acceptodds
Under review as a conference paper at ICLR 2027

Beyond Scalar IoU: Structured Verification from Rollout Groups for Video Temporal Grounding

Abstract

Reinforcement learning with verifiable rewards (RLVR) provides a natural frame- work for adapting pretrained models to video temporal grounding, where generated temporal intervals can be scored directly against ground truth intervals. Yet exist- ing overlap verifiers typically score each rollout independently, leaving the joint structure of the rollout group unused. We introduce SUTURE, which conditions verification on the rollout group and exploits its structure at two complementary scales: disagreement across rollouts controls how strongly the target is reweighted, while coverage at each position determines where reward mass is redistributed. We show that the resulting verifier admits an exact decomposition into the stan- dard IoU term and a covariance correction determined by the rollout group. A local gradient diagnostic finds a preference for responses covering relatively less supported target regions in the analyzed groups. Across five temporal grounding benchmarks, SUTURE improves grounding performance at every reported IoU threshold. Its trained policy also shows less video-start anchoring in reasoning traces: for later events, the first temporal mention more often overlaps the annotated target. Together, these results show that the joint structure of a rollout group can support a more informative temporal verifier

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video Temporal Grounding, Reinforcement Learning with Verifiable Rewards, Multimodal Large Language Model

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_5WZ7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces SUTURE, a rollout-group-conditioned verifier for video temporal grounding with reinforcement learning with verifiable rewards. The verifier redistributes weight within the annotated target using temporal group coverage, while pairwise disagreement controls the redistribution amplitude subject to a floor. Normalization preserves total target mass and leaves the IoU denominator unchanged. The paper derives an exact covariance-based reward decomposition and a fixed-group local GRPO direction identity. Experiments report improvements over TaRO across five benchmarks, with a second backbone evaluated on the three primary benchmarks. Fixed-amplitude controls, a separate three-run comparison, sensitivity analyses, and temporal-behavior diagnostics further characterize the method.

Critical strengths
  1. A simple and interpretable verifier modification. The separation between where to redistribute the target weight and how strongly to redistribute it is well motivated. The exact decomposition into ordinary IoU and a group-conditioned correction clarifies how this changes the learning signal. The matched-overlap ordering result and the local gradient analysis are appropriately qualified in the appendices: the latter permits a non-collinear correction but does not guarantee one for every group. This is more informative than an unexplained reward heuristic would be.

  2. Evidence beyond headline comparisons. The shared-setting Naive IoU and fixed-amplitude controls provide useful evidence for structured target weighting. Appendix F.1/Table C reports three-run means and sample standard deviations on the primary benchmarks, with the reported means favoring SUTURE. Group size and floor sensitivity are also examined in Appendix G.1/G.3. Together with the reported gains across benchmarks and the second-backbone results, these establish a credible positive case. The explicit limitations on selected qualitative examples, local gradient diagnostics, and token-count measurements are also valuable.

Critical weaknesses
  1. The specific benefit of disagreement-adaptive amplitude is not fully isolated. Table 3 and Appendix G.2 show that the adaptive rule outperforms the tested constants, but these ablations are explicitly single-run point estimates. A mean-matched constant does not match the adaptive amplitude distribution or its training-stage evolution. Appendix G.3 also reports that 82.8% of groups are floor-clamped at the analyzed late-training stage; this is not a whole-training statistic, but it makes the timing of the adaptive contribution particularly relevant. These results support the narrower empirical comparison while leaving uncertainty about whether the advantage specifically stems from assigning amplitude based on current group disagreement. This limits confidence in the claimed contribution of the adaptive component, rather than invalidating structured weighting itself.

  2. Some evaluation details are insufficiently specified for independent reproduction. Appendix E provides splits and counts for the three primary datasets, but the exact Ego4D-NLQ/TACoS evaluation splits, counts, complete second-backbone configuration, and checkpoint-selection criterion remain unclear. The long-video improvements in Table 2 are modest (0.4–1.0 percentage points), and lack reported across-run uncertainty. Separately, Figure 2 does not define its shaded bands, binning, target-start category cutoff, or treatment of outputs without valid temporal mentions. Appendix G.5 gives first-mention and hit conventions for a different comparison, without explicitly linking them to Figure 2. These omissions limit how confidently the broader performance and behavioral conclusions can be assessed.

