acceptodds
Under review as a conference paper at ICLR 2027

SlashVID: Training-Free Extreme Token Compression via Select-and-Aggregate for Video Temporal Grounding

Abstract

Video Temporal Grounding (VTG) using Large Vision-Language Models (LVLMs) requires dense temporal sampling to preserve boundary-sensitive evidence, yet the resulting large number of visual tokens renders inference prohibitively expensive. At extreme compression ratios, pruning-only baselines can remove all tokens from some frames and discard subtle boundary cues. We propose **SlashVID**, a training-free framework that decomposes extreme token compression into two complementary paths: aggregating redundant tokens to preserve temporal context and selecting a small set of query-relevant tokens to sharpen boundary evidence. Specifically, SlashVID employs Adaptive Segment-wise Guidance to partition each video into segments and allocate a fixed budget across the two paths, Bidirectional Temporal Aggregation Clustering to consolidate the unselected context, and Dynamic Max-Min Importance Selection to add diverse, query-guided evidence. Experiments on Charades-STA and ActivityNet show that SlashVID retains 80% of full-token grounding performance at a 2% retention ratio while achieving 9.6 prefill speedup, improving relative performance retention over the strongest baseline by an absolute 4.55%.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video Temporal Grounding, Extreme Token Compression, Large Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_nbTQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

To address the issue where temporal action localization often loses evidence of action boundaries when constrained by an extremely limited budget of visual tokens, the authors propose "SlashVID." Its core contributions include: ASG, which reserves a budget for each temporal segment and allocates quotas between two branches based on visual complexity; DMM-IS, which selects detailed features by considering visual saliency, query relevance, and diversity, while ensuring that at least one token per frame is provided as input to the LLM; and Bi-TAC, which aggregates unselected features to enrich the context using bidirectional temporal merging and clustering with a temporal uniformity penalty. The authors evaluate the localization performance at various token retention rates on the Charades-STA and ActivityNet datasets using Qwen3-VL-4B/8B, and provide comprehensive experiments covering module ablation, boundary error analysis, and computational efficiency. Furthermore, the paper demonstrates the framework's generalizability across different tasks and backbone models through video question-answering experiments using the LLaVA series and temporal localization experiments using Qwen3.5.

Critical strengths

First, ASG allocates a budget to the two subsequent branches based on the input frame length and the effective rank of the visual features. Second, within this allocated budget, DMM-IS selects tokens by considering visual saliency, query relevance, and diversity, while ensuring that at least one token is selected for each input frame. Third, Bi-Tac is introduced to aggregate unselected tokens into a small number of contextual representatives, utilizing bidirectional temporal merging combined with density-peak clustering that incorporates a temporal uniformity penalty. These three modules take into account information from various aspects, and their integration is, in my view, a major highlight of this paper.

Critical weaknesses

I believe the organization of the experimental tables is problematic. Table 1 focuses on the Qwen3-VL series, which is clearly the primary model family for this paper. However, Table 2 introduces the LLaVA series and shifts the task to VideoQA, while reporting results only for a 10% token retention rate; this deviates from the core task and alters too many variables at once. Table 3 introduces another Qwen family model but covers only a single benchmark. Consequently, the current results are insufficient to demonstrate that the proposed method maintains its relative advantages when applied to VTG tasks using other model families.

Questions for authors

① It is worth considering running experiments with models other than Qwen2.5-VL, such as LLaVA-Onevision and reporting the results on the Charades-STA and ActivityNet datasets using token retention rates of 5% and 2%. Please compare these results against those of FlashVID, and also include the results obtained using the full-input setting. ② To evaluate cross-task generalization, please conduct additional evaluations on VideoMME and LongVideoBench using the primary backbone, Qwen3-VL, across various token retention rates. This will allow for a more direct assessment of how well the method transfers to VideoQA while keeping the backbone constant. ③ Please report the number of tokens allocated to each of the two branches by the SlashVID method at the 5% and 2% retention budget levels.

4 comments

We thank the reviewer for recognizing the complementary roles of ASG, DMM-IS, and Bi-TAC. We agree that changing the backbone and task simultaneously makes the generalization evidence harder to interpret. We will reorganize the experiments into two groups: cross-backbone VTG evaluation with the task fixed, and cross-task evaluation with the primary backbone fixed.

