acceptodds
Under review as a conference paper at ICLR 2027

Video-OPD++: Geometry-Calibrated Corrective On-Policy Distillation for Temporal Video Grounding

Abstract

On-policy distillation has recently emerged as an effective post-training paradigm for temporal video grounding, combining student-induced training states with dense teacher feedback. However, each realized Video-OPD update remains centered on the token sampled by the student, without explicitly turning teacher-supported timestamp alternatives that could improve localization into training targets. The interval structure of temporal video grounding makes such alternatives directly evaluable: with the suffix fixed, a timestamp can be replaced and the resulting interval scored by temporal IoU (tIoU). Our analysis shows that, for more than three quarters of student trajectories, the teacher support contains an alternative timestamp action that improves tIoU. Motivated by this observation, we propose Video-OPD++, a geometry-calibrated corrective on-policy distillation framework. At each student-induced timestamp state, Video-OPD++ evaluates teacher-supported candidate actions using annotation-derived tIoU utilities under fixed-suffix interventions and reweights the teacher prior to construct a corrective distribution aligned with localization quality. This distribution replaces the original local sampled-action reward at timestamp positions, while a discounted future residual preserves feedback from subsequent token discrepancies. Across three temporal video grounding benchmarks, Video-OPD++ improves average mIoU over Video-OPD by 2.3 points. These results suggest that incorporating temporal video grounding geometry into the distillation target helps the student learn more accurate temporal boundaries.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Temporal Video Grounding, On-Policy Distillation, Multimodal Large Language Models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_YenqOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the post-training setting of Temporal Video Grounding (TVG) and identifies a sampled-action correction bottleneck in Video-OPD: although the method provides step-specific dense teacher feedback, each realized update is centered only on the single token sampled by the student and does not rank alternatives within the complement; a large number of improving alternative timestamp actions within the teacher support never enter the optimization objective (and these alternatives are already saturated in the high-probability region of the teacher at K = 8); meanwhile, the supervision signal is highly concentrated.

To address this, the paper proposes Video-OPD++, a geometry-calibrated corrective on-policy distillation framework: at the token positions of the first valid start–end timestamp pair, and conditioned on the student-induced states, the top-K candidates are taken deterministically from the teacher distribution; with the student suffix held fixed, a single-token substitution is performed, and the tIoU between the parsed interval of the modified completion and the annotated interval serves as the candidate utility (candidates that fail to parse or yield invalid intervals are excluded). The utilities are z-score standardized within each valid candidate support and used as the exponent for a multiplicative tilting of the teacher prior, yielding a distributional corrective target.

Main contributions: (1) identifying and quantifying the sampled-action correction bottleneck in on-policy distillation; (2) proposing the geometry-calibrated corrective distribution, which converts pointwise action evaluation into task-aligned distributional supervision; (3) proposing the token-partitioned training objective (corrective supervision replaces the sampled-action reward, complemented by a discounted future residual); (4) providing a KL-regularized closed-form interpretation of the target together with a complete gradient analysis.

Across three TimeLens benchmarks, the average mIoU of Video-OPD++ shows consistent improvements over OP-RKD, OP-FKD, GRPO, and Video-OPD.

Critical strengths
  • A systematic diagnosis that quantifies the sampled-action correction bottleneck in on-policy distillation. (i) The nominally step-wise dense sampled supervision is in fact concentrated on very few positions (effective supervised token count N_eff = 2.82; the three largest completion positions account for approximately 89% of the absolute sampled log-ratio mass); (ii) improving alternatives are widespread yet unused within the teacher support (in 76.3% of trajectories, a single-token valid substitution exists that raises tIoU from 0.385 to 0.522, with correction coverage of 62.0% / 72.7% at the start / end timestamp positions, respectively).

  • The verifiable tIoU reward is sunk from the trajectory level to the token level to reweight the teacher candidate scores, yielding a more precise learning target: each candidate's grounding quality directly modulates its share of the corrective distribution, so the student learns task-aligned preferences among plausible timestamp actions rather than a pointwise sampled log-density ratio.

Critical weaknesses
  • Limited methodological novelty: the core operations are a composition of standard components—renormalizing the teacher distribution over the candidate set, exponential utility reweighting, and distribution matching. Verifiable tIoU rewards have already been adopted at the trajectory level by RL baselines (Time-R1, TVG-R1); this work mainly sinks the same signal to the token level and fuses it with the teacher prior.

  • Several hyperparameters (λ, η, the loss coefficients) are not ablated.

  • Validation is confined to a single model family (Qwen3-VL-8B student / Qwen3-VL-32B teacher), which limits the external validity of the conclusions.

Questions for authors
  • Are the main results and key ablations single runs or multi-seed averages? Please provide variances or confidence intervals for the comparisons of γ = 0 vs. γ = 4, "replace vs. retain the sampled-action reward," and the Charades gain. If these differences fall within seed-level noise, the central claim that "geometry calibration is the primary performance lever" would not hold.

  • Does the reproduced Video-OPD match the officially reported numbers?

  • Selection protocol for γ: Was γ = 4 selected on an independent validation set, or chosen as the best value on the three test benchmarks? In the latter case, the main results suffer from tuning-on-test leakage.

