- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes Student-Curriculum Coupling (SCC) for on-policy distillation in temporal video grounding. It selects 100 teacher-qualified Anchor examples that the initial student handles well and 900 Frontier examples that it handles poorly. During training, a sampled prediction receives teacher scoring and distillation only when its temporal IoU is below 0.7. The paper evaluates a Qwen3-VL-8B student on three TimeLens benchmarks. It reports a 5.1% relative mean-recall improvement over Video-OPD with 60.0% fewer selected examples and 50.4% shorter training time. The appendices examine curriculum composition, routing, normalization, teacher choice, and rollout count.
Critical strengths
-
The paper demonstrates a useful accuracy-efficiency trade-off with a simple implementation. Holding the AF curriculum fixed, Table 6 improves mean recall at every threshold while reducing training time from 1:11:20 to 1:00:26. Table 2 explains the saving through 492 teacher-scored routes instead of 1,004. This comparison gives more direct evidence for routing than the headline comparison against the larger TVDF curriculum.
-
The experimental decomposition is substantial. Tables 3, 11, 14, 15, and 16 examine curriculum composition, the competence threshold, the Anchor/Frontier ratio, curriculum size, and normalization. These details make the practical contribution easier to assess and avoid overstating an unbiased reverse-KL interpretation for SCC itself.
Critical weaknesses
-
The claimed stabilization mechanism is not demonstrated by the reported schedule. Appendix B.2 and Tables 4-5 present each unique example once, apart from four final Frontier repeats. All 100 Anchors therefore have only one training presentation. An Anchor can fail its initial competence criterion at that presentation, as Section 3.3 defines, but the experiment does not then revisit it to establish recovery. Figure 4 also does not state whether successive competence values evaluate a fixed set or different scheduled examples. With one stochastic rollout per example, changes can reflect both sampling variation and changes in the evaluated cohort. The endpoint gains are useful, but they do not establish the stronger acquisition, suspension, and restoration account in Section 5.
-
The static control does not fully isolate the value of current-student routing. Appendix B.12 matches the total and per-step supervision counts, but selects 49 Anchors and 443 Frontiers. SCC instead supervises 11 Anchors and 481 Frontier presentations in Table 14. The methods thus differ in their active curriculum composition as well as their routing rule. A static gate based on initial student competence, or a random mask matching SCC's active region counts, would be more diagnostic. Repeated seeds are also absent. This matters because some gains over uniform AF training are small, including 0.21 points on Charades [email protected] in Table 11. The current results support the overall recipe more strongly than the specific causal role assigned to evolving supervision necessity.
Questions for authors
-
How are the competence curves in Figure 4 computed and smoothed? Are the same examples evaluated at successive checkpoints? Is there evidence that an Anchor's grounding performance recovers after its corrective update, beyond the initial-to-current threshold crossing?
-
Does SCC still outperform a static initial-student gate and a random routing mask that match its per-step counts and active Anchor/Frontier composition? Please report variation across independent runs for this comparison and uniform OPD on AF.
-
How many candidate examples require teacher and initial-student inference to construct AF? Table 6 measures training after curriculum construction. What are the shared and additional screening costs relative to TVDF, and how do they affect the total cost of obtaining a trained model?
7 comments
[Continuation of W1 & Q1]
3. Controlled Evaluation of Anchor Supervision
We added an Anchor-muted control that retains the same 100 Anchors, 900 Frontiers, and presentation schedule but disables teacher scoring and OPD for all Anchors. Frontier routing continues to use the control model's current predictions. Student initialization, teacher, normalization, and the remaining training settings are unchanged.
Downstream grounding performance. Table 1 reports mean mIoU across the three benchmarks and mean recall across the nine recall metrics. Values are percentages; bold marks the higher score at each checkpoint.
Table 1. Effect of disabling Anchor supervision on downstream grounding performance.
| Checkpoint | Setting | Mean mIoU (%) | Mean recall (%) |
|---|---|---|---|
| 20 | SCC | 53.48 | 54.14 |
| Anchor-muted SCC | 53.64 | 54.25 | |
| 40 | SCC | 55.71 | 55.65 |
| Anchor-muted SCC | 54.81 | 54.57 | |
| 59 | SCC | 55.66 | 55.64 |
| Anchor-muted SCC | 54.54 | 54.22 | |
| 79 | SCC | 55.71 | 55.80 |
| Anchor-muted SCC | 54.38 | 54.04 |
Disabling Anchor supervision weakens later-stage grounding performance. SCC maintains higher aggregate performance at checkpoints 40, 59, and 79 and outperforms the control on all 12 benchmark metrics at the final checkpoint.
