acceptodds
Withdrawn from ICLR 2027

GLANCE: Efficient Video Large Language Models via Screening and Parallel Redundancy Suppression

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Denser frame sampling improves temporal coverage in video large language models (Video-LLMs) but makes decoder prefill expensive. Visual-token compression must balance subset quality with selection cost: token-wise ranking can retain redundant evidence, whereas subset-aware selection incurs pairwise comparisons and sequential updates. We introduce , a training-free, query-agnostic video token compressor that compares redundancy within compact candidate pools. Informativeness and Uniqueness-guided Screening (IUS) combines within-group feature-response ranks with video-wide directional uniqueness to shortlist candidates. Parallel Redundancy Suppression (PRS) discounts similarity to earlier candidates in a fixed utility order, computing penalties in parallel without iterative selected-set updates. Across three Video-LLM backbones and four benchmarks, the complete configuration, with approximately 12.5% decoder-input retention and late visual-token withdrawal at nominal 10% retention, preserves 94.2–97.6% of full-token mean accuracy with – speedups in model-side time to first token (TTFT). In a matched Qwen3-VL comparison, PRS reduces candidate-selection time from 8.37 to 0.66 ms at 32 frames with close observed accuracy. Compressed 128-frame input also exceeds the 64-frame full-token baseline in three-task mean accuracy at lower TTFT on Qwen3-VL.Code is available at https://anonymous.4open.science/r/glance-CEE3

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
MLLM Inference acceleration, Training-free, Plug and play

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

GLANCE is a training-free, query-agnostic visual-token compressor for video language models. IUS combines within-group feature-norm ranks with video-wide directional uniqueness to construct candidate pools. PRS then computes similarity penalties against preceding candidates in a fixed utility order in parallel, avoiding the sequential updates of greedy subset construction. The complete system additionally expands the input budget and withdraws visual tokens in later decoder layers: nominal 10% retention corresponds to approximately 12.5% decoder-input retention and 9.4% layer-averaged visual-token retention. Across three backbones and four benchmarks, this configuration preserves 94.2%–97.6% of full-token mean accuracy with approximately 2.2–3.9× speedups in model-side time to first token. Matched selector comparisons, component ablations, and frame-scaling experiments further evaluate the design.

Critical strengths

The low-cost redundancy treatment has direct empirical support. Fixing the candidate order makes PRS amenable to batched execution and well suited to GPU inference. With the same IUS scores, candidate pools, quotas, and depth schedule, Table 3(c) reduces candidate-selection time from 8.37 ms for Greedy to 0.66 ms, with three-task mean accuracy of 60.2 versus 60.0. The paired analysis in Appendix D additionally supports a TTFT reduction on that cohort. The authors appropriately avoid claiming that PRS is more accurate than Greedy. Separating accuracy, module time, and overall latency makes the practical argument convincing.

Broad evaluation demonstrates a useful trade-off for temporal coverage. Results across three backbones, four benchmarks, and multiple compression levels provide stronger evidence than a single-model, single-budget evaluation. In Table 10, 128-frame GLANCE on Qwen3-VL achieves a three-task mean of 63.0 at 981.3 ms TTFT, compared with 62.4 at 1,383.0 ms for the 64-frame full-token model. This illustrates how compression savings can support denser temporal observation. Reporting the accompanying peak-memory increase from 23.6 to 27.4 GiB also makes the trade-off transparent.

Critical weaknesses

Cross-method comparisons do not fully isolate the new selector's contribution. Tables 2 and 14 usefully separate IUS–PRS from input expansion and late withdrawal, but competing methods have not been evaluated under the same schedule. For example, at the 20% setting on Qwen3-VL, the selector alone scores 62.3, matching V-CAST's displayed mean in Table 1, whereas the complete configuration reaches 63.0. The main results therefore support the accuracy–latency trade-off of the complete system, but do not attribute all of that advantage to IUS–PRS. Comparing the strongest baseline with matched input quotas and withdrawal depth would better identify the methodological contribution. This is an addressable evidence gap and does not invalidate the measured benefit of the complete configuration.

The robustness of the smaller main-table accuracy gains is not directly quantified. Appendix D does provide appropriate paired bootstrap analyses, but its accuracy diagnostics concern separate subsets and component comparisons; they are not confidence intervals for the main four-benchmark means or differences from the strongest external baselines. This matters for the approximately 0.4-point advantage on LLaVA-Video at 10% retention and for the task-level regressions observed when adding uniqueness. The current results establish competitive performance, but paired analyses of the full prediction sets would clarify whether the smaller advantages are robust.

Questions for authors

Could at least one strongest baseline be evaluated with the same input quotas and late-withdrawal schedule, reporting accuracy and TTFT against IUS–PRS? This would directly affect my assessment of the selector's incremental contribution.