Questions for authors
  1. Adaptive contribution. How stable is the fixed-versus-adaptive advantage under matched seeds and a common checkpoint-selection rule? Can you clarify when the amplitude is above its floor during training? A control preserving the stage-wise amplitude distribution while breaking its association with current group disagreement would help distinguish disagreement-specific allocation from scheduling effects; this is a suggested diagnostic, not a claim that the existing comparison is incorrect.

  2. Evaluation protocol. Please specify the long-video split versions/counts, the complete Qwen3 configuration, and the checkpoint-selection data and criterion. Which backbone and seed-pairing protocol does Table C represent, and is there uncertainty information for the long-video gains?

  3. Behavioral and implementation definitions. For Figure 2, please define the bins, central statistic, shaded bands, population counts, cutoff for classifying targets as “at video start,” and handling of missing/malformed mentions. Do the first-mention and positive-overlap hit conventions in G.5 also apply here? Separately, Appendix D.1 explains zero-length coverage; how do invalid parses enter the group, and how is pairwise IoU defined when the union has zero duration? These details could materially affect interpretation and reproducibility.

2 comments

We thank the reviewer for the detailed reading, including the appendices. Changes in the revised PDF are marked in blue, and equation and appendix references follow the revised PDF.

W1 / Q1. Contribution of the disagreement-adaptive amplitude

We ran both checks the reviewer proposed: matched seeds under a common checkpoint rule, and a control that keeps the amplitude distribution while breaking its link to each group's disagreement.

Matched seeds. We trained the mean-matched constant a=0.4a = 0.4 with the seeds of Table C (42, 1, and 2) and compare it with the three SUTURE runs already reported in Table C (Appendix F.1) of the submission. All runs use Qwen2.5-VL-7B-Instruct and the training settings of Table 3. Values are mean ± sample standard deviation over the three runs.

BenchmarkAmplitudemIoU[email protected][email protected][email protected]
QVHighlightsConstant a=0.4a = 0.459.8 ± 0.281.6 ± 0.667.2 ± 0.346.3 ± 0.3
QVHighlightsSUTURE61.4 ± 1.183.0 ± 1.069.3 ± 1.348.3 ± 2.0
ActivityNetConstant a=0.4a = 0.441.9 ± 1.061.5 ± 1.541.7 ± 1.222.3 ± 0.7
ActivityNetSUTURE44.3 ± 1.864.8 ± 2.344.9 ± 2.224.4 ± 1.5
Charades-STAConstant a=0.4a = 0.454.1 ± 1.180.1 ± 0.361.7 ± 2.434.0 ± 2.6
Charades-STASUTURE55.5 ± 0.480.8 ± 0.764.3 ± 0.937.1 ± 0.4

SUTURE has the higher mean on all twelve metrics, with mIoU gaps of 1.4 to 2.4 points.

Checkpoint rule. For every method and run, we select the checkpoint with the highest mIoU on TVGBench (Wang et al., 2025), following TaRO's checkpoint-selection protocol. TVGBench is not used for evaluation. The same checkpoint is used on every benchmark, and Appendix J.5 states this rule.

Within-batch distribution-preserving control. A control that permutes amplitudes across groups while preserving their within-batch distribution falls 3.3 mIoU below SUTURE. We compute the usual floored amplitudes for groups with uP>0u_{\mathcal{P}} > 0 and randomly permute these values among those groups before forming the weight fields. Each ϕP\phi_{\mathcal{P}} stays with its own group. The permutation exactly preserves the set of amplitudes within each batch while randomizing their assignment.

All three runs below use Qwen2.5-VL-7B-Instruct, the training settings of Table 3, and the checkpoint rule above. Values are single-run equal-weight averages over QVHighlights and Charades-STA, in percent. The fixed a=0.4a = 0.4 and SUTURE rows are already reported in Table 3 (Section 4.3) of the submission.

AmplitudemIoU[email protected][email protected][email protected]
Fixed a=0.4a = 0.4 (mean-matched)57.380.965.441.2
Shuffled within batch55.980.364.538.3
SUTURE (adaptive)59.282.967.644.1

The shuffled control falls below SUTURE on every metric. On ActivityNet it reaches 40.9 mIoU, against 44.3 ± 1.8 across the three SUTURE runs of Table C. Appendices J.1 and J.2 report these results.