① Cross-family VTG evaluation with InternVL-3.5-8B

To examine generalization beyond the Qwen family while keeping the VTG task fixed, we report additional results using InternVL-3.5-8B on the Charades-STA test set (3,720 queries). The comparison includes full-input inference and FlashVID/SlashVID at multiple retention settings.

RetentionMethod[email protected] (%)[email protected] (%)[email protected] (%)mIoU (%)Relative mIoU (%)
100%Full44.0921.058.9030.84100.00
12.5%*FlashVID35.7315.485.9725.1781.61
12.5%SlashVID35.4615.916.2425.3682.23
5%FlashVID29.8711.454.0920.1565.34
5%SlashVID33.1513.845.1922.6873.54
2%FlashVID24.848.362.6615.6550.75
2%SlashVID28.7911.183.8218.9261.35

At the reported 5% and 2% settings, SlashVID improves mIoU over FlashVID by 2.53 and 3.27 percentage points, respectively, while retaining 73.54% and 61.35% of full-input mIoU. SlashVID also achieves higher recall at all three IoU thresholds in both settings. These results provide additional evidence that its relative advantage under low retention extends to the tested InternVL backbone.

This additional cross-family evaluation currently covers Charades-STA, with 2.5% as the lowest reported retention setting; the requested ActivityNet and exact 2% comparisons remain unaddressed by this table.

② Cross-task generalization with the primary Qwen3-VL-4B backbone

To isolate task transfer, we keep Qwen3-VL-4B fixed and evaluate VideoMME and LongVideoBench at 25%, 5%, and 2% retention, alongside full-input inference.

VideoMME contains 900 questions per duration category, totaling 2,700 questions; overall accuracy is computed from total correct predictions. LongVideoBench uses the same 1,337-question set across all configurations.

VideoMME

RetentionMethodShort Acc. (%)Medium Acc. (%)Long Acc. (%)Overall Acc. (%)Relative Acc. (%)
100%Full73.6761.1149.4461.41100.00
25%FastVID66.2254.2246.1155.5290.41
25%FlashVID70.1157.7847.8958.5995.42
25%SlashVID69.8957.5648.4458.6395.48
5%FastVID46.2237.5636.6740.1565.38
5%FlashVID58.3347.3342.4449.3780.40
5%SlashVID61.2249.7844.4451.8184.38
2%FastVID30.6727.3328.5628.8546.98
2%FlashVID49.1139.7836.5641.8168.09
2%SlashVID52.6743.1139.6745.1573.52

LongVideoBench

RetentionMethodAccuracy (%)Relative Acc. (%)
100%Full58.56100.00
25%FastVID52.6689.91
25%FlashVID55.7295.15
25%SlashVID55.9595.53
5%FastVID37.7064.37
5%FlashVID48.0281.99
5%SlashVID50.7186.59
2%FastVID27.7547.38
2%FlashVID40.0968.45
2%SlashVID43.7574.71

SlashVID achieves the highest overall accuracy among the compared compression methods at every tested retention rate on both benchmarks. At 5%/2% retention, its improvements over FlashVID are 2.44/3.33 percentage points on VideoMME and 2.69/3.66 points on LongVideoBench. At 25% retention, the differences are small: SlashVID answers one additional VideoMME question and three additional LongVideoBench questions correctly. The clearest improvements occur at low retention, supporting cross-task applicability with the primary backbone fixed.

③ Token allocation between selection and aggregation

We clarify the allocation rule and provide a calculation under the stated resolution-sensitivity prior. For a video with MM original tokens per frame and target retention rr, the per-frame budget is

B=max⁡(1,⌊rM⌋).B=\max(1,\lfloor rM\rfloor).

