- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
Autoregressive video temporal grounding suffers from systematic endpoint accuracy asymmetry. The authors link this issue to error propagation from predicted start timestamps. They formulate start errors as sunk loss and remaining end errors as conditional regret. FinishStrong addresses this problem through two-stage post-training. Online-start SFT trains end predictions using model-generated start prefixes. Conditional-tree GRPO assigns end advantages through within-branch comparisons. FinishStrong-9B achieves 61.7% IoU on TimeLens-Bench and reduces conditional end MAE by 36.5% over standard SFT.
Critical strengths
-
Clear motivation. Controlled prefix interventions reveal substantial error propagation from start to end predictions.
-
Effective training framework. FinishStrong combines online prefix exposure with conditional credit assignment. FinishStrong achieves 61.7% overall IoU. Ablations isolate the contributions of both training stages and conditional credit assignment.
Critical weaknesses
- Limited generalization evaluation. Experiments focus on single-interval temporal grounding. Generalization to multi-moment grounding and general video understanding remains unexplored.
- Limited handling of severe start errors. Online-start SFT excludes starts beyond the GT end, while Conditional-tree GRPO provides zero end advantages in such cases. The impact of these errors remains unexplored.
- Insufficient empirical validation of the GRPO failure mechanism. The negative-advantage issue is theoretically motivated, but its prevalence during training is not quantified.
Questions for authors
-
Can FinishStrong extend to multi-segment temporal grounding, such as OMTG-Bench (ICML26) ? Does it preserve general video understanding performance on benchmarks such as Video-MME and Video-MME-v2?
-
How does FinishStrong handle predicted starts at or beyond the GT end, where positive IoU is unattainable? Can the authors analyze end prediction errors and interval duration bias across different start-error magnitudes and directions, compared with baseline methods?
5 comments
Response to Weakness 2 and Question 2
W2 & Q2-Part1: How does FinishStrong handle predicted starts at or beyond the GT end?
Thank you for highlighting this important case. We would like to clarify that both training stages explicitly account for these severe errors through complementary mechanisms: Online-start SFT retains GT start supervision, while Conditional-tree GRPO discourages such predictions through start-level credit assignment.
Online-start SFT. As formulated in Sec. 4.1 (Eq. (4)), when the predicted start is at or beyond the GT end, only the conditional end loss is masked. The GT start loss remains active for every sample, so these samples still provide supervision for learning the correct start.
Conditional-tree GRPO. Such starts remain in sampling and optimization. For a branch with , no valid end can achieve positive IoU. All end continuations therefore receive zero rewards and non-positive end advantages. However, the start advantage is
Thus, these zero-overlap starts are penalized relative to starts that achieve positive rewards (Sec. 4.2, Eqs. (5)-(6)).
To examine the practical impact, we track the frequency of these severe errors throughout both training stages.
Table R-3. Frequency of predicted starts at or beyond the GT end during training.
| Training Phase | Online-start SFT | Conditional-tree GRPO |
|---|---|---|
| Early | 4.991% | 4.357% |
| Middle | 5.935% | 2.073% |
| Final | 5.721% | 1.851% |
The rate remains relatively stable during Online-start SFT (4.991% to 5.721%). A likely explanation is that GT start supervision promotes the annotated start while broadly suppressing incorrect alternatives. Since most incorrect starts fall before the GT end, this correction mainly addresses these more frequent errors, with less targeted pressure on starts at or beyond the GT end. In contrast, Conditional-tree GRPO directly discourages these zero-overlap starts through start-level credit assignment, reducing their frequency from 4.357% to 1.851% (57.5% relative reduction).
Q2-Part2: How do end prediction errors and interval duration bias vary across different start-error magnitudes and directions?
Thank you for this insightful question. Sec. 5.3 (Table 2 and Fig. 4) evaluates conditional end prediction under identical, controlled start prefixes. To complement these tests, we analyze each model's own predictions under greedy decoding, grouped by start-error magnitude and direction.
For a predicted interval and a ground-truth interval , we define and the normalized start error . We report End MAE and the mean of the per-example normalized duration bias . Positive and negative bias indicate overestimation and underestimation of interval duration, respectively.