Could the existing full prediction sets be used to report video-clustered paired intervals for GLANCE versus the strongest baseline, at least at the 10% setting in Table 1? This does not require retraining and would clarify the stability of the smaller gains.

PRS can suppress a candidate because of an earlier candidate that is ultimately discarded. Can the authors quantify how often this occurs and whether it correlates with failures relative to Greedy? This would help establish the limits of the parallel approximation.

4 comments
AuthorsOpenReview ↗

Response to W1 & Q1 (Selector contribution under matched schedules)

W1:Cross-method comparisons do not fully isolate the new selector's contribution. Tables 2 and 14 usefully separate IUS–PRS from input expansion and late withdrawal, but competing methods have not been evaluated under the same schedule. For example, at the 20% setting on Qwen3-VL, the selector alone scores 62.3, matching V-CAST's displayed mean in Table 1, whereas the complete configuration reaches 63.0. The main results therefore support the accuracy–latency trade-off of the complete system, but do not attribute all of that advantage to IUS–PRS. Comparing the strongest baseline with matched input quotas and withdrawal depth would better identify the methodological contribution. This is an addressable evidence gap and does not invalidate the measured benefit of the complete configuration.

Q1:Could at least one strongest baseline be evaluated with the same input quotas and late-withdrawal schedule, reporting accuracy and TTFT against IUS–PRS? This would directly affect my assessment of the selector's incremental contribution.

Thank you for your insightful comments.

To isolate the selector's incremental contribution, we conducted a controlled comparison of IUS–PRS, V-CAST, and FlashVID's before-LLM selector across three shared schedules on Qwen3-VL-8B at a nominal 10% retention budget. Within each schedule, input token counts, withdrawal depth and rule, and visual-token counts at every decoder layer match exactly across all methods for every example, including integer rounding.

  • S0: approximately 10% input retention, without input expansion or layer-wise withdrawal.
  • S1: approximately 12.5% input retention, with all visual tokens removed prior to the final quarter of decoder layers, yielding approximately 9.375% layer-averaged retention.
  • S2: identical input packing as S1, followed by a shared FlashVID-style partial withdrawal rule, yielding approximately 10% layer-averaged retention.

Here, FlashVID-selector denotes its before-LLM compression evaluated under these common schedules. Aggregate mean accuracy weights Video-MME, LongVideoBench, MLVU, and MVBench equally.

Table R1. Shared-schedule accuracy and latency. Each cell reports the equally weighted four-benchmark mean accuracy (%) / model-side TTFT (ms).

Shared scheduleIUS–PRSV-CASTFlashVID-selector
S061.13 / 232.7759.34 / 236.6859.41 / 279.58
S161.43 / 235.8160.25 / 239.4660.27 / 284.00
S261.32 / 237.3060.23 / 240.0860.43 / 285.40

As shown in Table R1, IUS–PRS consistently achieves the highest four-benchmark mean accuracy and the lowest model-side TTFT under every shared schedule. Under S0—where neither input expansion nor withdrawal is applied—it outperforms V-CAST by +1.79 percentage points and FlashVID-selector by +1.72. Under S1, the corresponding gains are +1.18 and +1.16 points, respectively. Under S2, the gains are +1.09 and +0.89 points, respectively. Together, these controlled comparisons provide direct evidence of an incremental selector contribution under matched input and decoder budgets.

AuthorsOpenReview ↗

Response to W2 & Q2 (video-clustered paired intervals)

W2:The robustness of the smaller main-table accuracy gains is not directly quantified. Appendix D does provide appropriate paired bootstrap analyses, but its accuracy diagnostics concern separate subsets and component comparisons; they are not confidence intervals for the main four-benchmark means or differences from the strongest external baselines. This matters for the approximately 0.4-point advantage on LLaVA-Video at 10% retention and for the task-level regressions observed when adding uniqueness. The current results establish competitive performance, but paired analyses of the full prediction sets would clarify whether the smaller advantages are robust.

Q2:Could the existing full prediction sets be used to report video-clustered paired intervals for GLANCE versus the strongest baseline, at least at the 10% setting in Table 1? This does not require retraining and would clarify the stability of the smaller gains.

Thank you for your insightful comments.

To address this, we conducted video-clustered paired analyses at nominal 10% retention, comparing GLANCE with native FlashVID on Qwen3-VL-8B and LLaVA-Video-7B, and with V-CAST on Qwen3.8-27B, without retraining.

We evaluate both methods on Video-MME, LongVideoBench, MLVU, and MVBench.

We report the measured paired accuracy differences and 95% confidence intervals separately for each benchmark, together with the equally weighted four-benchmark mean. Separately for each backbone, the four benchmark comparisons and their mean form one Holm family of five tests. Sorting the unadjusted pp values in ascending order, we apply the Holm step-down correction:

p(i)Holm=min⁡(1,max⁡1≤k≤i[(6−k)p(k)]).p_{(i)}^{\mathrm{Holm}}=\min\left(1,\max_{1\le k\le i}\left[(6-k)p_{(k)}\right]\right).

