acceptodds
Withdrawn from ICLR 2027

A Bimodal Perspective on Video Grounding

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Video temporal grounding requires localizing events within videos based on natural language queries. Although temporal Intersection over Union (tIoU) is the de facto metric for both evaluation and Reinforcement Learning from Verifiable Rewards (RLVR), we show that applying it uniformly across event durations creates two previously underexplored failure modes. First, for short events, small absolute boundary shifts cause disproportionately large changes in tIoU, so temporally close predictions can receive misleadingly low scores. Second, when tIoU is directly optimized as an RLVR reward, a model can increase its expected reward by inflating predicted intervals; consequently, tIoU-based RLVR improves long-event localization while degrading short-event performance. These observations motivate a duration-aware framework that treats short and long events as distinct temporal regimes. We evaluate them separately with distance-based and overlap-based criteria, respectively. For training, we combine distance-IoU with boundary accuracy to provide smoother supervision for short events while preserving effective overlap supervision for long events and discouraging interval inflation. We further introduce a multi-rollout query sampler that uses duration-aware success criteria to retain hard-yet-solvable examples, together with BiVTG-100K, a refined dataset with improved coverage and annotation quality for short events. Experiments on the manually re-annotated Charades, ActivityNet, and QVHighlights benchmarks show that the proposed framework consistently improves duration-aware grounding and mitigates the short-event failures of tIoU-based RLVR.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video understanding, Video Temporal Grounding

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_zALyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the limitations of tIoU for both short and long events in video temporal grounding. The authors propose a duration-aware evaluation and sample selection strategy, and combines dIoU with boundary rewards for RLVR training. Experiments show the superior performance on several VTG benchmarks.

Critical strengths
  1. The paper provides a clear analysis of the different effects of tIoU on short and long events. The motivation is clear and reasonable.

  2. The method achieves superior performance on several benchmarks. The paper also provides extensive ablation studies on different reward functions and sampling strategies.

Critical weaknesses
  1. As shown in Table 3, the new filtered dataset brings considerable performance gains, especially for long events. For a fair comparison, the authors should report the results of the full method on both the original TimeLens-100K and the cleaned BiVTG-100K datasets under the same training settings.

  2. The paper finds that different rewards work better for short and long events. However, the final reward simply combines dIoU and boundary rewards with fixed weights, rather than introducing a more principled duration-aware design. This also introduces additional hyperparameters. Moreover, the combined reward does not consistently outperform individual rewards in Table 3. The novelty and motivation of this reward design are therefore not sufficiently strong.

Questions for authors

In Weakness.

1 comment
AuthorsOpenReview ↗

Response to Reviewer zALy

Thank you for your careful review and comments. Our responses are detailed below.


Q1: How much of the improvement comes from the training algorithm when the data are held fixed?

First, we have used BiVTG-100K for this comparison because hard-sample selection on the original TimeLens-100K can misclassify audio-dependent queries as difficult visual grounding examples before RLVR.

Table R.1 directly compares the baseline configuration with our method. Both use the same BiVTG-100K SFT checkpoint and the same training data, so the comparison isolates the joint contribution of sample strategy and reward design.

SFT dataSamplingRewardActivityNet ShortActivityNet LongCharades ShortCharades LongQVHighlights ShortQVHighlights Long
BiVTG-100KOne-Pass tIoUtIoU62.360.374.161.973.670.2
BiVTG-100KSome-Accept HybriddIoU + boundary66.660.978.459.678.967.0

Table R.1. Comparison using identical BiVTG-100K training data. Short-event columns report Dist ≤1×\leq 1\times percentage (↑), and long-event columns report [email protected] (↑).

With the training data held fixed, our method improves short-event grounding on all three benchmarks, consistent with its aim of mitigating the short-event degradation observed under tIoU-only RLVR. Long-event performance improves on ActivityNet and decreases on Charades and QVHighlights. Rows 1–3 of Table 3 show the separate contribution of data refinement, which improves long-event grounding at the SFT stage.

Together, the refined dataset, mixed reward, and sampling strategy improve over the prior state of the art on both short and long events across all three benchmarks.


Q2: What motivates the mixed reward, and where does the novelty lie?

The motivation for the mixed reward comes from the reward hacking behavior we observe under tIoU-based RLVR. The mixed reward itself is not complex for avoiding more hyper-parameters beyond weights. Our novelty claim centers on two underexplored observations and the following duration-aware framework they motivate.