The shuffled control also falls below the constant. One explanation consistent with the training statistics is that random reassignment can give a large amplitude to a group whose rollouts agree and cover most of the target. There ϕˉP\bar\phi_{\mathcal P} approaches −1-1, so the correction coefficient βP=aP/(1+aPϕˉP)\beta_{\mathcal P}=a_{\mathcal P}/(1+a_{\mathcal P}\bar\phi_{\mathcal P}) grows large. Over the first 234 steps, 23.5% of groups in the shuffled run have βP>0.9\beta_{\mathcal P}>0.9, against 6.9% under the adaptive rule.

When the amplitude is above the floor. With ufloor=0.3u_{\text{floor}}=0.3, the amplitude exceeds the floor for 88% and 60% of groups in epochs 1 and 2, and for 17% in epoch 3, the late stage analyzed in Appendix G.3. The adaptive rule therefore differs from a constant mainly in the first two epochs, and the shuffled control above, which keeps each step's amplitudes but reassigns them across groups, tests whether this assignment matters.

W2 / Q2. Evaluation protocol

These details were missing from the submission. Appendix J.5 of the revision states each of them in full.

  • Long-video splits. Ego4D-NLQ uses the v2 validation split (4,552 queries), since test annotations are not public, and TACoS uses the 25-video test split (4,001 queries).
  • Qwen3-VL-8B-Instruct. It follows TaRO's Qwen3 recipe and the GRPO settings of Appendix E.1. Table Q in Appendix J.5 lists the complete training and evaluation configuration.
  • Table C. All runs use Qwen2.5-VL-7B-Instruct with seeds 42, 1, and 2. The constant-amplitude runs in W1 / Q1 above use the same seeds.
  • Checkpoint selection. See the checkpoint rule in W1 / Q1 above.
  • Long-video uncertainty. We trained and evaluated TaRO and SUTURE with Qwen3-VL-8B-Instruct over three runs (seeds 42, 1, and 2) under the same checkpoint rule. The table reports mean ± standard deviation of R1 (%).
DatasetMethod[email protected][email protected][email protected]
Ego4D-NLQTaRO12.4 ± 0.77.2 ± 0.73.6 ± 0.3
Ego4D-NLQSUTURE13.5 ± 0.78.0 ± 0.64.4 ± 0.1
TACoSTaRO27.9 ± 0.617.3 ± 0.67.7 ± 0.4
TACoSSUTURE28.3 ± 0.218.3 ± 0.78.4 ± 0.8

SUTURE has a higher mean at all three thresholds on both datasets. The revised Table 2 reports these across-run statistics.

W2 / Q3. Behavioral and implementation definitions

The revised Figure 2 caption and Appendix J.5 give the complete definitions, including the bin edges and the number of queries in every bin and split. Figure 2 follows the conventions of Appendix G.5:

  • First temporal mention. The first "X s to Y s" span after <think>; Appendix J.5 gives the full parsing rule.
  • Panel (a). ActivityNet queries are grouped into seven bins of normalized target start s⋆/DVs^\star/D_V. Markers are medians and bands are interquartile ranges, over the same 14,005 queries for every method.
  • Panel (b). A target is at the video start when s⋆/DV≤0.02s^\star/D_V \le 0.02. A hit is any positive overlap with an annotated window, as in Appendix G.5.
  • Malformed responses. A query is kept only if all three methods produce a well-formed response with a valid first mention. Nearly all removed queries are Naive IoU responses without an <answer> tag, in line with its rate of reaching the 1,024-token limit (Table H).

Invalid parses and zero-length unions. An invalid response (no <answer> tag, no parsable interval, or an end that does not exceed the start) stays in the group as an empty prediction that covers no part of the target. Its pairwise IoU with every other member is 0. Valid intervals have positive length, so a union of zero duration occurs only between two invalid responses and follows the same rule. Sections 2.1 and 3.1 now state this convention.

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

This paper proposes SUTURE, a structured reward design for reinforcement learning-based video temporal grounding. The reward explores the structure of the groups of sampled rollouts by favoring less-supported correct overlap and redistributing strength according to the pairwise disagreement between the complete predicted intervals. The method is incorporated into an existing reinforcement learning framework without introducing additional inference-time modules. Experiments across multiple temporal grounding benchmarks and vision-language model backbones demonstrate consistent improvements over the evaluated baselines.