Table R2. Video-clustered paired comparisons at nominal 10% retention. Each cell reports GLANCE minus the native baseline in accuracy points and an unadjusted 95% confidence interval. The final row reports Holm-adjusted pp values for the four-benchmark mean. Baselines are native FlashVID for Qwen3-VL-8B and LLaVA-Video-7B, and V-CAST for Qwen3.8-27B.

BenchmarkQwen3-VL-8BLLaVA-Video-7BQwen3.8-27B
Video-MME+2.37 [1.11, 3.63]+1.56 [0.37, 2.78]+2.67 [1.22, 4.11]
LongVideoBench+1.57 [0.29, 2.85]+0.52 [0.26, 0.78]+0.22 [0.05, 0.40]
MLVU+1.81 [0.51, 3.16]+0.37 [0.14, 0.59]−0.41 [−2.01, 1.20]
MVBench−0.08 [−1.25, 1.10]+0.24 [0.12, 0.36]+0.34 [0.10, 0.58]
Four-benchmark mean+1.42 [0.79, 2.05]+0.67 [0.36, 0.98]+0.70 [0.16, 1.24]
Holm pp for mean0.000050.000150.03547

The four-benchmark mean differences are +1.42 points for Qwen3-VL-8B (95% CI [0.79, 2.05]; Holm p=0.00005p=0.00005), +0.67 for LLaVA-Video-7B ([0.36, 0.98]; p=0.00015p=0.00015), and +0.70 for Qwen3.8-27B ([0.16, 1.24]; p=0.03547p=0.03547). The mean improvements remain statistically significant after Holm correction across all three backbones (adjusted p<0.05p<0.05). The benchmark breakdown also retains task-specific regressions on MVBench for Qwen3-VL-8B and on MLVU for Qwen3.8-27B.

AuthorsOpenReview ↗

Response to Q3 (Suppression by discarded predecessors)

Q3: PRS can suppress a candidate because of an earlier candidate that is ultimately discarded. Can the authors quantify how often this occurs and whether it correlates with failures relative to Greedy? This would help establish the limits of the parallel approximation.

Thank you for your insightful comments.

To examine the impact of discarded predecessors, we conducted a retained-set recomputation diagnostic and a matched comparison with selected-set Greedy.

Selection diagnostic. We analyzed videos at both 32 and 128 frames using Qwen3-VL-8B under S1 at nominal 10% retention. For each original PRS-selected set SS, we recomputed redundancy penalties using only preceding candidates retained in SS, while keeping the candidate order and token budget fixed. This produces a counterfactual selection S′S' that excludes suppression by discarded predecessors.

The pooled replacement fraction, ∑v∣Sv′∖Sv∣∑v∣Sv∣\frac{\sum_v |S'_v \setminus S_v|}{\sum_v |S_v|}, is 6.04% at 32 frames and 6.29% at 128 frames. These values quantify changes in the retained subset induced by removing discarded-predecessor penalties, without implying equivalent losses in useful information or prediction accuracy.

Table R3. Selection changes, accuracy disagreements, and efficiency. The lower value in each latency pair is bold. Accuracy disagreements are evaluated on Video-MME, LongVideoBench, and MLVU at each frame count. Latencies are from the separate matched H100 profiling experiments in the submitted paper (Table 3(c), Table 16, and Appendix C.6).

Diagnostic32 frames128 frames
Retained positions replaced6.04%6.29%
PRS wrong / Greedy correct5750
PRS correct / Greedy wrong7853
Candidate selection: PRS / Greedy (ms)0.66 / 8.371.07 / 8.87
Complete compression: PRS / Greedy (ms)1.68 / 9.323.50 / 11.42
Model-side TTFT: PRS / Greedy (ms)288.75 / 299.771100.84 / 1107.54

Under identical token budgets and the S1 schedule, PRS and selected-set Greedy achieve comparable three-benchmark mean accuracy at 128 frames (65.81% versus 65.79%). Their predictions differ in both directions: Greedy corrects 50 PRS errors, while PRS correctly answers 53 questions missed by Greedy. Thus, the approximation affects individual predictions but has little observed impact on aggregate accuracy.

Meanwhile, PRS reduces candidate-selection latency from 8.87 to 1.07 ms (8.3$\times$) and complete compression latency from 11.42 to 3.50 ms (3.3$\times$) at 128 frames, with a smaller reduction of 6.70 ms in model-side TTFT. These timings come from the submitted paper's matched profiling setup, separate from the benchmark accuracy diagnostic.

Overall, these results indicate that discarded-predecessor suppression changes a limited fraction of retained tokens, while PRS maintains comparable aggregate accuracy to sequential Greedy with substantially lower subset-construction overhead.