1. Evaluation bias of tIoU for short events. As Figure 1 shows, a small absolute boundary shift can cause a large drop in tIoU for a short event, even when the predicted timestamp is close to the annotated one. We then correct the conventional evaluation by including boundary metrics that provide graded feedback on temporal error.

2. Interval inflation under tIoU-based RLVR. As Figure 2 shows, optimizing tIoU can encourage the model to predict excessively long intervals. Table 3 shows the associated trade-off. Training with tIoU alone improves long-event localization but can degrade short-event grounding. The mixed reward addresses this behavior by combining overlap-based supervision from dIoU with a boundary term that penalizes predictions that extend beyond the annotated interval, which is the motivation for our mixed reward.

The single-reward ablations in Table 3 show why the mixed reward does not consistently outperform each individual reward. Overlap-based supervision favors long events, while boundary supervision is particularly effective for pinpointing short events. The fixed mixture balances these objectives across the training set for this trade-off.


Thank you again for your comments and valuable questions.

Reviewer_3HzjOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper examines duration-dependent limitations of temporal IoU in video grounding, arguing that it penalizes short-event boundary errors excessively and encourages interval inflation during reinforcement learning. It proposes separate evaluation criteria for short and long events, a combined distance-IoU and boundary reward, multi-rollout sample selection, and the refined BiVTG-100K dataset. Experiments on three TimeLens-Bench datasets report improvements over TimeLens-8B and analyze the contributions of data refinement, sampling, and reward design.

Critical strengths
  1. Useful analysis of duration-dependent behavior. Separating short and long events exposes performance trade-offs that aggregate scores can obscure. The reward comparisons provide practical insight into how localization objectives affect different temporal scales.
  2. Systematic empirical evaluation. The paper evaluates three benchmarks and separately examines data refinement, sample selection, and reward shaping. Improvements on conventional metrics for ActivityNet and QVHighlights provide evidence of utility beyond the newly proposed evaluation criteria.
Critical weaknesses
  1. The proposed short-event metrics are insufficiently justified. The motivation emphasizes the excessive sensitivity of tIoU to small absolute errors, yet the proposed acceptance thresholds remain proportional to ground-truth duration. More importantly, these criteria can accept substantial under-coverage: for ground truth [10, 14] and prediction [11, 12], both Dist and Err satisfy their thresholds despite an IoU of only 0.25. This may be appropriate for approximate event spotting, but its suitability for temporal interval localization is not established. Without validation against annotation uncertainty, human judgments, or application requirements, improved scores do not adequately establish that the new criteria measure localization quality more faithfully.
  2. The reward-hacking interpretation is not sufficiently supported. Figure 2 demonstrates increasing predicted durations, but this observation alone does not establish that the model exploits the reward without improving localization. The analysis does not directly connect interval expansion, reward gains, and deterioration in independently assessed localization quality. Moreover, Table 3 shows that replacing tIoU with the final mixed reward improves short-event Dist accuracy while reducing long-event [email protected] on all three datasets. These results support an objective-dependent trade-off, but do not establish that the proposed method corrects the claimed reward-hacking mechanism. This distinction is central to the paper’s diagnosis and contribution.
Questions for authors
  1. Metric validity: What evidence supports the chosen duration-normalized thresholds as appropriate measures of short-event localization? In particular, how should predictions that satisfy both distance criteria but cover only a small fraction of the event be interpreted relative to annotation uncertainty or human judgments?
  2. Mechanism validation: Can the authors provide a controlled analysis linking interval expansion to increased reward and degraded localization quality on the same samples, and show that the mixed reward mitigates this behavior? Such evidence would help distinguish reward exploitation from a change in the localization trade-off.
3 comments
AuthorsOpenReview ↗

Response to Reviewer 3Hzj

Thank you for your positive comments and constructive feedback. Our responses are provided below.


Q1: The justification of the proposed short-event metrics.

In your example, the ground-truth interval [10,14][10,14] and the prediction [11,12][11,12] have centers only 0.50.5 s apart, while their tIoU is 0.250.25. For such a short event, this small center error indicates that the prediction pinpoints the occurrence closely. For an event whose full extent matters, tIoU captures how much of the annotated interval the prediction covers.

In Table 1, we classify events by their ground-truth duration. With a short-event cutoff of Tshort=5sT_{\text{short}} = 5 \mathrm{s}, the 44 s event in your example falls in the short-event group. We therefore evaluate it using Dist and Err, which measure the deviations of the predicted center and boundaries, respectively.