Table R-4. End prediction errors and interval duration bias across start-error regimes.
| Start-error regime | Model | N (%) | End MAE (s) ↓ | Mean Duration Bias |
|---|---|---|---|---|
| Early () | TimeLens | 1904 (20.25%) | 12.538 | 2.7700 |
| Video-OPD | 1778 (18.91%) | 13.331 | 2.7965 | |
| FinishStrong | 1365 (14.52%) | 11.902 | 2.3836 | |
| Slightly Early () | TimeLens | 1740 (18.50%) | 6.670 | 0.4021 |
| Video-OPD | 1678 (17.84%) | 5.263 | 0.4699 | |
| FinishStrong | 1705 (18.13%) | 4.525 | 0.3306 | |
| Near-GT / Late () | TimeLens | 5212 (55.42%) | 7.732 | 0.4838 |
| Video-OPD | 5200 (55.30%) | 5.221 | 0.4433 | |
| FinishStrong | 5774 (61.40%) | 4.714 | 0.2866 | |
| Severely Late () | TimeLens | 229 (2.44%) | 4.834 | 0.4749 |
| Video-OPD | 278 (2.96%) | 3.408 | 0.1409 | |
| FinishStrong | 390 (4.15%) | 2.605 | 0.0114 | |
| At or Beyond GT End () | TimeLens | 319 (3.39%) | 39.273 | 2.7503 |
| Video-OPD | 470 (5.00%) | 44.951 | 1.9653 | |
| FinishStrong | 170 (1.81%) | 41.154 | 1.8254 |
FinishStrong achieves the lowest End MAE across all four recoverable start-error regimes (). Compared with the strongest baseline in each regime, it reduces End MAE by 5.1%, 14.0%, 9.7%, and 23.6% for early, slightly early, near-GT/late, and severely late starts, respectively. It also exhibits a smaller positive mean duration bias in every regime, indicating reduced average overestimation of interval duration. Under severely late starts, this mean bias is 0.0114.
For starts at or beyond the GT end (), end errors remain large and no continuation can get positive IoU. FinishStrong produces such starts less frequently: 1.81%, compared with 3.39% for TimeLens and 5.00% for Video-OPD.
Response to Weakness 3
W3: Insufficient empirical validation of the GRPO failure mechanism. The negative-advantage issue is theoretically motivated, but its prevalence during training is not quantified.
Thank you for this insightful suggestion. We provide two complementary quantitative analyses of actual training rollouts to examine how end quality relates to the advantages assigned during training.
1) Shared trajectory-level advantages can penalize good ends following poor starts.
Standard GRPO assigns the same trajectory-level advantage to both start and end tokens. A poor start can lower the overall interval reward and lead to a negative advantage even when the subsequent end is accurate given that start. Applying this shared negative advantage to both decisions therefore also penalizes the conditionally good end.
We examine actual training rollouts across cIoU intervals. Higher cIoU indicates that the end realizes a larger fraction of the overlap attainable from its committed start. Table R-5 reports the rollout distribution and the negative-advantage frequency within each cIoU bin. GRPO uses joint trajectory advantages, while FinishStrong uses within-start end advantages.
Table R-5. Negative-advantage frequency across cIoU intervals.
| cIoU | GRPO | FinishStrong | ||
|---|---|---|---|---|
| Rollouts (%) | Negative Advantage (%) | Rollouts (%) | Negative Advantage (%) | |
| 13.21 | 97.08 | 12.14 | 55.75 | |
| 0.53 | 100.00 | 0.78 | 88.00 | |
| 1.45 | 96.67 | 1.61 | 88.46 | |
| 1.98 | 90.24 | 1.96 | 79.37 | |
| 3.33 | 88.41 | 3.72 | 68.33 | |
| 4.05 | 80.95 | 4.22 | 65.44 | |
| 8.73 | 75.69 | 9.16 | 67.12 | |
| 9.40 | 77.95 | 10.74 | 57.23 | |
| 10.70 | 70.27 | 10.18 | 43.90 | |
| 10.70 | 67.12 | 9.47 | 36.07 | |
| 5.83 | 61.98 | 5.84 | 27.66 | |
| 30.09 | 51.76 | 30.18 | 2.78 | |
| Total | 100.00 | -- | 100.00 | -- |
Notably, even among nearly conditionally optimal end predictions (, i.e., good ends), 51.76% receive negative advantages under GRPO and are undesirably penalized along with their imperfect starts during training. Under FinishStrong, this proportion drops to only 2.78%. By comparing end predictions following the same start, Conditional-tree GRPO substantially reduces the conflict between joint trajectory credit and conditional end quality.