Dear Reviewer pbqc,

We hope this message finds you well. We sincerely thank you for the time and effort you have devoted to reviewing our paper.

We have posted our response to your comments, where we did our best to address each of your concerns. As the discussion period is approaching its end, we would be grateful if you could kindly take a look when convenient, and let us know whether our response addresses your concerns or whether anything remains unclear. We would be happy to provide further clarifications promptly within the remaining time.

Thank you very much again for your valuable feedback, which has helped us improve the paper.

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

In this paper, the authors propose GLANCE, a training-free, query-agnostic video token compressor that compares redundancy within compact candidate pools by utilizing Informativeness and Uniqueness-guided Screening measurements to achieve this. Moreover, Parallel Redundancy Suppression is proposed to suppress long-history token candidates. In this end, this paper leverages Video-LLM backbones like Qwen3-VL to conduct evaluations on various benchmarks to obtain competitive performance while maintaining efficiency and speedups.

Critical strengths

The experimental results with Qwen3-VL and LLaVA-Video on four video benchmarks seem competitive and lead to improved inference efficiency speedups.

Critical weaknesses
  1. The main concern of this paper lies in AI-generated paper writing and presentation, obviously. I can point out some examples while lots of others left for further investigate.

    (1) Regarding Fig. 2, the framework is full of words and texts while the display is also quite blur to read, that raise the concern of all AI-generated framework presentation.

    (2) Lines 157, 211, 212, 252, 274, 343, 404, 431, and 464, etc., there are quite a lot of semicolons in the main paper.

    (3) 234, jth should be j^{th} in LaTeX formatting.

    (4) Line 254, 320&321, 322&323, 352, 370&371&372&373, 376&377, and 412&422&423, etc., I believe these sentences need to be investigated for typos, as they are very difficult to understand.

  2. What is the essential motivation of this paper, which exposes another concern of this paper: the fragmented and confused contributed novelty for the videoLLM visual token reduction task?

  3. What are the main differences from previous works regarding the computed metrics, when compared with diversity scores or something else?

  4. In the experiments, why choose the video-minute-long subset instead of the full set?

Questions for authors

This paper raises serious readability problems and needs careful human revision. Moreover, the motivation and essential novelty are very ambiguous and need to be demonstrated and clarified. Also, please refer to my detailed questions above.

6 comments
AuthorsOpenReview ↗

Response to Weakness 1 (Writing, Figure 2, and notation)

Response to Weakness 1 (Writing, Figure 2, and notation)

The main concern of this paper lies in AI-generated paper writing and presentation, obviously. I can point out some examples while lots of others left for further investigate.

(1) Regarding Fig. 2, the framework is full of words and texts while the display is also quite blur to read, that raise the concern of all AI-generated framework presentation.

(2) Lines 157, 211, 212, 252, 274, 343, 404, 431, and 464, etc., there are quite a lot of semicolons in the main paper.

(3) 234, jth should be j^{th} in LaTeX formatting.

(4) Line 254, 320&321, 322&323, 352, 370&371&372&373, 376&377, and 412&422&423, etc., I believe these sentences need to be investigated for typos, as they are very difficult to understand.

We appreciate you pointing out these specific presentation and readability issues.

Figure 2. The figure explains the method through three linked stages: scoring the tokens, screening a candidate pool, and applying PRS to that pool. Its worked example shows how a candidate with high initial utility can be discounted because it resembles an earlier candidate. The top part connects the retained tokens to the inference configuration. The labels identify quantities defined in Equations (1)–(5), and the caption (lines 185–191) specifies the schematic nature of the example. The three stages correspond to Section 3 and Algorithm 1.

Punctuation and notation. At line 157, the two clauses distinguish budget allocation across groups from candidate scoring within a group. Lines 211–213 distinguish the cost of the video-wide cue from explicit candidate comparisons and explain the normalization difference from VidCom2. We will typeset the ordinal at line 234 as jthj^{\mathrm{th}} and clean up the punctuation throughout the text.

Definitions and complexity. Lines 250–254 give the complete selector complexity, identify candidate pools as the only source of quadratic comparisons, and state how cost scales with the number of groups. At nominal 10% retention with input expansion, kf≈P/8k_f\approx P/8 and mf≈P/4m_f\approx P/4, giving mf2/P2≈1/16m_f^2/P^2\approx 1/16. Line 254 specifies the meaning of this ratio: “This compares similarity-matrix sizes, not measured speedups.”

Experimental reporting. Lines 320–323 map the three backbones to 32, 64, and 32 frames, respectively, and state the shared evaluation settings. Line 352 points to the LLaVA-Video 20% results in Appendix A.2. Lines 368–373 distinguish the selector-only means (61.1% and 62.3%) from the complete configuration (61.4% and 63.0%), explicitly attributing part of the gains to the schedule. Section 4.4 names the three frame-scaling benchmarks, and lines 412–413 report accuracy, TTFT, and peak memory together to describe that trade-off. Finally, lines 422–423 explain the complementary scoring cues and explicitly state that token-level evidence retention remains unmeasured. These passages identify the quantities being compared and their scope, which we will make more transparent in the revision.