Critical strengths

(1) A well-motivated limitation of scalar IoU rewards: a single overlap score does not fully characterize the quality of predicted temporal boundaries, and an improving reward supervision is proposed by providing more structured feedback. (2) A simple method with no additional inference-time overhead. (3) Consistent empirical improvements.

Critical weaknesses

Insufficient discussion of closely related group-aware reward formulations Some prior works has already explored constructing rewards directly from group-level properties. In particular, GAPO (EMNLP 2025) introduces group-aware rewards based on output frequency and coverage, while DRA-GRPO (Findings of ACL 2026) adjusts rewards according to the semantic density of sampled reasoning paths. These works are conceptually relevant to SUTURE because they also exploit group-level structure to distinguish otherwise equally rewarded predictions and encourage underrepresented but valid outputs.SUTURE's temporal support weighting and IoU-compatible reward construction appear to be meaningful task-specific distinctions. [1] GAPO: Group-Aware Reinforcement Learning for Output Diversity in Large Language Models [2] DRA-GRPO: Your GRPO Needs to Know Diverse Reasoning Paths for Mathematical Reasoning

Questions for authors

(1) How does SUTURE differs from these existing group-aware reward formulations such as GAPO and DRA-GRPO? (2) Compare SUTURE against a simpler group novelty reward such as following: Rinovel=IoU(Pi,T)+λIoU(Pi,T)(1−1G−1∑j≠iIoU(Pi,Pj))R_i^{\mathrm{novel}} = IoU(P_i,T) + \lambda IoU(P_i,T) \left( 1-\frac{1}{G-1} \sum_{j\ne i}IoU(P_i,P_j) \right)

1 comment

We thank the reviewer for the positive assessment and for pointing us to GAPO and DRA-GRPO. Changes in the revised PDF are marked in blue, and equation and appendix references follow the revised PDF.

Q1. Relation to GAPO and DRA-GRPO

The revised Related Work (Appendix A) discusses both. GAPO (Anschel et al., 2025) uses the frequency of valid responses within the sampled group to promote output diversity. DRA-GRPO (Chen et al., 2026) uses semantic similarity among completions to promote exploration of distinct reasoning paths, including paths that reach the same correct answer. Both adjust rewards using response-level properties of the group.

SUTURE uses position-level coverage inside the annotated target interval. The group shows which correct temporal regions are covered less consistently, and the verifier redistributes target weight toward those regions. Three properties follow.

  1. The best answer stays the same. For a given target interval, a prediction that exactly matches it always receives the maximum overlap reward of 1, whatever the rest of the group predicts, and no other prediction reaches it (Remark 1, after Proposition 1). The group changes only how partial overlaps are scored. In GAPO and DRA-GRPO, which output receives the highest reward depends on the rest of the group.
  2. Only the correct part counts. Target weight is redistributed only inside the annotated interval, so differing from the group outside it earns no direct reward.
  3. Where the overlap falls matters. For two predictions in the same group with the same overlap and the same length outside the interval, Corollary 1 shows that the one whose correct overlap has lower average group coverage receives the higher reward. A similarity score between whole responses mixes in differences outside the target and does not indicate where inside the target the overlap lies. Above a floor, the adjustment grows stronger as the group disagrees more, and Proposition 1 gives the resulting correction relative to IoU exactly.

Q2. Comparison with a group novelty reward

We implemented the exact group novelty reward proposed by the reviewer and evaluated it with λ=1.0\lambda = 1.0:

Ri=IoUi+λ,IoUi(1−1G−1∑j≠ipIoU(Pi,Pj)),λ=1.0.R_i = \mathrm{IoU}_i + \lambda\\,\mathrm{IoU}_i \Big(1 - \frac{1}{G-1}\sum_{j \neq i} \mathrm{pIoU}(P_i, P_j)\Big), \qquad \lambda = 1.0.

Here pIoU\mathrm{pIoU} is the IoU between two predicted intervals and G=8G = 8 is the group size. Multiplying by IoUi\mathrm{IoU}_i restricts the bonus to rollouts that overlap the target, in line with the reviewer's emphasis on underrepresented but valid outputs. All runs use Qwen2.5-VL-7B-Instruct and the training settings of Table 3, with each checkpoint selected by the highest TVGBench mIoU, following TaRO's checkpoint-selection protocol.

