Grounding with Confidence: Controllable Generative Video Temporal Grounding

1 INTRODUCTION

• Applications such as moderation may favor recall, while recommendations and targeted clipping may favor precision, motivating adjustable acceptance criteria over the same candidate pool.These criteria let applications change returned intervals without regenerating timestamps.Source: §1 · p. 1
• Ranking matters because generation order need not reflect interval reliability, and absent-event queries require deciding whether to return any candidate at all.A plausible timestamp may be returned even when matching content is absent.Source: §1 · pp. 1–2
• Existing confidence alternatives have trade-offs: sequence likelihood does not directly supervise localization correctness, repeated sampling requires multiple answers, and a visual verifier adds inference.The proposed score is trained for individual temporal predictions and available from the original grounding pass.Source: §1 · p. 2
• A lightweight head scores intervals from pooled decoder states, enabling ranking, threshold-based selection, and rejection without regenerating timestamps or using an inference-time verifier.The same candidate pool can support different acceptance criteria.Source: §1 · p. 2
• GT-anchored supervision trains the generator to cover annotated occurrences, while offline verifier scores and online temporal-overlap labels train confidence; set-level RL optimizes coverage and selected-set quality.Unsupported candidates teach the head to reject without becoming generator timestamp targets.Source: §1 · pp. 2–3
• +4.47 and +4.64 points over generation order in query-macro Recall@0.5 at 10% and 25% global return budgets, respectively.These gains are reported on fixed OMTG-Bench candidates.Source: §1 · p. 3

2.1 INTERVAL CONFIDENCE AND ACCEPTANCE

• The model emits a variable-length interval set, and a span-pooled decoder-state head scores each interval during the original decode.At inference, it reuses cached timestamp-generation states without adding confidence tokens; teacher-forced replay is used only for training.Source: §2.1 · p. 3
• Generation-order NMS forms pool P; confidence then supports thresholded variable-cardinality selection, including empty sets, or ranking under a return budget.Thresholding and budgeted selection operate on the same completed decode; neither changes interval boundaries or recovers candidates removed by NMS.Source: §2.1 · p. 3

2.2 LEARNING INTERVAL CONFIDENCE

• Offline replay trains the confidence head on fixed candidates using verifier soft labels, soft binary cross entropy, and within-query ranking.Replay includes supported, unmatched, GT-completion, and duration-based negative candidates; unsupported candidates teach rejection without becoming generator targets.Source: §2.2 · p. 4
• Online adaptation uses binary overlap labels on NMS survivors, with detached hidden states and a supervised loss that updates only the head.These local labels indicate overlap with a GT occurrence, not an interval’s marginal value after other candidates are accepted.Source: §2.2 · pp. 4–5

2.3 TRAINING THE CANDIDATE GENERATOR

• Generator targets retain every GT interval as an anchor and add bounded supported boundary alternatives; set-level RL then optimizes coverage and selected-set quality.RL uses the same NMS and confidence selector as evaluation, with a warm start and a reward combining recall, selected-set F1, and union tIoU.Source: §2.3 · p. 5

3.1 EXPERIMENTAL SETUP

• The OMTG-Bench evaluation uses 320 queries and 1,173 official GT intervals; selectors reuse cached predictions, with equal global return counts in the fixed-pool comparison.The main OMTG models use generation-order NMS at tIoU 0.3 and the official visual/512-token budget.Source: §3.1 · pp. 5–6, §3.2 · p. 6

3.2 SELECTION UNDER EQUAL RETURN BUDGETS

• Confidence improves official query-macro recall over Order by 4.47 and 4.64 points at 10% and 25% retention, respectively.Both selectors rank the same fixed candidate pool under equal global return budgets.Source: §3.2 · p. 6, Table 1 · p. 6

3.3 THRESHOLD SELECTION AND TRANSFER

• On OMTG-Bench, the two transferred operating points achieve micro precision of 71.71% (654/912) and 86.89% (391/450), respectively, and neither target transfers.Their micro recall values are 55.75% and 33.33%, respectively; the relative contributions of threshold-selection variability, cohort composition, and changes in the score–correctness relation remain unresolved.Source: §3.3 · p. 7