AuthorsOpenReview ↗

Response to Weakness 2 (Essential motivation and methodological contribution)

What is the essential motivation of this paper, which exposes another concern of this paper: the fragmented and confused contributed novelty for the videoLLM visual token reduction task?

We appreciate you pointing out these specific presentation and readability issues.

The essential motivation is to preserve useful video evidence while keeping the cost of constructing the retained subset low enough to realize a net inference benefit. This is the tension described in the Introduction (lines 45–100): independent token ranking is inexpensive but can repeatedly select similar evidence, while selection that updates redundancy against an evolving subset introduces sequential work. Denser frame sampling makes both the token budget and the selection overhead more consequential.

GLANCE addresses this through two connected stages. IUS screens the full token population into small candidate pools, using feature-response rank and video-wide directional uniqueness. PRS resolves candidate redundancy in parallel, using a fixed utility order. Screening limits where pairwise comparisons occur, and the fixed reference order removes the sequential selected-set updates. The method is independent of the question and decoder attention maps. This is the concrete meaning of the paper's “screen broadly, compare selectively” principle.

The novelty lies in this selection procedure and its measured quality–cost trade-off: inexpensive global screening followed by fixed-order parallel suppression within budget-controlled pools. Section 2 places it in the context of earlier methods using feature norms, cosine distance, diversity, and score suppression. Sections 3.1–3.2 specify how GLANCE constructs and combines its scores and redundancy references; the contribution is the resulting algorithm rather than an individual distance primitive.

The existing ablations directly test this design. In Table 3(a,b), adding PRS to IUS raises the four-benchmark mean from 60.3% to 61.4% on Qwen3-VL and from 60.6% to 61.0% on LLaVA-Video. Table 3(d) tests the screening pool: increasing it from kfk_f to 2kf2k_f raises the Qwen3-VL mean from 60.3% to 61.4%, while larger pools increase TTFT without improving the displayed mean. Most directly, Table 3(c) fixes IUS scores, quotas, candidate pools for Greedy/PRS, and decoder depth while comparing the subset-construction rules:

Table W2. Existing evidence for the selection quality–cost trade-off. Reproduced from Table 3(c) of the submitted paper. Accuracy is the three-task mean over Video-MME, LongVideoBench, and MLVU. Selection times come from the separate profiling cohort specified in Appendix C.6.

Subset constructionMean accuracy (%)Selection time (ms)
IUS + direct Top-kk59.30.15
IUS + selected-set Greedy60.08.37
IUS + PRS60.20.66

With the same candidate pools and inference settings, PRS reduces selection time from 8.37 to 0.66 ms while yielding closely comparable observed accuracy (60.2% versus 60.0%). This directly tests the motivation for parallel refinement. The contribution established by this comparison is lower construction cost; Section 4.6 and Appendix D distinguish this from a claim of accuracy superiority over Greedy. Section 3.3 separately identifies input expansion and late withdrawal as inference settings following prior work, and Table 2 isolates their contribution from IUS–PRS.

AuthorsOpenReview ↗

Response to Weakness 3 (Differences from existing diversity metrics)

Sections 2 and 3 distinguish the metrics by their definitions, reference sets, and roles in selection. IUS combines a token's relative response strength within its visual group with a video-wide directional statistic. PRS then evaluates redundancy among the screened candidates. Following Section 3, each of the FF visual groups contains PP feature vectors of dimension DD, so N=FPN=FP, and kfk_f is the output quota for group ff.

First, Equation (1) defines If,p=af,p/(P−1)I_{f,p}=a_{f,p}/(P-1), where af,pa_{f,p} is the ascending, zero-based rank of the original feature norm within group ff. This makes response strength relative to its group and invariant to a common positive rescaling of that group's features. It is a screening proxy, without requiring attention extraction.

Second, Equation (2) defines

v^f,p=vf,p∥vf,p∥2.\hat v_{f,p}=\frac{v_{f,p}}{\lVert v_{f,p}\rVert_2}.

cg=1N∑f′,p′v^f′,p′.c_g=\frac{1}{N}\sum_{f',p'}\hat v_{f',p'}.