The group mean of the bracketed pairwise-disagreement factor is exactly our disagreement statistic uPu_{\mathcal P}, with aP=max⁡{uP,ufloor}a_{\mathcal P}=\max\{u_{\mathcal P},u_{\mathrm{floor}}\}, where uflooru_{\mathrm{floor}} is the amplitude lower bound defined in Section 3.1. The methods therefore share a pairwise-disagreement signal but use it differently: the novelty reward directly scales each rollout's IoU, whereas SUTURE uses the group mean to control the strength of a shared, mass-preserving weight field within the target.

The group novelty row is new, and the SUTURE row is already reported in Table 1 of the submission.

[email protected] / 0.5 / 0.7QVHighlightsActivityNetCharades-STA
Group novelty82.5 / 68.1 / 47.663.4 / 43.9 / 24.879.2 / 62.2 / 35.1
SUTURE84.1 / 70.1 / 50.666.8 / 47.1 / 26.281.6 / 65.1 / 37.6

SUTURE outperforms the novelty baseline at all three thresholds on all three benchmarks (Appendix J.1).

Reviewer_bqX2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces SUTURE for Reinforcement Learning with Verifiable Rewards (RLVR) in video temporal grounding. Different independent scalar Intersection-over-Union (IoU) scores, SUTURE conditions verification on the joint support pattern of a rollout group by redistributing target weight to under-supported correct regions and scaling amplitude based on pairwise disagreement. Experiment results show that, SUTURE achieves consistent localization gains across five video temporal grounding benchmarks while mitigating video-start anchoring in MLLM reasoning traces.

Critical strengths
  1. It is good to see that, this paper provides an analytical derivation proving that SUTURE decomposes into standard IoU and a group-level covariance correction. It bridges individual metric evaluation with population-level group statistics.
  2. SUTURE extracts fine-grained spatial/temporal structure from the group's collective coverage pattern, which is different from standard RLVR methods that treat rollout groups merely as a means for scalar normalization.
  3. The method is validated across five benchmarks covering both short and long videos, demonstrating consistent improvements over standard scalar IoU baselines.
Critical weaknesses

This paper shows weakness in following aspects:

  1. The presentation should be significant improved. Many terms are used without clear definition, e.g., the s_star and e_star in the first equation is not defined. The first equation is used without an number. This makes some technical details hard to follow.
  2. Conditioning reward calculation on fine-grained rollout coverage patterns across multiple rollouts adds verification complexity during RL training. The paper lacks a convincing discussion of the computational overhead introduced by SUTURE.
  3. The mapping from pairwise disagreement to amplitude scaling introduces hyperparameters. A more convincing sensitivity analysis regarding how these scaling factors affect policy collapse or variance should be provided.
Questions for authors
  1. The design choice of hyperparameters should be studied.
  2. The authors are suggested to comapre SUTURE with a non-uniform weighting that does not use the rollout group, e.g, test triangular, Gaussian, or boundary-distance weighting. This experiment is important to support the claim that rollout-group structure is essential.
  3. A more convincing experiment and discussions on efficiency and computational overhead should be provided.
2 comments

We thank the reviewer for the concrete suggestions. Changes in the revised PDF are marked in blue, and equation and appendix references follow the revised PDF. We start with Q2, which the reviewer marked as important.

Q2. Non-uniform weighting without the rollout group

We trained controls that replace the group-dependent weight profile with a fixed target-position profile. Amplitude (a=0.4a = 0.4), mass normalization, and the IoU denominator are unchanged, so each rollout is scored independently of the group.

For an annotated target [s⋆,e⋆][s^\star,e^\star], let L=e⋆−s⋆L=e^\star-s^\star and ν=(t−s⋆)/L∈[0,1]\nu=(t-s^\star)/L\in[0,1]. All three controls use raw target weights w(t)=1+aψ(ν)w(t)=1+a\psi(\nu) with a=0.4a=0.4. The center-peaked triangle is ψcenter(ν)=1−2∣2ν−1∣\psi_{\mathrm{center}}(\nu)=1-2\lvert2\nu-1\rvert, and the boundary-emphasizing profile (Static boundary) is ψedge(ν)=2∣2ν−1∣−1\psi_{\mathrm{edge}}(\nu)=2\lvert2\nu-1\rvert-1. The triangle is an affine function of the distance to the nearest target boundary, so it is also the boundary-distance variant. The Gaussian profile is ψgauss(ν)=2exp⁡(−(ν−12)2/(2σ2))−1\psi_{\mathrm{gauss}}(\nu)=2\exp\big(-(\nu-\tfrac12)^2/(2\sigma^2)\big)-1 with σ=0.2\sigma=0.2 in normalized position.

