- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes DOVE, a training-free video token compression framework targeting sub-1% layer-averaged retention. The method constructs a quality-weighted evidence geometry from multiple visual and contextual cues, applies regularized D-optimal subset selection, and aggregates discarded information through a fixed-length residual readout. Evaluation on four benchmarks and three Video LLM backbones reports consistent gains over four selected training-free baselines in the targeted extreme-compression regime.
Critical strengths
Strong empirical performance: DOVE consistently outperforms the evaluated training-free baselines across three Video LLM backbones and four benchmarks under 0.5–1% token retention.
Critical weaknesses
-
Narrow and insufficiently motivated scope: The evaluation is restricted to 0.5–1% layer-averaged token retention, but the paper does not establish a realistic use case where such extreme compression, despite substantial information loss, would be preferable to a moderate operating point.
-
Incomplete validation and insufficient robustness analysis: The comparison omits relevant compression baselines such as DivPrune ( CVPR 2025), DyCoke (CVPR 2025), FLoC (ICLR 2026), STTM (ICCV 2025), LLaVA-PruMerge (ICCV 2025), LLaVA-Scissor (Arxiv 2025), and PruneVID (ACL 2025). It also excludes benchmarks specifically designed for extreme long-video understanding, such as MLVU(CVPR 2025), LVBench (https://lvbench.github.io/) and InfiniBench (EMNLP 2025). Moreover, DOVE introduces numerous handcrafted cues and hyperparameters without sufficient component-wise ablations or sensitivity analyses, leaving its robustness and generality unclear.
Questions for authors
Clarification requests
-
What practical deployment scenario motivates the 0.5–1% layer-averaged retention regime, and how was this range selected over more moderate compression levels?
-
How were the numerous cue weights and hyperparameters chosen, and were any evaluated benchmarks used during their development or tuning?
-
What advantages does DOVE provide over existing diversity-based token selection methods such as DivPrune and FLoC? Direct comparisons would be important to determine whether the proposed D-optimal objective offers meaningful benefits beyond prior diversity- and coverage-aware selection strategies.
-
The experiments use only 32 or 64 sampled frames. How does DOVE perform when the number of input frames is increased to 128, 256, or denser sampling such as 1 FPS before applying extreme token compression? This evaluation would help determine whether DOVE can effectively preserve sparse evidence from genuinely long videos, rather than only compressing a small set of uniformly sampled frames.
Suggestions for future work: None. These questions concern evidence needed to assess the paper’s current claims.
12 comments
Response to Reviewer TW46 [Part 2/5]
Q3 & W2: Comparisons with DivPrune, FLOC, PruneVID, Merging Baselines, and Long-Video LVBench
What advantages does DOVE provide over existing diversity-based token selection methods such as DivPrune and FLOC? Direct comparisons would be important to determine whether the proposed D-optimal objective offers meaningful benefits beyond prior diversity- and coverage-aware selection strategies. It also excludes benchmarks specifically designed for extreme long-video understanding, such as MLVU, LVBench, and InfiniBench.
Response:
- Direct Evaluations of DivPrune (CVPR 2025) & PruneVID (ACL 2025) (
Table 2): Under strict rebuttal compute constraints—having prioritized full five-benchmark runs of DivPrune (CVPR 2025) and PruneVID (ACL 2025), -frame scaling ( tokens), hour-long LVBench, noise stress-testing, and -setting sensitivity sweeps—we could not additionally run empirical evaluations for FLOC (ICLR 2026),DyCoke,STTM,LLaVA-PruMerge,LLaVA-Scissor, orMLVU/InfiniBenchbeyondLVBench, and will thoroughly discuss all of these in revised Section 2 (Related Work, noting that spatiotemporal/connected-component merging and dynamic KV pruning are represented empirically by FlashVID, VisionZip, PruneVID, and FastV). Across all four primary benchmarks (Table 2), DOVE beats DivPrune on every dataset at ( vs. Avg., points) and leads at ( vs. ), while outperforming PruneVID by points at ( vs. ) and points at ( vs. ). - Why Regularized D-Optimality Outperforms Pairwise Dispersion (DivPrune) & Exemplar Coverage (FLOC) (
Table 3): Pairwise max-min dispersion (DivPrune) and submodular facility-location / exemplar-coverage objectives (FLOC) rely on pairwise similarities: facility-location saturates once nearby tokens share a representative, while unweighted max-min dispersion treats noisy outliers as isolated clusters. Crucially, pairwise metrics cannot detect higher-order linear subspace collinearity—multiple tokens lying in an already-spanned 2D/3D subspace remain pairwise separated yet contribute zero new rank for linear VLLM value projections. DOVE's ridge-regularized explicitly maximizes quality-weighted joint subspace volume. Even under the exact same DOVE pool , geometry, and readout (Table 3), pairwise/metric-coverage selectors (Farthest-first,FPS) and unregularizedk-DPP MAP() consistently trail regularized D-optimality (). - Evaluation on Hour-Long LVBench (
Table 4): On LVBench (hour-long videos demanding multi-event retrieval), DOVE ranks first across all compressed methods at both (, over PruneVID and over DivPrune) and ().