For segment kk, containing TkT_k frames,

=T_k\max(1,\lfloor\alpha_k B\rfloor), \qquad N_{\mathrm{agg},k} =T_kB-N_{\mathrm{sel},k}.$$ The prior contains 100 Charades-STA video units with 58 frames and $M=256$, and 500 ActivityNet units with 118 frames, where the counts for $M=64/100/152$ are $75/420/5$. Each unit contains two segments. The frame-weighted $\alpha$ is $0.52\pm0.12$. | Dataset | Target retention | Mean total budget | Realized retention (%) | Selection tokens | Selection share (%) | Aggregation tokens | Aggregation share (%) | | --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | | Charades-STA | 5% | 696.00 | 4.6875 | 359.60 | 51.67 | 336.40 | 48.33 | | Charades-STA | 2% | 290.00 | 1.9531 | 139.20 | 48.00 | 150.80 | 52.00 | | ActivityNet | 5% | 556.96 | 4.9622 | 265.97 | 47.75 | 290.99 | 52.25 | | ActivityNet | 2% | 219.48 | 1.9554 | 118.47 | 53.98 | 101.01 | 46.02 | Token counts are averages per video unit. Realized retention is calculated as total allocated tokens divided by total original visual tokens. The final selection share differs from the mean $\alpha$ because allocation applies integer rounding and a minimum of one selected token per frame. For example, when $M=152$ at 2% retention, $B=3$, and $\alpha=0.42/0.67$ yields one/two selected tokens per frame. These figures describe **allocation under the specified prior**, independently of the duration-bucket latency analysis. The minimum selection quota ensures at least one selected token per frame at the LLM input by construction; empirical branch-output statistics require verification against execution logs.

I am maintaining my original score and am even considering lowering it.

The authors have not yet incorporated the requested experimental results into the manuscript, and I still find the results presented in the paper confusing. At a minimum, the additional results should be included in the appendix.

I have already specified the experiments needed, including evaluating the method’s generalizability on video tasks with Qwen3-VL-8B. However, the authors have still not fully addressed these requests. I therefore cannot tell whether any unfavorable results were obtained but left unreported.

Dear Reviewer, Thank you for your candid follow-up. We sincerely apologize for not uploading the revised PDF alongside our initial response, as well as for the misunderstanding regarding the model scale. First, as your initial question referred to generalization experiments on Qwen3-VL without specifying a particular model scale, we evaluated the cross-task video benchmarks on Qwen3-VL-4B because it serves as the primary backbone across our main tables. We prioritized maintaining consistency with that backbone and therefore did not immediately run the evaluation on Qwen3-VL-8B. We take your concerns very seriously and are addressing both points right away:

  1. Revised Manuscript: We are finalizing the revised PDF to present the experimental results more clearly and incorporate all additional rebuttal results into the main text and appendix (with all changes highlighted). We will upload it shortly.

  2. Qwen3-VL-8B Video Evaluation: We have launched the requested generalization experiments on Qwen3-VL-8B. Once the runs are complete, we will post the full results here and include them in the revised appendix.

Thank you for your patience and for holding our work to a high standard.

Reviewer_jbhYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper deals with token compression in video temporal grounding with VLM to improve efficiency. As VTG is to find the start and end of queried events, it uses boundary aware methods. First segment the video based on dissimilarity, allocate budget in segment-wise. The budget is composed of query-dependent parts and aggregation of the remaining parts. The query guided parts also use diversity as one of the criteria.

Critical strengths

It is good that this paper chose a token pruning methods that are related with the given task. The intuitions are clear, algorithms are chosen based on ablation study. The result shows meaningful improvements from the other methods.

Critical weaknesses

The idea itself is reasonable. But the system is composed of many modules and algorithms. Even though they are compared with alternatives and the performance is better than others, it looks like those selections are based trial and errors on test data, rather than important new ideas or founding. So, this is a well working system for the test benchmarks, I am not sure that it will work well for general cases.

Questions for authors

Do we need a accurate video temporal grounding with single stage? Why not using two stage approach where the first stage find the rough location with minimum tokens and the second stage refine the timestamps with more tokens on the candidate region?

1 comment

We thank the reviewer for raising concerns about generalization and the choice of a single-stage evaluation protocol.

On the inference of test-set-driven trial and error. We respectfully disagree with this inference. Experimental feedback is necessary to assess whether the proposed modules deliver their intended effects. Such evaluation does not, by itself, constitute trial-and-error optimization on test data. In our work, hyperparameters are not repeatedly retuned to improve test-set scores. The number of components alone therefore does not substantiate the suggestion of test-set-driven tuning.