All controls use Qwen2.5-VL-7B-Instruct and the training settings of Table 3, with each checkpoint selected by the highest TVGBench mIoU, following TaRO's checkpoint-selection protocol. The R1 values below are percentages averaged with equal weight over QVHighlights and Charades-STA. The Naive IoU, fixed a=0.4a = 0.4, and SUTURE rows are already reported in Table 3 (Section 4.3) of the submission; the three static profiles are new. Appendix J.1 gives the full control comparison.

Verifier[email protected][email protected][email protected]
Naive IoU78.661.937.2
Static center (triangle)79.163.239.5
Static Gaussian (σ=0.2\sigma = 0.2)79.663.840.6
Static boundary79.163.239.4
Fixed a=0.4a = 0.4 (group-conditioned)80.965.441.2
SUTURE (adaptive)82.967.644.1

At the same fixed amplitude a=0.4a=0.4, the group-conditioned field outperforms all three static profiles at all three IoU thresholds, and adaptive SUTURE improves these averages further.

A second control tests the rollout group from the other direction. It keeps SUTURE's adaptive amplitudes but reassigns them among groups with nonzero disagreement in each batch, and it falls 3.3 mIoU below SUTURE (Appendix J.2), so the group also matters for how strongly the weight moves.

W1. Presentation

We agree that several symbols were used before being defined. The revised manuscript numbers every displayed equation and defines symbols at first use: s⋆s^\star and e⋆e^\star (the annotated start and end times) in Section 2.1, mIoU in Section 4.1, and the baseline reward terms in Appendix A.

W2 / Q3. Computational overhead

SUTURE reuses the timestamps from the G=8G = 8 rollouts already generated per input, computing pairwise interval IoU and coverage on the CPU without extra video encoding or model forward passes. We measured full training runs with Qwen2.5-VL-7B-Instruct on four B200 GPUs, using 16 inputs per step (128 rollouts). Appendix J.4 reports these costs.

Total training timeSeconds per stepReward computation per step
Naive IoU13.81 h106.20.0 s
SUTURE13.91 h107.00.1 s
TaRO18.30 h140.730.4 s (temporal reward)

Reward computation is about 0.1% of a SUTURE step, and the verifier runs only during training, so inference cost is unchanged. TaRO's temporal reward needs two extra VLM forward passes for every rollout, since its IoU gate applies only when the reward is assigned, which adds 4.4 h to training (32% longer than SUTURE).

W3 / Q1. Hyperparameters and training stability

As defined in Section 3.1, uflooru_{\text{floor}} sets the minimum redistribution amplitude in aP=max⁡{uP,ufloor}a_{\mathcal{P}}=\max\{u_{\mathcal{P}},u_{\text{floor}}\}, where uPu_{\mathcal{P}} is the group disagreement. The mapping from disagreement to amplitude has no scale factor: uP=1−AvgPairIoU(P)u_{\mathcal{P}} = 1 - \mathrm{AvgPairIoU}(\mathcal{P}) is used directly, and uflooru_{\text{floor}} is its only hyperparameter. We use 0.3, the best value in the sweep below, and 0.5 performs comparably.

This floor sweep is already reported in Table F (Appendix G.3) of the submission, and the Naive IoU row in Table 3 (Section 4.3). All R1 values are percentages averaged with equal weight over QVHighlights and Charades-STA with Qwen2.5-VL-7B-Instruct.

Setting / uflooru_{\text{floor}}[email protected][email protected][email protected]
Naive IoU78.661.937.2
081.065.041.4
0.1581.064.840.7
0.3 (default)82.967.644.1
0.581.367.443.9
0.780.465.241.6

Sensitivity. Across all tested floor settings, the gains over Naive IoU are at least 1.8, 2.9, and 3.5 percentage points at [email protected], [email protected], and [email protected], respectively. The floor mainly changes the size of the reward correction: late in training, the mean target-weight variation ScorrS_{\text{corr}}, already reported in Appendix G.3 of the submission, rises from 0.066 at ufloor=0u_{\text{floor}}=0 to 0.115 at 0.3 and 0.366 at 0.7.