=\frac{1}{N}\sum_{f',p'}\left(1-\hat v_{f,p}^{\top}\hat v_{f',p'}\right).$$ The mean $c_g$ is not renormalized. Thus $g$ is the mean cosine distance to all sampled video tokens and can be computed in $O(ND)$ without a video-wide pairwise matrix. Each token contributes a unit direction to the mean. As discussed in Section 2 and lines 209–213, this differs from using cosine similarity to means of the original, unnormalized features, as in VidCom2. V-CAST adds cosine distance to the frame mean to a min–max-normalized feature norm. IUS instead uses a within-group rank, a video-wide mean of unit directions, and the multiplicative utility $U=R(I\odot R(g))$ in Equation (4), where $R$ is video-wide min–max normalization. Third, the global statistic does not determine redundancy within the final subset. For each group, IUS retains $m_f=\min(2k_f,P)$ candidates. With their utility order fixed, Equation (5) gives $$\rho_1=0.$$ $$\rho_j=\max\left(0,\max_{i < j}x_j^{\top}x_i\right)\quad(j > 1).$$ $$s_j=u_j(1-\rho_j).$$ All candidate penalties are computed by batched similarities and a lower-triangular mask, followed by selection of the highest $k_f$ adjusted scores. This compares candidates within each visual group, while the video-wide cue enters through IUS. **Table W3. Reference sets and execution rules.** The prior-work descriptions follow Section 2 of the submitted paper. | **Construction** | **Reference for distinctiveness / redundancy** | **Selection rule** | | --- | --- | --- | | VidCom2 scoring | Frame and video means of the original features | Combines cosine-similarity cues and uses uniqueness for budget allocation | | V-CAST scoring | Frame-mean cosine distance and min–max-normalized feature norm | Adds the two cues, with trajectory curvature used for budget allocation | | Selected-set diversity methods | Currently selected tokens, as in DivPrune, IDPruner, and CDPruner | Sequential selection with updates to the selected-set criterion | | IUS screening | Within-group norm rank and the mean of all video-token unit directions | Multiplicative utility, uniform group quotas, then a small shortlist | | PRS refinement | All earlier candidates in a fixed utility order within the shortlist | Parallel penalties, then Top-$k_f$ by adjusted score | The distinction from selected-set diversity is therefore operational as well as mathematical: PRS fixes its redundancy reference set so that penalties can be computed together. Section 3.2 (lines 243–246) and Appendix C.4 explicitly describe the resulting approximation: an earlier candidate may contribute to a penalty even if it is ultimately discarded. Section 4.6 tests this choice against selected-set Greedy. The complete selector cost is $O(ND+FP\log P+\sum_f m_f^2D)$, as given in Section 3.2. Quadratic comparisons occur only inside the candidate pools.
Reviewer_aicJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

GLANCE is a training-free, query-agnostic visual token compressor for video language models. Informativeness and Uniqueness-guided screening combines within-group feature-norm ranks with video-wide directional uniqueness to construct small candidate pools. Parallel Redundancy Suppression discounts candidates according to their similarity to earlier candidates in a fixed utility order, avoiding sequential selected-set updates. The complete inference configuration additionally expands the input token budget and withdraws visual tokens in late decoder layers. Experiments across three backbones and four multiple-choice QA benchmarks show favorable accuracy–latency trade-offs, with additional frame-scaling and component analyses.

Critical strengths
  1. The paper addresses the practical cost of subset construction, rather than treating token selection as free. Screening before candidate-level comparison is simple and computationally motivated.

  2. Evaluation spans three backbones, four benchmarks, multiple retention settings, and frame counts. The matched greedy comparison directly investigates the quality–cost trade-off of parallel suppression. The manuscript distinguishes the selector from its inference schedule and reports important limitations, including unmatched schedules and the scope of statistical diagnostics.

Critical weaknesses
  1. The main comparisons do not fully isolate the selector’s contribution. Input expansion and late visual-token withdrawal are part of the complete configuration, but competing selectors are not rerun under the same schedule. The existing GLANCE-only ablations are useful, yet a matched-schedule comparison with the strongest baselines is needed to attribute the gains.

  2. All evaluated benchmarks use multiple-choice QA. Retaining answer-discriminative evidence may suffice without preserving comprehensive video content. Detailed or dense captioning, evaluated for event coverage, factual accuracy, temporal ordering, and hallucinations, would strengthen the evidence for broad information preservation. PRS suppresses a candidate using all preceding candidates, including those ultimately discarded. This can penalize evidence complementary to the final subset. The greedy comparison is informative, but the frequency and consequences of this approximation remain insufficiently characterized.

Questions for authors
  1. Under identical input token counts, withdrawal depth, and withdrawal rule, how does IUS–PRS compare with the strongest competing selectors? Would the relative advantage remain at matched measured latency?

  2. Can the authors provide detailed or dense-captioning results that assess omitted events and fine-grained factual information, beyond multiple-choice answer accuracy?

  3. How often is a retained-set decision affected by a predecessor that is itself discarded? Can failure cases or a diagnostic comparison with selected-set greedy suppression clarify when PRS over-suppresses useful candidates?