Retention on the fixed Anchor set. We additionally evaluate the same 100 Anchors from the AF curriculum at checkpoints 20, 40, 59, and 79, using five fresh rollouts per example under unchanged sampling settings. Table 2 reports mean tIoU and the percentage satisfying , computed over 500 evaluation rollouts per checkpoint for each model. These evaluations do not contribute to training.
Table 2. Grounding performance on the fixed set of 100 Anchors.
| Checkpoint | Setting | Mean tIoU | Success rate (%) |
|---|---|---|---|
| 20 | SCC | 0.9373 | 95.2 |
| Anchor-muted SCC | 0.9258 | 93.4 | |
| 40 | SCC | 0.9388 | 94.0 |
| Anchor-muted SCC | 0.9148 | 91.4 | |
| 59 | SCC | 0.9324 | 94.2 |
| Anchor-muted SCC | 0.9108 | 91.0 | |
| 79 | SCC | 0.9326 | 94.4 |
| Anchor-muted SCC | 0.9089 | 90.8 |
SCC achieves higher mean tIoU and success rate at all four checkpoints. From checkpoints 20 to 79, its success rate decreases by 0.8 percentage points, compared with 2.6 points for the Anchor-muted control. At the final checkpoint, SCC achieves 94.4% versus 90.8% success.
Together, these comparisons support the contribution of Anchor supervision to sustained downstream performance and retention on the selected Anchor set during continued training. The fixed-Anchor evaluation holds cohort composition constant and tracks performance across checkpoints, extending the evidence beyond the threshold crossings used for routing.
These analyses assess performance retention across checkpoints but do not fully isolate the contribution of corrective supervision on individual Anchors to subsequent performance recovery. We have clarified this scope in Section 7 (lines 507–515). We have also specified the computation and smoothing procedure in the Figure 4 caption (lines 428–431), clarified the interpretation of the training curves in Section 5 (lines 409–416), and incorporated the controlled evaluations in Section 5 (Table 4, lines 433–454) and Appendix B.16 (Tables 25–26, lines 1418–1468).
W2 & Q2: Matched Static Controls and Multi-Seed Robustness
Response: The original static control did not fully separate supervision composition from the routing rule. We therefore added both requested static controls and evaluated SCC, the two controls, and uniform Video-OPD on AF across three random seeds.
1. Controls with Matched Supervision Allocation
Both controls use the same 1,000 AF examples and 1,004-presentation schedule as SCC. They match the total supervision budget and the number of supervised Anchor and Frontier presentations at each step. The reference SCC allocation contains 492 supervised presentations: 11 Anchors and 481 Frontiers. Student initialization, teacher, optimization settings, and loss normalization are unchanged.
- Matched Random Routing (AF) randomly selects the required number of presentations within each step and curriculum region.
- Initial-Score Routing (AF) selects the required number of presentations with the lowest offline initial-student tIoU scores within each step and region, implementing static selection based on initial competence under the same supervision quotas.
Both routing masks are fixed before training. Only SCC's aggregate supervision counts are reused; neither control uses its example-level routing decisions or current-student scores. Both controls generate fresh on-policy trajectories during training.
These comparisons hold supervision budget, timing, and Anchor–Frontier composition fixed while testing whether current-student assessment improves the selection of supervised presentations.
2. Three-Seed Results
All AF results in Table 3 report mean ± sample standard deviation across three independent training runs at checkpoint 79. Published single-run Video-OPD (TVDF) results are included for reference. Values are percentages; bold indicates the highest mean for each metric.
Table 3. Matched static controls and three-seed TVG performance.
