and presenting it as a heuristic with a diagnostic ratio only.\n\nTwo smaller points. The ActivityNet experiments are run on the 65.5% of videos that survive the 150-second cap; the paper states the drop but the headline numbers should be framed as conditioned on that truncated population. And the motivating claim that ground truth is a distribution over intervals is not actually matched by the guarantee, which covers a single labeled interval (the envelope when there are multiple). That is a presentational mismatch, minor but worth noting.\n\nOverall: this is a modest but genuine contribution, and the main theorem is sound. The per-boundary hole is load-bearing for one variant but does not sink the paper's central claim. A serious editor should send it to review with the expectation that the authors fix the two-parameter argument.","headline":"Solid split-conformal wrapper for temporal grounding, with one real hole: the two-parameter variant is not proven to cover at 1-alpha, so its efficiency comparisons are at unmatched coverage.","tokens_in":30368,"tokens_out":1761,"would_cite":true,"duration_ms":19168,"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 post-hoc conformal wrapper gives any video temporal grounder a finite-sample coverage guarantee.","keywords":["video temporal grounding","conformal prediction","distribution-free coverage","uncertainty quantification","black-box model","temporal nonconformity score","region prediction","calibration wrapper"],"falsifier":"Take a long video, use early clips as calibration and later clips as test for the same query type, and run Algorithm 1. Because the clips are temporally dependent, exchangeability fails; if the measured coverage falls below $1-\\alpha$ by more than the total-variation defect $\\Delta$ computed from Proposition 8, the bound is violated.","tokens_in":29345,"feed_emoji":"🎬","tokens_out":7158,"duration_ms":66908,"temperature":0.7,"pith_summary":"The paper argues that any fixed video temporal grounder—whether a trained localizer or a black-box video–language model—can be wrapped post-hoc so that its output becomes a temporal region containing the true moment with probability at least $1-\\alpha$. The wrapper calibrates a single scalar, the quantile of a temporal nonconformity score over held-out labels, then widens every prediction by that amount. Because the guarantee is finite-sample and distribution-free under exchangeability, no retraining and no white-box access are needed. Across three benchmarks and five grounders, realized coverage tracks the requested target, and the calibration step exposes grounder-specific failure modes that point metrics hide.","feed_headline":"Wrap any video grounder in a coverage guarantee","feed_subtitle":"A post-hoc conformal wrapper turns any temporal grounder into a region predictor with user-set coverage, no retraining.","key_machinery":"The central object is the temporal nonconformity score: for an interval grounder, $s = \\max\\big((\\hat s - s^*)/\\ell, \\ (e^* - \\hat e)/\\ell\\big)$, the larger of the two per-side overshoots normalized by predicted length; for a relevance-signal grounder, $s = -\\min_{t \\in I^*} f_X(t)$. These scores are fed through split-conformal calibration: $\\hat q$ is the $\\lceil (n+1)(1-\\alpha)\\rceil$-th smallest calibration score, and the output region is the nested family $\\mathcal{C}_{\\hat q}(X)$ (interval widened by $\\hat q\\ell$ per side, or the super-level set $\\{t: f_X(t) \\ge -\\hat q\\}$). The theory in Theorem 5 shows that within this one-parameter family the calibrated region is the shortest one with marginal coverage at least $1-\\alpha$, and Proposition 8 bounds the coverage loss by a total-variation defect when exchangeability fails.","core_discovery":"The central discovery is that the object of uncertainty quantification for temporal grounding can be the temporal region itself, not the point interval. Given any grounder, define the score of a query–video pair as the smallest widening of the predicted interval that makes it contain the true moment (or, for a relevance signal, the negative of its minimum over the true moment). Calibrating the finite-sample-corrected quantile of this score on a held-out set and widening each test prediction by that quantile yields a region that contains the true interval with probability at least $1-\\alpha$ (Proposition 4). The paper also proves the calibrated region is asymptotically the shortest in its one-parameter family, gives conditions under which coverage holds within event-length strata, and bounds the coverage loss when moments from one video break exchangeability. Empirically, the wrapper delivers the requested coverage on nearly every configuration, and the calibrated boundary parameters diagnose each grounder's error structure.","pith_inferences":["The same wrapper could extend to other interval-valued outputs, such as audio event detection or spatio-temporal tube localization, where a single scalar quantile would certify simultaneous boundary containment.","The paper assumes one ground-truth interval per query, but its own motivation highlights annotator disagreement; a natural extension would certify coverage against a quantile of the annotation interval distribution rather than a single label.","The certified width itself is a per-query difficulty signal; the paper's downstream-utility analysis shows it can drive selective prediction, but the sign of the width-quality correlation is dataset-dependent, so deployments