4 comments
AuthorsOpenReview ↗

Response to W1 & Q1 (Selector contribution and matched schedules)

W1:The main comparisons do not fully isolate the selector's contribution. Input expansion and late visual-token withdrawal are part of the complete configuration, but competing selectors are not rerun under the same schedule. The existing GLANCE-only ablations are useful, yet a matched-schedule comparison with the strongest baselines is needed to attribute the gains.

Q1:Under identical input token counts, withdrawal depth, and withdrawal rule, how does IUS–PRS compare with the strongest competing selectors? Would the relative advantage remain at matched measured latency?

Thank you for your insightful comments.

To better isolate the selector's contribution, we conducted additional experiments comparing IUS–PRS with V-CAST and FlashVID's pre-LLM selector under three matched inference schedules on Qwen3-VL-8B at nominal 10% retention. All methods share identical input token counts, withdrawal rules and depths, and layer-wise visual-token budgets, including integer rounding.

  • S0: approximately 10% input retention, without input expansion or layer-wise withdrawal.
  • S1: approximately 12.5% input retention, with all visual tokens removed prior to the final quarter of decoder layers, yielding approximately 9.375% layer-averaged retention.
  • S2: identical input packing as S1, followed by a common FlashVID-style partial withdrawal rule, yielding approximately 10% layer-averaged retention.

Here, FlashVID-selector denotes its before-LLM compression evaluated under these shared controls. Aggregate mean accuracy weights Video-MME, LongVideoBench, MLVU, and MVBench equally.

Table R1. Controlled selector comparison. Each cell reports four-benchmark mean accuracy (%) / model-side TTFT (ms).

Shared scheduleIUS–PRSV-CASTFlashVID-selector
S061.13 / 232.7759.34 / 236.6859.41 / 279.58
S161.43 / 235.8160.25 / 239.4660.27 / 284.00
S261.32 / 237.3060.23 / 240.0860.43 / 285.40

Under S0—where neither input expansion nor withdrawal is applied—IUS–PRS outperforms V-CAST by +1.79 percentage points and FlashVID-selector by +1.72. Under S1, the corresponding performance gains reach +1.18 and +1.16 .

Precisely matching measured latency across selectors is challenging because their selection overheads differ and token budgets change discretely. We therefore fix the input token counts, withdrawal depths and rules, and layer-wise visual-token budgets, and compare the resulting accuracy and measured TTFT. Under these controlled conditions, IUS–PRS simultaneously achieves the highest four-benchmark mean accuracy and the lowest model-side TTFT under each of the three shared schedules (Table R1), demonstrating the strongest observed accuracy–latency trade-off among the evaluated selectors under these budgets and schedules.

AuthorsOpenReview ↗

Response to W2 & Q2 (Detailed captioning and information preservation)

W2:All evaluated benchmarks use multiple-choice QA. Retaining answer-discriminative evidence may suffice without preserving comprehensive video content. Detailed or dense captioning, evaluated for event coverage, factual accuracy, temporal ordering, and hallucinations, would strengthen the evidence for broad information preservation.

Q2:Can the authors provide detailed or dense-captioning results that assess omitted events and fine-grained factual information, beyond multiple-choice answer accuracy?

Thank you for your insightful comments.

We evaluate detailed descriptions on the VideoDetailCaption dataset using Qwen3-VL-8B and Qwen3.8-27B with 32 input frames, and LLaVA-Video-7B with 64 input frames. GLANCE (IUS–PRS) and V-CAST share the same S1 schedule at nominal 10% and 20% retention within each backbone. Generation uses the official prompts, greedy decoding, and a common 2,048-token output limit. Reference descriptions are supplied only during evaluation.

All caption metrics are evaluated by gpt-5.6-luna at temperature zero. EC, FA, TCO, and VC use the same 64 uniformly sampled chronological frames, reference descriptions, and anonymous generated captions across methods.

Event coverage (EC). Percentage of visually supported events in a common inventory that are fully covered by the generated caption. The inventory is constructed from the frames and reference before inspecting generated captions.

Factual support (FA). Percentage of supported factual claims among claims judged either supported or contradicted by the frames. Claims labeled not observed, unclear, or nonvisual are excluded from this denominator.

Temporal coverage and order (TCO). Percentage of visually verified event pairs with a clear before-relation for which the caption fully covers both events and expresses their order correctly.

Visual contradiction (VC). Percentage of contradicted claims under the same denominator as FA, used as a hallucination proxy.

Semantic-match accuracy (YesAcc). Percentage of yes judgments among valid yes/no judgments of semantic agreement with the reference, assessed using the question, reference, and generated description without video frames.

Table R2. Caption accuracy and four-dimensional audit on VideoDetailCaption. All entries are percentages under the shared S1 schedule. Best values within each backbone and retention setting are bold. YesAcc: yes/no semantic-match accuracy; EC: event coverage; FA: factual support; TCO: temporal coverage and order; VC: visual contradiction.