3.4 END-TO-END GROUNDING AND COMPONENT ANALYSIS

• RL gains 3.52 points in tF1@0.3, 2.85 in tF1@0.5, and 2.49 in tIoU over SFT.The models share the backbone, visual budget, parser, NMS, and one-decode inference setup.Source: §3.4 · p. 7, Table 2 · p. 8
• Raw recall rises from 67.85 to 73.64 as RL reduces mean candidate count from 11.18 to 8.99, about 20%.Confidence filtering adds 0.32 points on the fixed RL pool, with paired 95% interval [−0.14, 0.76].Source: §3.4 · p. 7, Table 3 · p. 8
• Post-NMS AUROC rises from 0.754 to 0.818 with the online head on fixed L200 candidates and decoder states.Its 10%-budget recall gain is 1.62 points (95% interval [0.14, 2.83]); gains at 25% and in test-oracle F1 remain uncertain.Source: §3.4 · p. 7

3.5 RANKING AND REJECTION ON SINGLE-INTERVAL DATASETS

• At test-F1-optimal thresholds, RL reduces native TimeLens2-4B false-alarm rates from 66.04–91.17% to 13.24–21.35% and improves F1 by 1.40–5.63 points.SFT already reduces false alarms and exceeds RL on QVHighlights, so the additional benefit of RL varies by dataset.Source: §3.5 · p. 8, Table 4 · p. 8
• Confidence improves Charades mIoU by 2.02 points when selecting the highest-confidence rather than first raw interval, but cross-dataset consistency remains unresolved.ActivityNet and QVHighlights point estimates are positive, but their intervals include zero.Source: §3.5 · p. 8

Video temporal grounding

• Discriminative temporal-grounding methods learn matching or saliency scores, while generative approaches produce timestamps through language decoding or dedicated interval prediction.Prior generative work includes Huang et al. (2024), Ren et al. (2024), Wang et al. (2024), Wu et al. (2025), and Pramanick et al. (2025).Source: §4 · p. 9

Grounding reliability and rejection

• Prior reliability and rejection work addresses absent queries, irrelevant queries, test-time adaptation, or localization–verification alignment through distinct methods.Examples include RaTSG, OpenVMR, Moment of Untruth, Generalized Video Moment Retrieval, RA-RFT, HRVTG, Event Verification Tuning, and CAVE.Source: §4 · p. 9

Confidence and selective prediction

• Confidence research includes feature-based readouts, selective prediction, calibration, verbalized confidence, sampled-answer uncertainty, and hidden-state entropy probes.For temporal grounding, URPA uses rollout variance to weight adaptation rewards.Source: §4 · p. 9

5 LIMITATIONS

• The confidence head is a lightweight MLP over pooled decoder states, and confidence supervision updates the readout without directly shaping underlying representations.The paper identifies this as one practical realization within a limited architectural and optimization space.Source: §5 · p. 9
• Confidence-based selection cannot recover candidates missing from the pool or suppressed by NMS.Selection depends on the available candidate pool.Source: §5 · p. 9
• Probability calibration and robustness across domains remain future research directions.Source: §5 · p. 9

6 CONCLUSION

• Fixed-pool OMTG-Bench experiments show improved recall under equal return budgets, with the largest gains at tight budgets.Source: §6 · p. 9
• A single decoded response supports ranking, threshold-based selection, and rejection without an inference-time verifier.Source: §6 · p. 9
• The same candidate intervals can serve different return budgets and acceptance requirements through post-decoding policy adjustment.Source: §6 · p. 9

A SUPPLEMENTARY SELECTION RESULTS AND STATISTICAL CHECKS

• Official recall averages each query's matched fraction over all 320 queries after Hungarian matching, which maximizes total IoU before counting pairs at tIoU ≥0.5.The RL reward uses thresholded maximum-cardinality matching, so all-survivor recall is not a formal upper bound.Source: A SUPPLEMENTARY SELECTION RESULTS AND… · p. 13
• Paired source-video bootstrap intervals for Table 5 reuse 3,000 draws, each sampling 287 videos with replacement and retaining their queries.Each scorer is reranked on replicated NMS candidates and receives the same recomputed quota.Source: A SUPPLEMENTARY SELECTION RESULTS AND… · p. 13, Table 5 · p. 13
• At 50% retention (657 intervals), Confidence has 528 matched intervals and 80.37% micro precision, versus Order's 454/69.10% and Verifier's 497/75.65%.Confidence answers 257 of 320 queries, compared with Verifier's 275; the micro counts are diagnostic rather than official query-macro recall.Source: A SUPPLEMENTARY SELECTION RESULTS AND… · p. 13
• At 10% retention, Order tie-breaking is sensitive to random cross-query ties, and the Confidence-versus-Verifier micro-precision interval remains inconclusive.The bootstrap analyses are exploratory, exclude checkpoint-selection and training-seed uncertainty, and have no multiple-comparison correction.Source: A SUPPLEMENTARY SELECTION RESULTS AND… · p. 13
• Single-answer comparisons select either the first valid interval or the highest-confidence interval from the same frozen response, with ties resolved by generation order.The Table 6 comparisons use fixed decoded pools on mixed-query TimeLens runs; both selectors reuse saved responses.Source: A SUPPLEMENTARY SELECTION RESULTS AND… · p. 13, Table 6 · p. 13
• Confidence improves single-answer point estimates across the evaluated datasets, but only the Charades mIoU interval excludes zero.The smaller ActivityNet and QVHighlights gains remain uncertain.Source: Table 6 · p. 13, A SUPPLEMENTARY SELECTION RESULTS AND… · p. 14