We recognize that generalization and robustness require empirical evidence, and we address these concerns through controlled ablations, joint sensitivity analysis, and transfer evaluations. Specifically, we jointly vary the merging threshold τm\tau_m and segmentation threshold τs\tau_s on Charades-STA with Qwen3-VL-4B at 2% retention:

τm\τs\tau_m \backslash \tau_s0.8500.8750.9000.9250.950
0.7043.3043.4143.7443.3843.13
0.7544.2145.3445.6945.3444.40
0.8044.3345.4745.8345.9044.58
0.8544.2045.3645.7345.7745.51
0.9043.5644.7344.1244.8243.13

Across all 25 configurations, mIoU ranges from 43.13 to 45.90. Within the neighborhood τm∈[0.75,0.85]\tau_m\in[0.75,0.85], $tau_s\in[0.85,0.95]$, all 15 configurations achieve 44.20–45.90 mIoU. The originally reported default setting, (τm,τs)=(0.80,0.90)(\tau_m,\tau_s)=(0.80,0.90), achieves 45.83 mIoU. We retain this setting rather than replacing it with the highest-scoring configuration in the sensitivity sweep. These results support local robustness to simultaneous threshold variation and complement the one-dimensional sensitivity analyses.

We will explicitly document how the defaults were determined and which configurations were held fixed across evaluations. Without specific evidence of test-set-driven parameter selection, characterizing our development process as trial-and-error optimization goes beyond what the available evidence supports. We therefore respectfully ask that concerns about generalization be distinguished from claims about test-set tuning.

On two-stage coarse-to-fine grounding. We agree that coarse localization followed by local refinement is a reasonable approach. However, the question addressed by this paper is how to preserve grounding evidence under an extremely constrained visual-token budget. All methods in our main VTG comparison are evaluated under the same single-stage grounding protocol, with the same backbone, frame sampling, decoding settings and so on. This shared protocol provides a controlled comparison of compression strategies. The reported improvements therefore establish the effectiveness of SlashVID within this setting.

We recognize two-stage coarse-to-fine grounding as a reasonable alternative, and SlashVID could be incorporated into either stage. However, its potential benefits do not invalidate a shared single-stage protocol, which isolates the effect of compression on grounding quality. Under a matched cumulative visual-token budget, coarse localization and refinement compete for the same limited resources. Reducing the first-stage budget may produce inaccurate candidate windows, leaving the refinement stage unable to recover evidence outside them. Expanding the windows or retaining multiple candidates increases coverage but spreads the remaining budget over a larger region, reducing local token density. A second LLM invocation also introduces additional overhead, while newly sampled frames require additional visual encoding. We therefore view two-stage grounding as a complementary system-level extension whose potential benefits do not, by themselves, undermine the validity of our single-stage evaluation protocol.

We therefore respectfully disagree that demonstrating the superiority of single-stage over two-stage grounding is necessary to substantiate our contribution. The relevant question for the reported experiments is whether SlashVID improves grounding quality and efficiency under the shared evaluation protocol.

Reviewer_AJ9tOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SlashVID, a training-free framework for extreme video token compression in video temporal grounding. SlashVID combines Adaptive Segment-wise Guidance for segment-level budget allocation, Dynamic Max-Min Importance Selection for query-aware token selection, and Bidirectional Temporal Aggregation Clustering for preserving contextual information from unselected tokens. Experimental results show that SlashVID outperforms the existing methods on the various benchmark tests.

Critical strengths
  • This paper is clearly written and easy to follow.
  • The proposed algorithm is well-motivated for extreme compression scenario, and technically sound.
  • Extensive experiments and fine-grained ablations across multiple models, benchmarks, and retention ratios demonstrate a competitive accuracy–efficiency trade-off.
Critical weaknesses
  • More experimental evidence on long-horizon temporal grounding and the robustness of the many heuristic design choices would strengthen the paper, as these aspects are particularly relevant to extreme-compression settings.
Questions for authors
  1. A qualitative visualization on real examples, showing the temporal segmentation, adaptive branch allocation, selected tokens, and aggregated representatives, would be helpful. Comparing the resulting boundary predictions with those of pruning-only or aggregation-only baselines would further clarify how each component contributes.

  2. The VTG evaluation is limited to Charades-STA and ActivityNet-Captions. Results on long-horizon benchmarks, such as Ego4D-NLQ, would better support the motivation. Reporting accuracy, temporal coverage, and runtime as video length increases would also be useful.