Additionally, Appendix Table 5 reports results with a cutoff of Tshort=3sT_{\text{short}} = 3 \mathrm{s}. Under this definition, the same event falls in the long-event group and is evaluated by tIoU. The score of 0.250.25 reflects that the prediction covers one quarter of the annotated interval.

Our method improves over the baseline under both cutoffs as shown in Table 1 and Table 5 (0-2s or 0-4s defined as short events). Thus, while the definition of a short event determines which metric applies to this particular example as the boundary case, the average improvement persists under both definitions on evaluated benchmarks.


Q2.1: How does interval inflation relate to reward and localization quality during training?

To supplement Figure 2, we track predictions at four checkpoints during RLVR with tIoU as the main reward. Table R.2 reports the training reward, the test performance on Charades, and the mean and median ratios of predicted duration to ground-truth duration (Pred/GT).

Training stepReward rtIoUr_{\text{tIoU}}Long-event [email protected] ↑Short-event Dist ≤1×\leq 1\times ↑Mean Pred/GTMedian Pred/GT
400.2456.378.21.691.20
800.3060.476.11.791.29
1200.3559.675.91.881.33
1600.3958.275.81.871.33

Table R.2. Training reward on hard samples from TimeLens-100K and test performance on Charades during tIoU-only RLVR.

Under tIoU-based training, reward gains are accompanied by longer predicted intervals and declining short-event localization accuracy. Long-event localization improves early in training, but those gains taper off even as the training reward continues to rise. This pattern supports the observation that optimizing tIoU alone encourages interval inflation and short-event localization degradation.


Q2.2: Does the mixed reward mitigate interval inflation?

Table R.3 extends the reward ablation in Table 3 with duration statistics for each reward setting after RLVR training.

RewardLong-event [email protected] ↑Short-event Dist ≤1×\leq 1\times ↑Mean Pred/GTMedian Pred/GT
tIoU60.676.71.761.25
dIoU60.976.91.741.25
Boundary58.878.31.601.11
dIoU + boundary (final)59.678.41.641.14

Table R.3. Localization performance and predicted-to-ground-truth duration ratios on Charades after 100 RLVR steps.

Adding the boundary reward to dIoU reduces interval inflation and improves short-event localization. The final mixed reward also yields shorter, more accurately sized predictions than tIoU alone, directly addressing the overextension observed during tIoU-only training. As you point out, Table 3 shows that this improvement indeed comes with a trade-off in long-event localization.

Thank you for the additional experiments. Tables R.2 and R.3 strengthen the evidence for interval inflation under tIoU-based training and its mitigation by the mixed reward, partially addressing my mechanism-validation concern.

However, my main concerns remain. The response does not establish why predictions with very low event coverage should be accepted as successful interval localization, and the cutoff analysis does not directly address this issue. I also still view the methodological contribution mainly as a task-specific combination of existing reward-shaping, sampling, and data-curation techniques.

AuthorsOpenReview ↗

Response to Reviewer 3Hzj

Thank you for the follow-up and for acknowledging that the additional experiments strengthen the evidence for interval inflation and its mitigation.

Regarding low coverage, our point is that, for short events, low tIoU does not necessarily imply unsuccessful localization, as illustrated in Figure 1. This is particularly relevant when human annotations have approximately one-second granularity rather than frame-level precision at 30 FPS. In such cases, overlap- or coverage-based evaluation can overstate localization errors relative to the precision of the annotations. We therefore consider absolute temporal deviations a more appropriate basis for assessing short-event localization. The short-event cutoff and acceptance thresholds can be varied to report results under different tolerance levels, and we will also consider reporting absolute boundary error alongside Dist and Err.

Regarding the methodological contribution, our central message is to expose the limitations of tIoU in both evaluation and RLVR training. We acknowledge that the proposed solution is straightforward, and we focus on whether it addresses the identified problems, rather than on the formal novelty or complexity of its components.

Thank you again for the discussion. It has helped us clarify both the rationale for our evaluation choices and the scope of our contribution.

Reviewer_Sq9VOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the limitations of Temporal Intersection over Union (tIoU) in the context of video temporal grounding, specifically regarding its sensitivity to event duration. The study highlights that for short-duration events, tIoU imposes an excessively heavy penalty for boundary errors; conversely, in reinforcement learning settings, reliance on tIoU can bias models toward producing overly long predictions. To address these issues, the authors propose a dual-mode framework that distinguishes between short and long events: it employs metrics based on distance and boundary errors for short events, while utilizing overlap-based metrics for long events. During training, the approach incorporates a distance-aware IoU and boundary-based rewards, alongside a multi-rollout sample selection strategy. Furthermore, the authors constructed a dataset, BiVTG100K, through data cleaning and augmentation. The effectiveness of the proposed method is validated on re-annotated versions of the ActivityNet Captions, Charades-STA, and QVHighlights benchmark datasets.