Policy collapse / variance. Beyond the submitted sweep, we now compute four statistics per group for every floor and average them over all groups in each of the three epochs: the rollout disagreement uPu_{\mathcal{P}}, the within-group standard deviation of the training reward σr\sigma_r, which normalizes the GRPO advantage, and the within-group IoU standard deviation and mean IoU. Naive IoU is measured the same way.

Setting / uflooru_{\text{floor}}uPu_{\mathcal{P}}: epoch 1 / 2 / 3Epoch-3 σr\sigma_rEpoch-3 IoU stdEpoch-3 mean IoU
Naive IoU0.587 / 0.381 / 0.2070.0860.0800.512
00.569 / 0.378 / 0.1910.0820.0770.512
0.150.576 / 0.381 / 0.1930.0830.0770.511
0.3 (default)0.547 / 0.387 / 0.1970.0870.0770.508
0.50.566 / 0.259 / 0.1050.0520.0430.525
0.70.579 / 0.360 / 0.1910.0970.0740.513

Under Naive IoU, rollout disagreement falls by about two thirds over training; we use this as the reference for how much diversity GRPO training normally loses. At the default floor, disagreement follows a similar trajectory, and the epoch-3 reward spread σr\sigma_r stays at the Naive IoU level, so the redistribution neither inflates reward variance nor shrinks the spread that normalizes the advantage. Floors of 0, 0.15, and 0.7 also stay near the Naive IoU level. The 0.5 run is the exception: from epoch 2 its rollouts agree more, but it reaches the highest epoch-3 mean IoU among all settings, and its final mIoU remains above Naive IoU. This narrowing does not recur at 0.7, so a larger floor does not by itself reduce diversity. At the default floor, we therefore see no sign of policy collapse or added reward variance relative to Naive IoU.

Appendix J.3 extends the existing Appendix G.3 analysis with these per-epoch diagnostics.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We thank all reviewers for their careful reading. The reviewers found the verifier simple and interpretable (5WZ7), valued the exact decomposition into IoU and a group-conditioned correction (bqX2, 5WZ7), and noted consistent gains across benchmarks and backbones (bqX2, VdtR, 5WZ7). The revised PDF marks every change in blue, and the response to each reviewer gives the details.

Controls for the rollout group and the adaptive amplitude (bqX2, VdtR, 5WZ7). We added three static target profiles that do not use the rollout group, the group novelty reward proposed by VdtR, and a control that shuffles SUTURE's amplitudes across groups. On the equal-weight average over QVHighlights and Charades-STA, SUTURE is above every new control and the fixed a=0.4a = 0.4 ablation of Table 3 on all four metrics (Appendix J.1). Over three matched seeds, it also has a higher mean than the fixed a=0.4a = 0.4 control on every metric of all three primary benchmarks (Appendix J.2).

Other additions.

  • Long videos (5WZ7). TaRO and SUTURE with Qwen3-VL-8B-Instruct over three runs: SUTURE has the higher mean at every threshold on Ego4D-NLQ and TACoS (Table 2).
  • Training stability (bqX2). Per-epoch rollout diversity and reward spread for every disagreement floor; at the default floor, both stay close to the Naive IoU level (Appendix J.3).
  • Training cost (bqX2). SUTURE trains in 13.91 h, against 13.81 h for Naive IoU and 18.30 h for TaRO on the same four GPUs; reward computation takes 0.1 s of a 107 s step (Appendix J.4).
  • Evaluation protocol (5WZ7). The checkpoint rule, long-video splits, the full Qwen3-VL configuration, and complete Figure 2 definitions (Appendix J.5).
  • Related work and presentation (VdtR, bqX2). The relation to GAPO and DRA-GRPO (Appendix A), numbered equations, and symbols defined at first use.
AuthorsOpenReview ↗

Gentle Reminder and Thanks

Dear Reviewers,

We have posted our responses and the additional results above, and we wanted to kindly check whether anything remains unclear. We would be happy to answer any further questions.

Thank you once again for your thoughtful reviews!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.