  3. The proposed method contains several heuristic components and a relatively large number of hyperparameters. While the paper provides one-dimensional sensitivity analyses for some parameters, these experiments do not examine interactions among them. In particular, Table 19 shows a 3.65-point mIoU variation across values of the merging threshold, which is comparable to some of the reported improvements over the strongest baseline. Could the authors provide further evidence that the gains are robust to joint parameter variation and that the same default configuration transfers across datasets and backbones without dataset-specific tuning?

  4. minor: L1253 VTG datasets TACoS (?) and Ego4D-NLQ (?) : wrong citation

3 comments
AuthorsOpenReview ↗

Response 1: Qualitative visualization on real example

Thank you for this helpful suggestion. We agree that connecting the compression process to the resulting temporal predictions can clarify the role of each component. Using real inference traces, we analyze video D8FTI from Charades-STA through four stages: temporal segmentation, adaptive branch allocation, selected tokens, and aggregated representatives. We also compare boundary predictions under the same visual-token budget.

The video is 31.05 seconds long, and the query is “person open a cabinet,” with a ground-truth interval of [1.40, 7.00] s. The cabinet-opening action occurs early in the video, while the person and cabinet remain visible afterward. The model therefore needs to distinguish the target action from the continued presence of the associated person and object.

For temporal segmentation, the video is divided into two compression segments containing 13 and 18 effective visual time units, respectively. The split at approximately 12.96 s is introduced by the minimum-segment rule. These segments organize the compression process; their boundaries do not represent the target action’s start and end times.

For adaptive branch allocation, each segment receives a total budget proportional to its length, while the adaptive mechanism adjusts the allocation between selection and aggregation within that segment. The realized allocations are:

SegmentαSelected tokensAggregated representativesRealized selection / aggregation ratio
S10.30267825% / 75%
S20.801083675% / 25%

The first segment primarily uses aggregation, whereas the second primarily uses direct selection. We report both α and the realized allocations to distinguish the allocation parameter from the actual branch proportions after budget discretization.

For selected tokens, DMM-IS retains 134 original visual tokens. Their patch locations show which local visual evidence is preserved directly. For aggregated representatives, the model produces 114 representatives, whose source support reveals how information is combined across time. For example, one representative centered at approximately 4.32 s has source support spanning all 13 time units in the first segment and involving 1,437 original tokens. This illustrates that a representative can incorporate information beyond its center frame, without implying equal source weights or lossless information preservation.

To connect these representations to grounding performance, we compare selection-only (pruning-only), aggregation-only, and complete SlashVID under identical sampled frames, queries, decoding settings, and in-layer pruning settings. All three methods receive 248 input visual tokens, with each single-branch baseline assigning its entire budget to its respective branch.

MethodPredicted interval (s)IoUAbsolute start error (s)Absolute end error (s)
Selection-only (pruning-only)[14.70, 20.00]0.00013.3013.00
Aggregation-only[10.00, 17.00]0.0008.6010.00
SlashVID[1.30, 10.40]0.6150.103.40

Under the same budget, both single-branch baselines predict intervals after the annotated action. SlashVID shifts the prediction back to the correct early portion of the video, reducing the start-boundary error to 0.10 s. Its end boundary remains 3.40 s late, so this example demonstrates a substantial improvement in localization, although the boundaries are not fully recovered. The measured visual-token retention rate for this example is 248/9,300, or approximately 2.67%.

A supplementary example, 1TZK5, further illustrates this complementary behavior. For the query “one person stands by the window eating food,” the ground-truth interval is [5.70, 11.40] s. Selection-only, aggregation-only, and SlashVID achieve IoUs of 0.000, 0.377, and 0.475, respectively. Compared with aggregation-only, SlashVID reduces the start/end errors from 5.70/1.40 s to 4.40/0.90 s. Furthermore, the third and fourth segments each contain five time units and receive 40 output tokens, but their selection/aggregation allocations are 10/30 and 30/10, respectively. This shows that the branch allocation is not determined solely by segment length.

We will add visualizations of more real examples to the revised PDF. The figures will align sampled frames, temporal segments, adaptive branch allocations, selected patch locations, and aggregated representatives with their source support along a shared timeline. They will also show the ground-truth intervals alongside predictions from all methods under the same budget, making the compression process and its relationship to temporal localization easier to inspect.