Critical strengths
  1. The paper points out that an identical absolute boundary error has a different impact on the tIoU (temporal Intersection over Union) for short versus long events; it further observes through experiments that training with a tIoU-based reward can lead to an expansion of the predicted intervals. This insight is valuable for understanding evaluation biases and training behaviors in the context of temporal action localization in videos.

  2. The authors conduct experiments across three benchmarks, analyze the roles of data cleaning, sampling strategies, and reward functions, and verify the effectiveness of their proposed method.

Critical weaknesses
  1. The innovation of the proposed method is limited, and its distinction from existing approaches remains unclear. The core framework primarily combines established techniques—namely dIoU/GIoU, boundary error penalties, and success-rate-based sample selection. The strategy of retaining a subset of successful rollouts is conceptually similar to the dynamic sampling approach used in DAPO. Although the paper cites DAPO, it fails to adequately articulate the unique contribution of the proposed sampling formula; furthermore, it lacks a direct comparison under an equivalent computational budget. It should be clarified whether the primary contribution lies in a novel optimization method or in a specific training recipe tailored for video localization.

  2. The proposed evaluation metrics may introduce new biases and lack validation regarding their effectiveness. For short-duration events, the metrics Dist≤|T| and Err≤|T| remain dependent on the length of the ground-truth interval. They do not correspond to a consistent absolute temporal tolerance and could potentially accept prediction intervals that are excessively contracted. For instance, if the ground-truth interval is [10, 11] and the prediction is [10, 10.01], both Dist and Err would fall within the thresholds, yet the tIoU would be only 0.01. While this does not necessarily imply that the metrics are "wrong," the authors must demonstrate that such predictions genuinely align with user requirements for short-event localization. It is recommended to incorporate human evaluation, as well as additional metrics such as absolute boundary error, interval coverage, and the ratio of predicted length to ground-truth length.

  3. Existing ablation studies suggest that the primary performance gains may stem from data cleaning rather than the proposed reinforcement learning (RL) method. In fact, the final method shows a slight performance decline on two of the datasets. Consequently, the current evidence better supports the claim that the method "mitigates the degradation caused by pure tIoU-based RL" rather than demonstrating "superiority over SFT in short-event localization." It is recommended to include Err metrics for the SFT baseline, report results across multiple random seeds, and evaluate the individual contribution of each algorithmic component while keeping the dataset and training budget constant.

  4. The explanation regarding "reward hacking" and the reproducibility of the method remain insufficiently addressed. While the lengthening of prediction intervals is an interesting phenomenon, an increase in the average interval length alone is not definitive proof that the model is exploiting reward loopholes through interval expansion; this could also be related to boundary uncertainty. It is necessary to demonstrate the relationship between prediction length, actual boundary error, and reward fluctuations, and to verify whether the new reward formulation effectively suppresses this behavior. Furthermore, the approach to reward normalization, the final mixing weights, and the settings for parameters such as α, γ, and τ require further clarification.

  5. The conclusion in Appendix A contains a clear mathematical error—for instance, on page 16, line 853.

  6. The formula for the "length of the union" in Appendix A does not hold when the intervals are disjoint.

  7. The textual explanation of the temperature parameter contradicts the sampling formula provided.

  8. The assertion that "there is no learning signal if all outcomes are successes or all are failures" does not apply to the continuous rewards used in this paper. The argument presented on page 6 (lines 295–298) is valid only if the actual rewards within a group are identical. Since this paper employs continuous rewards—such as tIoU, dIoU, and boundary rewards—the reward values ​​can differ even when all rollouts fall below the acceptance threshold. For example, if rewards are 0.1, 0.2, and 0.6, they are all below the 0.7 threshold; yet, there remains a relative advantage among them, providing a learning signal. Therefore, a binary acceptance state cannot be directly equated to a lack of informative signal in the GRPO rewards.

  9. There are inconsistencies between Table 9 and Table 4 that require verification. Based on the 4,500 samples for TimeLens in Table 9 and the counts across tIoU intervals, the calculated values ​​for [email protected] and [email protected] are approximately 68.7 and 58.2, respectively, whereas Table 4 reports 68.9 and 58.4. The authors should verify whether the same checkpoints, sample sets, and binning rules were used. This is a matter of reporting consistency and does not, by itself, imply that the experiments are flawed.