| Benchmark | Method | mIoU | [email protected] | [email protected] | [email protected] |
|---|---|---|---|---|---|
| Charades-TimeLens | Video-OPD (TVDF) | 52.0 | 73.1 | 45.8 | 32.4 |
| Video-OPD (AF) | 52.09 ± 0.33 | 73.26 ± 0.06 | 47.00 ± 0.42 | 32.30 ± 0.60 | |
| Matched Random Routing (AF) | 51.80 ± 0.33 | 72.64 ± 0.39 | 46.48 ± 0.56 | 31.96 ± 0.39 | |
| Initial-Score Routing (AF) | 52.00 ± 0.38 | 73.17 ± 0.57 | 46.79 ± 0.70 | 32.07 ± 0.68 | |
| SCC (AF) | 52.78 ± 0.11 | 73.82 ± 0.12 | 48.15 ± 0.25 | 33.09 ± 0.25 | |
| ActivityNet-TimeLens | Video-OPD (TVDF) | 47.3 | 60.5 | 45.6 | 35.8 |
| Video-OPD (AF) | 49.86 ± 0.51 | 63.61 ± 0.67 | 47.96 ± 0.40 | 37.58 ± 0.51 | |
| Matched Random Routing (AF) | 49.33 ± 0.72 | 63.08 ± 0.94 | 47.25 ± 0.89 | 37.02 ± 0.78 | |
| Initial-Score Routing (AF) | 49.64 ± 0.52 | 63.30 ± 0.79 | 47.64 ± 0.39 | 37.34 ± 0.59 | |
| SCC (AF) | 50.81 ± 0.17 | 64.88 ± 0.38 | 49.15 ± 0.07 | 38.65 ± 0.36 | |
| QVHighlights-TimeLens | Video-OPD (TVDF) | 61.0 | 73.8 | 60.3 | 50.4 |
| Video-OPD (AF) | 62.34 ± 0.68 | 75.25 ± 0.72 | 62.41 ± 1.06 | 52.17 ± 0.81 | |
| Matched Random Routing (AF) | 61.80 ± 0.71 | 74.89 ± 1.08 | 61.41 ± 0.67 | 51.61 ± 0.84 | |
| Initial-Score Routing (AF) | 62.00 ± 0.86 | 75.12 ± 0.79 | 61.86 ± 1.02 | 51.31 ± 1.07 | |
| SCC (AF) | 63.79 ± 0.09 | 76.99 ± 0.37 | 64.18 ± 0.33 | 54.27 ± 0.36 |
SCC achieves higher mean performance on all 12 metrics than both matched static controls and uniform Video-OPD on AF. Averaged across the nine recall metrics, SCC improves mean recall by 1.87 percentage points over Matched Random Routing, 1.62 points over Initial-Score Routing, and 1.29 points over uniform Video-OPD (AF).
For the highlighted Charades-TimeLens [email protected], SCC achieves 33.09 ± 0.25, compared with 32.30 ± 0.60 for uniform Video-OPD (AF), a mean improvement of 0.79 percentage points.
3. Interpretation and Manuscript Revision
The gains persist when supervision budget, timing, and Anchor–Frontier composition are matched. The random control addresses the alternative explanation that SCC benefits simply from allocating more supervision to Frontiers. The initial-score control additionally tests whether initial difficulty can provide an equally effective fixed selection. SCC's stronger results in both comparisons support the value of using current student performance to determine which presentations receive supervision.
We have incorporated the control protocols and three-seed results into Appendix B.13 (Table 23, lines 1326–1376).
Dear Reviewer G7E7,
We appreciate your careful review and constructive feedback. Below, we clarify the interpretation of online training dynamics and Anchor reactivation, and present additional evidence from controlled evaluations of capability retention, matched static routing, multi-seed robustness, and construction-inclusive efficiency.
W1 & Q1: Anchor Stabilization and Competence Measurements
Response: Figure 4 summarizes online training rollouts on the examples presented at each step. We clarify its computation and interpretation below and provide an Anchor-muted control evaluated on both the TVG benchmarks and a fixed Anchor set to assess performance retention.
1. Computation and Smoothing of Figure 4
At each training step, we compute the current student's rollout tIoU for all presented examples, including both supervised and suspended presentations. Each competence point pools presentations from the current step and the preceding four steps. The Overall, Anchor, and Frontier curves report mean tIoU over their corresponding presentations within this window. Each presentation contributes equally, so steps contribute in proportion to their presentation counts. We provide the unsmoothed curves in Appendix B.17 (Figure 6, lines 1471–1489).
The curves do not repeatedly evaluate a fixed cohort, and the checkpoint guides are visual reference marks only. Figure 4 describes online grounding scores and supervision activity; changes can reflect student evolution, differences in the examples presented, and rollout variability.
2. Meaning of Anchor Reactivation
In our approximately single-pass implementation, reactivation is defined relative to the offline initial assessment. Each Anchor has one training presentation. It initially meets the competence criterion and receives teacher scoring and OPD if its current rollout falls below τS=0.7 at that presentation. Reactivation therefore denotes an initial-to-current threshold crossing and does not require an earlier online presentation. This definition is stated in Section 3.3 (lines 242–247) and Appendix A.3 (lines 877–882).
A single sampled failure does not establish capability loss. The criterion identifies an observed task deficit that triggers supervision. Appendix B.14 (Figure 5, lines 1377–1400) compares this single-rollout decision with three- and five-rollout assessments using mean tIoU. Additional rollouts provide no consistent accuracy advantage across benchmarks and increase training time, supporting the practical single-rollout choice.