2) Advantages relative to end-localization error.
We also examine whether the assigned advantages encourage more accurate ends. Let and let be the mean error within the comparison set. We measure two types of disagreement:
- L: the set contains an end with lower-than-average error but a negative advantage, discouraging a relatively accurate end.
- H: the set contains an end with higher-than-average error but a positive advantage, encouraging a relatively inaccurate end.
Percentages are calculated over the comparison sets listed below. Both FinishStrong rows use its actual within-start end advantages.
Table R-6. Frequency of disagreement between end error and advantage sign (%).
| Method | Error-comparison scope | L | H | Both |
|---|---|---|---|---|
| GRPO | Flat rollout group | 76.06 | 41.33 | 32.43 |
| FinishStrong | Full rollout tree | 38.95 | 24.48 | 11.54 |
| FinishStrong | Same-start branch | 3.60 | 2.05 | 0.99 |
Under GRPO, 76.06% of groups contain a relatively accurate end that receives a negative advantage, while 41.33% contain a relatively inaccurate end that receives a positive advantage; both occur in 32.43% of groups. These opposing training signals are substantially less frequent in the analyzed FinishStrong trees, with corresponding frequencies of 38.95%, 24.48%, and 11.54%. Within its same-start branches—the comparison sets used to assign end advantages—the frequencies are 3.60%, 2.05%, and 0.99%.
The controlled G/H ablation in Table 3 further isolates the effect of end-advantage assignment: within-start end advantages improve Free IoU by 1.9 points and conditional cIoU by 2.45 points, while reducing conditional End MAE by 0.73s. These results complement the rollout analyses and support the benefit of conditional end-advantage assignment.
Thank you again for your constructive feedback. Please let us know if any concerns remain. We would be happy to provide further clarification or information.
Response to Weakness 1 and Question 1
We sincerely thank you for the thoughtful review and constructive suggestions. We appreciate your recognition of the clear motivation established through controlled prefix interventions and the effectiveness of our training framework. Below, we address your comments point by point. All additions and revisions are included in the revised manuscript and its appendices.
Thank you for this insightful suggestion! Our additional experiments show that FinishStrong extends effectively to multi-moment grounding while largely preserving general video understanding capabilities.
1) Generalization to multi-moment grounding.
Following your suggestion, we reproduce OMTG-4B (ICML 2026) [1] on OMTG-Bench and apply FinishStrong to this model, extending our evaluation beyond single-interval temporal grounding.
Interestingly, we observe a similar endpoint asymmetry in multi-moment grounding as in the single-moment: our reproduced OMTG-4B has an End MAE of 4.36s, 27.5% higher than its Start MAE of 3.42s. Applying FinishStrong reduces End MAE by 0.82s (18.8%) and Start MAE by 0.25s (7.3%), while improving all six grounding metrics by 1.84-2.91 percent. These results demonstrate improved localization of both boundaries and support the applicability of FinishStrong to multi-moment grounding.
Table R-1. Results on OMTG-Bench.
2) Preserving general video understanding capabilities.
We further evaluate FinishStrong-9B on 15 video understanding benchmark settings, covering video perception, temporal reasoning, spatial understanding, and long-video understanding, under both non-thinking and thinking inference.
Overall, FinishStrong remains broadly comparable to its base model Qwen3.5-9B across these evaluations under both inference settings. It also improves several benchmarks, including VideoMMMU (+2.34 points), Video-MME v2 (+0.70 to +1.48 points), VSI-Bench (+1.15 points), TempCompass (MC) (+1.01 points), and MLVU (MCQ) (+0.57 points).
Table R-2. General video understanding performance of Base and FinishStrong.
[1] Qi Xu et al. Towards One-to-Many Temporal Grounding. ICML, 2026. Paper