AuthorsOpenReview ↗

Response 2: Results on long-horizon benchmarks

Thank you for highlighting the importance of evaluation on long-horizon benchmarks. We considered this setting and conducted preliminary experiments on Ego4D-NLQ. However, these experiments revealed a limitation of the underlying model: under the evaluated setup, the LVLM does not yet demonstrate reliable long-horizon temporal grounding, even without token compression. Compression results on this benchmark therefore need to be interpreted in the context of the Full baseline’s observed performance.

Specifically, we evaluated Qwen3-VL-4B on 4,552 queries using the no-think setting. All Recall metrics and mIoU (all) are computed over the entire query set, with an IoU of zero assigned to samples whose predictions cannot be parsed.

MethodVisual-token retentionValid parse (%)[email protected] (%)[email protected] (%)[email protected] (%)mIoU (all, %)
Full100%97.6110.115.452.507.03
SlashVID5%97.438.944.812.226.22

Although Full achieves a valid parse rate of 97.61%, its mIoU is only 7.03%, and its [email protected] is only 5.45%. Thus, successfully producing a parseable answer does not translate into accurate temporal localization: substantial grounding errors are already present before compression. At 5% visual-token retention, SlashVID achieves an mIoU of 6.22%, preserving approximately 88.5% of Full’s mIoU. However, this relative retention is measured against a baseline with low absolute accuracy.

With both methods achieving less than 10% mIoU in this experiment, differences within this low-accuracy regime provide limited evidence of effective long-video grounding. These differences remain useful for quantifying relative performance retention, but they do not establish that the compressed model can reliably perform the task. Our concern is the low absolute accuracy of the uncompressed baseline, rather than any general validity threshold at 10% mIoU. This distinction is particularly relevant under extreme compression, such as 5% visual-token retention, where a high relative retention rate alone is insufficient to demonstrate useful long-horizon localization. We therefore view these Ego4D-NLQ results as preliminary evidence, rather than a basis for strong claims about effective or scalable long-video grounding.

AuthorsOpenReview ↗

Response 3 for Q3 and Q4

Thank you for raising this concern. We agree that one-dimensional sensitivity analyses do not capture interactions between hyperparameters. The 3.65-point mIoU variation in Table 19 also warrants further investigation, particularly given its magnitude relative to some of the reported gains.

To examine these interactions, we jointly vary the merging threshold (τm\tau_m) and segmentation threshold (τs\tau_s) on Charades-STA using Qwen3-VL-4B at 2% visual-token retention. The resulting mIoU scores (%) are:

τm\τs\tau_m \backslash \tau_s0.8500.8750.9000.9250.950
0.7043.3043.4143.7443.3843.13
0.7544.2145.3445.6945.3444.40
0.8044.3345.4745.8345.9044.58
0.8544.2045.3645.7345.7745.51
0.9043.5644.7344.1244.8243.13

Across all 25 configurations, mIoU ranges from 43.13 to 45.90, a spread of 2.77 points. Within the neighborhood defined by τm∈{0.75,0.80,0.85}\tau_m \in \{0.75, 0.80, 0.85\} and all five tested τs\tau_s values, the 15 configurations achieve 44.20–45.90 mIoU, with a smaller spread of 1.70 points.

The originally reported default configuration, (τm,τs)=(0.80,0.90)(\tau_m,\tau_s)=(0.80,0.90), achieves 45.83 mIoU, only 0.07 points below the highest score in this sweep. We retain this original default without replacing it with the highest-scoring configuration. These results support local robustness to simultaneous variation of the two thresholds and complement the existing one-dimensional analyses. They also indicate that threshold choices still matter; we do not interpret this experiment as establishing insensitivity to all hyperparameters or their interactions.

Regarding transfer, all datasets and backbones evaluated in the paper use the same default hyperparameter configuration, without dataset-specific or backbone-specific tuning. Thus, the reported results across datasets and backbones were obtained using a shared configuration, rather than separately optimized settings. We will explicitly clarify this protocol and include the joint sensitivity results in the revised manuscript. Together, these experiments provide evidence of local threshold robustness and the transferability of the fixed default configuration within the evaluated settings.

Thank you for your careful reading and for pointing out these citation errors. We will correct the references for TACoS and Ego4D-NLQ at line 1253 in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.