- 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?