- 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
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 original tokens per frame and target retention , the per-frame budget is
For segment , containing 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:
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.
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.
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.
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
LongVideoBench
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.