B THRESHOLD TRANSFER AND COMPONENT ANALYSIS

• On OMTG, thresholds calibrated to 80% and 90% empirical micro precision yielded 71.71% and 86.89% micro precision, respectively, so neither target transferred.Their OMTG micro recall was 55.75% and 33.33%, respectively.Source: B THRESHOLD TRANSFER AND COMPONENT… · p. 14
• Calibration used 512 source-video-disjoint videos, but this was not a fresh confirmation set because it had appeared in earlier project analyses.Source: B THRESHOLD TRANSFER AND COMPONENT… · p. 14
• In retrospective five-fold video-held-out checks within the same source cohort, only 3/5 folds met the 80% target and 2/5 met the 90% target.Pooled held-out precision was 80.00% and 89.94%, respectively.Source: B THRESHOLD TRANSFER AND COMPONENT… · p. 14

B.1 NMS AND CONFIDENCE SELECTION

• On RL candidates, confidence selection raises precision from 66.25 to 67.23 and F1 from 67.07 to 67.39, while recall changes from 70.51 to 70.10; the paired F1 gain is not established.The paired F1 change is +0.32 points, with a 95% source-video bootstrap interval of [−0.14, 0.76].Source: §B.1 · p. 14, Table 7 · p. 14
• Confidence-sorted NMS lowers tF1@0.5 from 67.39 to 66.61 at the same test-oracle threshold, supporting retention of generation-order NMS.On the inspected OMTG cohort, exploratory grouped threshold selection gives 67.37 versus 66.41, with a paired interval of [−1.76, −0.21] points.Source: §B.1 · p. 14

B.2 FULL SYSTEM AND CANDIDATE-GENERATION REFERENCE

• After NMS and confidence selection, the SFT and RL pools return 3.89 and 3.74 candidates per query, compared with 3.65 for the GT-only reference.The GT-only reference reaches 61.55 Recall@0.5, while raw SFT and RL pools return 11.18 and 8.99 candidates with 67.85 and 73.64 recall; its training setup differs, so it does not isolate candidate-target construction.Source: §B.2 · p. 15, Table 8 · p. 15

B.3 HEAD ADAPTATION ON FIXED CANDIDATES AND DECODER STATES

• At 10% and 25% return budgets, online-minus-SFT recall differences are +1.62 and +1.38 points, with the 10% gain supported and the 25% difference uncertain.Both comparisons use the frozen candidate pool; the paired 95% intervals are [0.14, 2.83] and [−0.59, 3.11].Source: §B.3 · p. 15, Table 9 · p. 15
• The online head has a test-oracle F1 difference of +0.204 points over SFT, but its paired interval includes zero, leaving final-set F1 improvement unresolved.The inference control rescored cached candidates with frozen heads and does not isolate head-training choices or establish online adaptation as necessary for final F1.Source: §B.3 · p. 15