should check that correlation before using width as a deferral rule.","On small, clustered calibration pools like QVHighlights, the paper's own results suggest that group-conditional or Mondrian calibration should be the default even when marginal coverage looks acceptable, because a few undercovered strata can drag the marginal average below target."],"forward_implications":["Any existing video temporal grounder can be upgraded to a reliable region predictor with only a labeled calibration set and a single sort, making uncertainty quantification accessible for black-box APIs.","The calibrated region is asymptotically the tightest possible within its score family, so the wrapper adds no unnecessary temporal slack.","Under a scale model where relative boundary error is independent of event length, the normalized score yields length-conditional coverage that a seconds-valued score does not provide.","When calibration and test distributions differ, the coverage loss is bounded by the total-variation defect, and duration-bucketed reweighting can restore much of it.","Calibration acts as a diagnostic: the per-boundary calibrated parameters reveal asymmetric error (for example, a VLM with error-free onsets and poorly localized offsets), information point metrics do not carry."],"supporting_citations":[{"why":"Supplies the split-conformal calibration method and the finite-sample coverage guarantee that Proposition 4 instantiates.","marker":"[17]"},{"why":"Provides the two-sided conformalized quantile regression construction that the boundary-widening score generalizes.","marker":"[20]"},{"why":"Gives the risk-controlling calibration procedure used for the miss-rate variant.","marker":"[22]"},{"why":"Supplies the covariate-shift weighting used to repair cross-dataset coverage.","marker":"[33]"},{"why":"Gives the non-exchangeability bound that Proposition 8 generalizes to the temporal setting.","marker":"[34]"},{"why":"Establishes the impossibility of distribution-free conditional coverage, motivating the paper's marginal and group-conditional guarantees.","marker":"[30]"}],"fun_headline_variants":["Wrap any video grounder to emit a region with coverage guarantee","Post-hoc conformal wrapper gives any video grounder a coverage guarantee","Turn any video grounder into a region with a probability guarantee","Certify coverage for any video temporal grounder without retraining","Model-agnostic conformal wrapper gives any grounder a coverage guarantee"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The guarantee collapses if the calibration and test query–video pairs are not exchangeable, and the paper's own experiments show this happens in practice: cross-dataset transfer can lose up to 18.8 points of coverage, and a VLM's within-dataset subgroups can be non-exchangeable after conditioning.","fun_headline_variants_meta":{"raw":{"variants":["Wrap any video grounder to emit a region with coverage guarantee","Post-hoc conformal wrapper gives any video grounder a coverage guarantee","Turn any video grounder into a region with a probability guarantee","Certify coverage for any video temporal grounder without retraining","Model-agnostic conformal wrapper gives any grounder a coverage guarantee"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000952,"raw_usage":{"total_tokens":4078,"prompt_tokens":980,"completion_tokens":3098,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":3009}},"tokens_in":596,"tokens_out":3098,"duration_ms":22309,"temperature":1.0,"reasoning_tokens":3009,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T04:41:19.136280+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a long video, use early clips as calibration and later clips as test for the same query type, and run Algorithm 1. Because the clips are temporally dependent, exchangeability fails; if the measured coverage falls below $1-\\alpha$ by more than the total-variation defect $\\Delta$ computed from Proposition 8, the bound is violated.","supporting_citations":[{"cited_title":"Distribution-free predictive inference for regression.Journal of the American Statistical Association, 113(523):1094–1111, 2018","cited_arxiv_id":null,"evidence_quote":"Supplies the split-conformal calibration method and the finite-sample coverage guarantee that Proposition 4 instantiates."},{"cited_title":"Conformalized quantile regression","cited_arxiv_id":null,"evidence_quote":"Provides the two-sided conformalized quantile regression construction that the boundary-widening score generalizes."},{"cited_title":"Distribution- free, risk-controlling prediction sets.Journal of the ACM, 68(6):1–34, 2021","cited_arxiv_id":null,"evidence_quote":"Gives the risk-controlling calibration procedure used for the miss-rate variant."},{"cited_title":"Conformal prediction under covariate shift","cited_arxiv_id":null,"evidence_quote":"Supplies the covariate-shift weighting used to repair cross-dataset coverage."},{"cited_title":"Conformal prediction beyond exchangeability.Annals of Statistics, 51(2):816–845, 2023","cited_arxiv_id":null,"evidence_quote":"Gives the non-exchangeability bound that Proposition 8 generalizes to the temporal setting."},{"cited_title":"The limits of distribution-free conditional predictive inference.Information and Inference, 10(2):455–482, 2021","cited_arxiv_id":null,"evidence_quote":"Establishes the impossibility of distribution-free conditional coverage, motivating the paper's marginal and group-conditional guarantees."}],"review_version":1}