  • Video-OPD++ computes candidate tIoU utilities from the annotated temporal intervals; do GRPO, OP-RKD, OP-FKD, and Video-OPD have equivalent access to annotations (e.g., is the trajectory reward of GRPO likewise derived from tIoU)? If the baselines do not consume annotations while the proposed method does, how much of the 2.3-point improvement should be attributed to "additional annotated supervision" versus "improved objective form"? The γ = 0 ablation (+0.2 only) suggests the former is the main factor—please clarify whether the gains belong to the objective design or to the introduction of the annotation signal.

  • Is there any video or event overlap between the 2,500 TVDF-selected training examples (sourced from HiREST, QerYD, DiDeMo, etc.) and the three TimeLens test benchmarks? Is the ground truth used for candidate utilities from the same source as the evaluation annotations (the TimeLens-corrected annotations)?

  • How were λ = 0.25 and η = 1.0 determined? Were there unreported pilot sweeps? If the results are sensitive to λ, are the conclusions about the future-residual component (especially the counterexample on Charades) robust?

Reviewer_258aOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, video OPD++is proposed to solve the problem of on policy decomposition in the task of temporal video grounding. The author believes that the conventional on policy distillation method mainly guides the student model based on the token level probability distribution of the teacher model, but for the timestamp token, the higher probability does not necessarily mean that the positioning result is more accurate. Therefore, token level distillation alone may not make full use of the geometric quality differences between timestamp candidates. To solve this problem, the author proposes a geometry calibrated corrective on policy disintegration, which includes:

  1. teacher goal distribution of geometric calibration: replace the timestamp candidates by fixing the subsequent tokens, analyze the generated time interval, calculate the positioning utility of the candidates by using the temporary IOU between it and the real interval, and then integrate the utility into the teacher distribution.
  2. future residual correction: combine the local geometric utility of the timestamp position with the future reward residual, so that the training target takes into account the positioning quality of the current timestamp candidate and the impact of the subsequent generation process.
  3. selective correction: it mainly adjusts the distillation target of the timestamp position, and the other token positions still retain the original video OPD training mode. The experiment was carried out on three datasets: charades timelens, activitynet timelens and qvhighlights timelens. Compared with video OPD, the Miou of video OPD++on the three datasets increased by 0.6%, 3.3% and 3.0%, respectively, with an average increase of about 2.3%.
Critical strengths
  1. The method design is clear and interpretable. The author introduces the geometric utility into the teacher distribution, and controls the deviation of the target distribution from the original teacher distribution through KL regularization. To a certain extent, the design takes into account the geometric feedback related to the knowledge and task of the teacher model.
  2. The experiment covers multiple data sets, and the results are generally positive. The author reported the improvement of Miou relative to video OPD on three datasets, and the improvement of activitynet timelens and qvhighlights timelens was more obvious.
Critical weaknesses
  1. The novelty of reward weighted distillation has not been fully confirmed. The central goal can be interpreted as the teacher reweighting before the task specific utility. This paper does not fully prove that this special formula and the goal of token Division provide substantial advantages. Simpler alternatives, such as reward weighted cross entropy, utility based candidate ranking, or comparable training budgets under direct time IOU optimization, are available.
  2. The effectiveness of fixed suffix intervention has not been fully verified. Replacing the timestamp token while retaining the original suffix may produce sequences that will not occur naturally under the autoregressive strategy after the intervention. Therefore, the resulting temporary utility may differ from the utility obtained by regenerating the suffix. The discounted future residuals cannot directly prove the reliability of fixed suffix candidate ranking. An experiment comparing the results of fixed suffix sorting and regenerated suffix will greatly strengthen this paper.
Questions for authors

How does the proposed objective compare with reward-weighted cross-entropy, hard best-candidate supervision, and direct temporal-IoU optimization under the same rollout and compute budgets?

Reviewer_8fVTOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This study explores the Video-OPD and finds a limitation in how it supervises timestamp generation. For example, Video-OPD uses dense feedback of the sampled tokens, but it does not explicitly show which token should be preferred, so the student model does not make good use of the teacher model information. In contrast, this study optimizes Video-OPD by using the teacher model to provide candidate tokens and evaluate their probabilities. The tIOU between the candidate and GT is used to fuse and supervise the student model for learning. The experimental results show that this method is improved on three different datasets compared to the previous method.

Critical strengths

1.The study finds a limitation of existing methods: while they provide dense feedback on sampled tokens, they fail to explicitly indicate which alternative tokens should be preferred. So student model are unable to effectively use the information provided bt teacher model. 2.To address this limitation, the authors propose a relatively simple method that combines teacher-supported candidate tokens with tIoU to guide student model training, and the experimental results are promising.

Critical weaknesses

1.While this study utilizes GT during the training, it does not report the effects of directly using GT timestamps to post-train the student model. Consequently, it is not clear whether the performance improvement stems from the teacher model or the GT labels. 2.The experimental results lack a baseline weighted by GT tIOU, which can help clarify benefits of incorporating teacher probabilities.

Questions for authors

I think additional experimental results are needed to p[rove the contributions of this study, including:

1.How does the student model perform when trained using GT timestamps directly? 2.What is the impact of constructing candidate weights based solely on tIoU? 3.Could the authors clarify whether the performance gains stem from the proposed method, or primarily from the inclusion of GT?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.