Questions for authors

Please provide responses to the questions listed under "Critical Weaknesses."

3 comments
AuthorsOpenReview ↗

Response to Reviewer Sq9V (1/3)

Thanks for the detailed comments. Our responses are provided below.


Q1: The novelty of the proposed method, and its distinction from DAPO.

Thank you for pointing out the ambiguity in how we presented our contribution. Our primary contribution is a training recipe tailored to video temporal grounding (VTG). It follows from two underexplored observations. First, the same absolute boundary shift can cause a much larger drop in tIoU for a short event with a low-variance value landscape (Figure 1). Second, optimizing tIoU directly during RLVR can encourage excessively long predictions, improving long-event localization while degrading short-event grounding (Figure 2 and Table 3). The proposed evaluation framework, combined reward, and sampling rules that follow from these observations address the above issues.

The specific contribution of our sampler is how it determines whether a rollout is acceptable in the VTG task. We share DAPO’s motivation for selecting informative queries. In math and coding tasks, answer verification or test cases provide a natural criterion for judging an outcome. In video temporal grounding, a prediction can be close to the annotated timestamp yet have little temporal overlap, so a single tIoU threshold can reject nearly correct short-event predictions. We therefore judge short-event rollouts by boundary error and long-event rollouts by tIoU, then use the number of acceptable rollouts to identify hard but solvable queries.

We will clarify that our novelty claim concerns the revealed limitations of tIoU and the reward hacking phenomenon, with the resulting duration-aware training recipe including a task-specific criterion of rollout quality.


Q2: The justification of the proposed short-event localization metrics.

In your example, the ground-truth interval is [10,11][10,11], and the prediction is [10,10.01][10,10.01]. The prediction starts at the annotated onset, and its center is within half a second of the annotated center, even though its tIoU is only 0.010.01. For a timestamp-based event, accepting this prediction under Dist and Err is our intended behavior: it gives credit for pinpointing the transient event. Its low tIoU captures a different property of the same prediction, namely that it covers only a small fraction of the annotated interval, which is essentially limited by the annotator's granularity (1s) of TimeLens-Benchmark.

We agree that it would be valuable to test how well this metric aligns with user judgments. As you mentioned, we will consider reporting absolute boundary error and the ratio of predicted to ground-truth duration alongside Dist and Err. We will also consider a human-preference study on a subset of short events, comparing the rank correlation between human judgments and metrics. This would directly test which metrics best reflect human preferences for short-event localization.


Q3: How much of the improvement comes from data refinement, and how much comes from the RLVR changes?

Applying the previous approach to BiVTG-100K corresponds to Row 5 of Table 3. In Table R.1, we compare that configuration with our method using the same BiVTG-100K SFT checkpoint and training data. This isolates the joint contribution of the sampling strategy and reward design.

SFT dataSamplingRewardActivityNet ShortActivityNet LongCharades ShortCharades LongQVHighlights ShortQVHighlights Long
BiVTG-100KOne-Pass tIoUtIoU62.360.374.161.973.670.2
BiVTG-100KSome-Accept HybriddIoU + boundary66.660.978.459.678.967.0

Table R.1. Comparison using identical BiVTG-100K training data. Short-event columns report Dist ≤1×\leq 1\times percentage (↑), and long-event columns report [email protected] (↑).

With the training data held fixed, our sampling and reward changes improve short-event grounding on all three benchmarks. Long-event performance improves on ActivityNet and decreases on Charades and QVHighlights. The clearest algorithmic contribution is therefore the recovery of short-event accuracy relative to tIoU-only RLVR. Rows 1–3 of Table 3 show the separate dataset contribution: data refinement improves long-event grounding at the SFT stage.

Together, the refined dataset, mixed reward, and sampling strategy improve over the prior state of the art on both short and long events across all three benchmarks.

AuthorsOpenReview ↗

Response to Reviewer Sq9V (2/3)

Q4: What supports the reward-hacking interpretation, and how were some hyperparameters chosen?

The details of the reward-hacking phenomenon and mitigation. Table R.2 tracks the tIoU reward, predicted-to-ground-truth duration ratios, and Charades test performance across RLVR checkpoints. As the training reward rises, predicted intervals become longer relative to the ground truth and short-event grounding precision declines. Long-event grounding precision improves early in training, then declines even as the reward continues to rise.