SelectorRetentionYesAcc ↑EC ↑FA ↑TCO ↑VC ↓
Qwen3-VL-8B
GLANCE (IUS–PRS)10%43.0923.8887.407.9712.60
V-CAST10%41.6823.4587.066.9212.94
GLANCE (IUS–PRS)20%51.1031.0088.7110.3311.29
V-CAST20%47.7029.5588.649.3711.36
Qwen3.8-27B
GLANCE (IUS–PRS)10%44.2920.2386.675.5013.33
V-CAST10%42.0820.0785.885.1514.12
GLANCE (IUS–PRS)20%45.4922.0688.117.6411.89
V-CAST20%45.0921.0986.546.3513.46
LLaVA-Video-7B
GLANCE (IUS–PRS)10%22.248.1665.721.1434.28
V-CAST10%15.435.9265.191.0234.81
GLANCE (IUS–PRS)20%22.449.3567.621.5332.38
V-CAST20%18.647.2864.290.9135.71

GLANCE consistently outperforms V-CAST on all five captioning metrics across all three backbones at both 10% and 20% retention, demonstrating stronger preservation of detailed video information.

AuthorsOpenReview ↗

Response to W3 & Q3 (Suppression by discarded predecessors)

W3:PRS suppresses a candidate using all preceding candidates, including those ultimately discarded. This can penalize evidence complementary to the final subset. The greedy comparison is informative, but the frequency and consequences of this approximation remain insufficiently characterized.

Q3:How often is a retained-set decision affected by a predecessor that is itself discarded? Can failure cases or a diagnostic comparison with selected-set greedy suppression clarify when PRS over-suppresses useful candidates?

Thank you for your insightful comments.

To examine the impact of discarded predecessors, we conducted a retained-set recomputation diagnostic and a matched comparison with selected-set Greedy.

Selection diagnostic. We analyzed videos at both 32 and 128 frames using Qwen3-VL-8B under S1 at nominal 10% retention. For each original PRS-selected set SS, we recomputed redundancy penalties using only preceding candidates retained in SS, while keeping the candidate order and token budget fixed. This produces a counterfactual selection S′S' that excludes suppression by discarded predecessors.

The pooled replacement fraction, ∑v∣Sv′∖Sv∣∑v∣Sv∣\frac{\sum_v |S'_v \setminus S_v|}{\sum_v |S_v|}, is 6.04% at 32 frames and 6.29% at 128 frames. These values quantify changes in the retained subset induced by removing discarded-predecessor penalties, without implying equivalent losses in useful information or prediction accuracy.

Table R3. Selection changes, accuracy disagreements, and efficiency. The lower value in each latency pair is bold. Accuracy disagreements are evaluated on Video-MME, LongVideoBench, and MLVU at each frame count. Latencies are from the separate matched H100 profiling experiments in the submitted paper (Table 3(c), Table 16, and Appendix C.6).

Diagnostic32 frames128 frames
Retained positions replaced6.04%6.29%
PRS wrong / Greedy correct5750
PRS correct / Greedy wrong7853
Candidate selection: PRS / Greedy (ms)0.66 / 8.371.07 / 8.87
Complete compression: PRS / Greedy (ms)1.68 / 9.323.50 / 11.42
Model-side TTFT: PRS / Greedy (ms)288.75 / 299.771100.84 / 1107.54

Under identical token budgets and the S1 schedule, PRS and selected-set Greedy achieve comparable three-benchmark mean accuracy at 128 frames (65.81% versus 65.79%). Their predictions differ in both directions: Greedy corrects 50 PRS errors, while PRS correctly answers 53 questions missed by Greedy. Thus, the approximation affects individual predictions but has little observed impact on aggregate accuracy.

Meanwhile, PRS reduces candidate-selection latency from 8.87 to 1.07 ms (8.3$\times$) and complete compression latency from 11.42 to 3.50 ms (3.3$\times$) at 128 frames, with a smaller reduction of 6.70 ms in model-side TTFT. These timings come from the submitted paper's matched profiling setup, separate from the benchmark accuracy diagnostic.

Overall, these results indicate that discarded-predecessor suppression changes a limited fraction of retained tokens, while PRS maintains comparable aggregate accuracy to sequential Greedy with substantially lower subset-construction overhead.

Dear Reviewer aicJ,

We hope this message finds you well. We sincerely thank you for the time and effort you have devoted to reviewing our paper.

We have posted our response to your comments, where we did our best to address each of your concerns. As the discussion period is approaching its end, we would be grateful if you could kindly take a look when convenient, and let us know whether our response addresses your concerns or whether anything remains unclear. We would be happy to provide further clarifications promptly within the remaining time.

Thank you very much again for your valuable feedback, which has helped us improve the paper.

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

Related papers

Discussion (0)

Sign in to comment.