Response to Reviewer TW46 [Part 3/5]
Table 2: Direct Comparison with DivPrune, PruneVID, and Prior Baselines on LLaVA-OneVision ( Frames)
| Method | VMME-S | VMME-M | VMME-L | VMME-All | Ego-Sub | Ego-Tot | LongVB | MVBench | Avg. | Rel. () | |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Vanilla | |||||||||||
| FastV$^\dagger$ | |||||||||||
| VisionZip | |||||||||||
| FastVID | |||||||||||
| FlashVID | |||||||||||
| PruneVID (ACL'25) | |||||||||||
| DivPrune (CVPR'25) | |||||||||||
| DOVE (Ours) | |||||||||||
| FastV$^\dagger$ | |||||||||||
| VisionZip | |||||||||||
| FastVID | |||||||||||
| FlashVID | |||||||||||
| PruneVID (ACL'25) | |||||||||||
| DivPrune (CVPR'25) | |||||||||||
| DOVE (Ours) |
FastV retains visual tokens in Layers -- ( / layer-averaged visual tokens).
Table 3: Controlled Comparison Against Diversity & Coverage Objectives Under Identical Pool & Geometry ()
| Selection Objective (Shared , Geometry & Readout) | VideoMME | EgoSchema | LongVideoBench | MVBench | Avg. Score | Rel. Acc. () |
|---|---|---|---|---|---|---|
| -DPP MAP (Unregularized determinant) | ||||||
| Farthest-first (Minimax metric coverage) | ||||||
| FPS (Iterative point coverage) | ||||||
| Regularized D-optimal (DOVE, Ours) |
Table 4: Evaluation on Hour-Long Video Benchmark LVBench (LLaVA-OneVision, Frames)
| Retention | Vanilla () | FastV$^\dagger$ | VisionZip | FastVID | FlashVID | PruneVID | DivPrune | DOVE (Ours) |
|---|---|---|---|---|---|---|---|---|
Response to Reviewer TW46 [Part 1/5]
We sincerely thank your valuable comment.
Q1, Q4 & W1: Practical Motivation for the 0.5%--1% Regime & Scaling to 128 Sampled Frames
Response:
Table 1): Following Q4, we scaled LLaVA-OneVision's input from its official 32-frame setting to 128 frames (N=25,088 raw visual tokens, where uncompressed quadratic LLM prefilling exceeds single-32GB-GPU serving envelopes) compressed by DOVE at R=1% with zero hyperparameter changes:VMME-Long(46.7→46.2, already retaining 94.5%--95.5% of 32f Vanilla's 48.9), LLaVA-OneVision's 32-frame instruction-tuning context prior and DOVE's untuned temporal-bin cap Kt=min(12,T) (Table 7, mapping both 32 and 128 frames into 12 coarse temporal intervals) keep long-split scores stable within single-split variance, whereas multi-event and motion-intensive benchmarks (EgoSchema,LongVideoBench,MVBench, andVMME-Short/Medium55.1→62.9 / 50.3→54.2) improve substantially.Table 1: Frame-Density Scaling from 32 to 128 Sampled Frames Under Extreme Compression on LLaVA-OneVision