Training stepReward rtIoUr_{\text{tIoU}}Long-event [email protected] ↑Short-event Dist ≤1×\leq 1\times ↑Mean Pred/GTMedian Pred/GT
400.2456.378.21.691.20
800.3060.476.11.791.29
1200.3559.675.91.881.33
1600.3958.275.81.871.33

Table R.2. Training reward on hard samples from TimeLens-100K and test performance on Charades during tIoU-only RLVR.

Table R.3 tests whether the reward choice affects this behavior. Adding the boundary term to dIoU reduces both the mean and median duration ratios and improves short-event localization. These comparisons support our interpretation that tIoU-based training encourages interval inflation and that the mixed reward mitigates it.

RewardLong-event [email protected] ↑Short-event Dist ≤1×\leq 1\times ↑Mean Pred/GTMedian Pred/GT
tIoU60.676.71.761.25
dIoU60.976.91.741.25
Boundary58.878.31.601.11
dIoU + boundary (final)59.678.41.641.14

Table R.3. Localization performance and predicted-to-ground-truth duration ratios on Charades after 100 RLVR steps.

Hyperparameter selection. We chose α\alpha for the boundary reward and γ\gamma for the center reward by examining how smoothly each reward decreases as localization error increases. For example, Table 7 in Appendix D.3 illustrates one of the resulting reward values. For the sampling temperature τ\tau, we examined the difficulty distribution of the selected queries in Figure 4(a) and chose a value that retains hard but solvable examples.


Q5: The upper bound in Appendix A on page 16, line 853.

Let r=Err/∣T∣r=\mathrm{Err}/|T|. The bound on p. 16 gives tIoU≤1/(1+r)\mathrm{tIoU}\leq 1/(1+r). If r>(1−λ)/λr>(1-\lambda)/\lambda, then 1/(1+r)<λ1/(1+r)<\lambda. The prediction therefore falls below the acceptance threshold even when the intervals overlap. This is the implication used in the λ=0.7\lambda=0.7 example on line 859 and in the subsequent discussion of short events. We will revise line 853 to state the conclusion precisely as tIoU<λ\mathrm{tIoU}<\lambda.


Q6: Does the bound hold for disjoint intervals?

Yes. The expression on p. 16 gives the length of the smallest interval enclosing both spans. It equals the union length when the spans overlap or touch and includes the intervening gap when they are disjoint. We will use the general identity ∣T∪T^∣=∣T∣+∣T^∣−∣T∩T^∣|T\cup\hat{T}|=|T|+|\hat T|-|T\cap\hat{T}| and make the disjoint case explicit. For disjoint spans, tIoU=0\mathrm{tIoU}=0. Disjointness also implies Err≥∣T∣\mathrm{Err}\geq |T|, so the lower bound max⁡(0,1−Err/∣T∣)\max(0,1-\mathrm{Err}/|T|) is zero. The upper bound is nonnegative, and the boxed inequality therefore holds in this case as well.

AuthorsOpenReview ↗

Response to Reviewer Sq9V (3/3)

Q7: The textual explanation of the temperature parameter contradicts the sampling formula provided.

Thank you for pointing out the ambiguity. Eq. 3 uses an inverse temperature τ\tau and assigns a query with nn accepted rollouts the weight w(n)=exp⁡(τ/n)w(n)=\exp(\tau/n). The implementation uses a temperature TT and computes w(n)=exp⁡(1/(nT))w(n)=\exp(1/(nT)). These weights are identical when τ=1/T\tau=1/T. We will retain Eq. 3 and revise the statement in Appendix B.1.


Q8: The argument presented on page 6 (lines 295–298).

In practice, short events often yield little variation in tIoU across rollouts, leaving GRPO with a weak learning signal. For example, with a one-second target ([10,11]) and integer-second predictions, the possible positive tIoU values are 1,1/2,1/3,1/4,…1, 1/2, 1/3, 1/4, \ldots, and many distinct predictions receive the same 0. Therefore, for predictions of short events, their tIoU rewards are often identical and provide no relative learning signal because of this low variance. This practical motivation is what we intended to convey in lines 295–298, and we will revise the wording accordingly.


Q9: Why do the TimeLens-8B scores in Tables 4 and 9 differ?

Thank you for pointing out the ambiguity in the tables. Table 4 reports the TimeLens-8B results quoted from the original paper. Table 9 summarizes our own inference run of TimeLens-8B, for which we can obtain per-query predictions in addition to the original average scores. Because the tables report separate runs, their aggregate scores can differ.

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

Related papers

Discussion (0)

Sign in to comment.