• The SFT model uses GT anchors with at most two alternatives per ground-truth occurrence, up to 32 candidates, and no low-IoU filler.Real alternatives require maximum GT tIoU of at least 0.5 and synthetic alternatives at least 0.6; inconsistent-duration or overfull queries are excluded rather than partially annotated.Source: C REPRODUCING THE REPORTED MODELS · p. 16
• Confidence replay uses a fixed candidate sequence with matched and unmatched proposals, GT completions, and duration-based negatives, while replay contributes zero language-model loss.The replay contains 594,620 candidate occurrences per epoch and reuses continuous verifier scores; synthetic generator alternatives do not receive fabricated verifier scores.Source: C REPRODUCING THE REPORTED MODELS · p. 16
• The offline verifier is a separately fine-tuned Qwen3.5-2B model whose binary-normalized Yes probability supplies soft targets for candidate clips.It also provides the Table 1 baseline, but is not used to construct RL labels or at inference.Source: C REPRODUCING THE REPORTED MODELS · pp. 16–17
• The generator objective weights GT, alternative, and formatting losses as 0.6, 0.3, and 0.1 when alternatives are present, and 0.9 and 0.1 otherwise.Confidence replay uses within-query pairwise ranking loss, weighted 0.25 within head loss; the head loss is also weighted 0.25 in the joint SFT objective.Source: C REPRODUCING THE REPORTED MODELS · p. 17
• From RL step 51, the confidence head receives one BCE update on valid NMS survivors labeled by whether maximum training-GT tIoU reaches 0.5.The head scores sampled intervals before the actor update, and detached features are reused from the actor’s old-log-probability pass.Source: C REPRODUCING THE REPORTED MODELS · p. 17
• Greedy evaluation allows 512 new tokens and 32 unique candidates, with confidence read from interval-token states during the same decode.There is no second video encoding, inference replay, or generated confidence token; invalid boundaries are not clipped, swapped, or invented.Source: C REPRODUCING THE REPORTED MODELS · p. 17

• Each model decodes once; generation-order NMS precedes threshold selection, while single-answer ranking compares raw candidates.OMTG official metrics include all queries in the denominator, malformed outputs count as empty, and invalid boundaries are filtered without timestamp changes.Source: D EVALUATION PROTOCOLS · p. 17
• Reported global F1-optimal thresholds are test-oracle operating points, and native baselines retain their original output rules.For the main OMTG threshold scan, ties in official tF1@0.5 are resolved by tIoU, EtF1, fewer retained candidates, then a higher threshold.Source: D EVALUATION PROTOCOLS · p. 17
• The TimeLens2-4B and 8B baselines are rerun with their original prompt and evaluated on all 320 OMTG queries without NMS or confidence selection.Invalid or incomplete responses remain in the denominator as empty predictions.Source: D EVALUATION PROTOCOLS · p. 18
• Synthetic unmatched queries pair each matched query with a cross-video mismatch, and an incorrect positive counts as both a false positive and a false negative.Synthetic mismatches are not human-verified absences; the protocol permits empty responses and treats nonempty or invalid unmatched responses as false alarms.Source: D EVALUATION PROTOCOLS · p. 18

E QUALITATIVE CASES BY SELECTION REQUIREMENT

• Ranking selects an existing bottle-cap boundary alternative with tIoU 0.905 rather than the earlier C2 at 0.600, without refining boundaries.The highest confidence score supplies the single answer from raw candidates in generation order, with no NMS or thresholding.Source: Figure 4 · p. 19
• A threshold can retain four workout occurrences while removing an extra fragment, raising F1 from 0.889 to 1.000 while recall remains 1.000.The two set-filtering cases use generation-order NMS followed by a global test-oracle threshold; the same threshold supports different answer counts.Source: Figure 5 · p. 20
• In query 132, confidence drops an interval with maximum GT IoU 0.400 even though its overlap with another proposal is only 0.294 and it survives NMS.Query 114 illustrates that thresholding can retain four separate occurrences rather than imposing a top-1 or fixed top-k budget.Source: E QUALITATIVE CASES BY SELECTION… · p. 20
• A threshold can reject both synthetic mismatches while accepting both matched queries, although the generator produced candidates for all four queries.The displayed cases use benchmark-wide test-oracle thresholds, without per-case tuning or NMS; synthetic unmatched labels are not official exhaustively verified negatives.Source: Figure 6 · p. 21
• The top candidate can still be rejected when a query is unsupported: mismatched-query scores are 0.089 and 0.023, versus 0.963 and 0.975 for matched queries.A maximum always exists for a nonempty pool; the threshold determines whether to return that candidate at all.Source: E QUALITATIVE CASES BY SELECTION… · p. 21

E.1 LIMITS OF SELECTION ON A FIXED CANDIDATE POOL

• Selection cannot recover an occurrence absent from decoded proposals, and NMS can remove a usable boundary before confidence filtering.In Q149 no candidate overlaps the first occurrence, capping subset-selection recall at 2/3; in Q58, NMS suppresses a higher-confidence, higher-GT-IoU candidate.Source: §E.1 · p. 22