{"id":"ae59371b-1795-4eb1-81e4-a9e533987ad5","arxiv_id":"2607.23265","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Wavelet-based temporal and spatial token condensation preserves 99.6% of full-token video question-answering accuracy at 10x compression without retraining.","lead":"WaveZip is a training-free method that uses wavelet transforms to cut the number of visual tokens fed into video language models, keeping 99.6% of full accuracy at a 10x compression. It is a practical efficiency recipe for long-video understanding that anyone with an off-the-shelf LVLM can bolt on.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Frozen BLIP-ITM scorer's alignment with target-LVLM importance is unverified; if misaligned, the 10x budget can discard critical evidence.","rationale":"The reader's weakest assumption — that BLIP-ITM cross-modal relevance and saliency are faithful proxies for the target LVLM's needs — is also the load-bearing concern from my reading. The method's spatial and temporal compression decisions are entirely gated by these external signals, yet the paper provides no direct evidence tying them to the LVLM's internal importance. The strongest claim (99.6% retention at 10x) is an average over three benchmarks on one backbone, and the main table shows WaveZip losing 1.7 points on LongVideoBench at ρ=0.1; the abstract's 'consistently outperforming' also relies on averages. The scorer alignment issue is more fundamental than missing error bars or threshold τ: if the proxy is misaligned, the mechanism can discard evidence that the LVLM would use, and no amount of wavelet tuning fixes that. The ablations (Figs. 4, 8, Table 4) show the components help relative to uniform/raw baselines under the same scorer, but they do not test whether the scorer is the right source of guidance. I therefore agree with the reader's assessment and recommend keeping the CONDITIONAL verdict: the paper is plausible and well-ablated, but the central proxy assumption needs a direct test. My proposed experiment provides that test without requiring new training or major reimplementation.","tokens_in":15268,"tokens_out":6407,"duration_ms":101289,"concrete_test":"On 200 random examples from VideoMME and LongVideoBench with LLaVA-OneVision-7B, compute LVLM-based per-frame importance (e.g., attention rollout over visual tokens, or gradient-norm saliency) and Spearman-correlate it with BLIP-ITM r_t. Then rerun WaveZip at ρ=0.1 replacing BLIP r_t with the LVLM-importance ranking while keeping spatial M_t fixed. If the LVLM-importance allocation improves average accuracy by more than ~1.5 points, the frozen scorer is a material bottleneck; if the difference is within noise, the proxy assumption is validated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"WaveZip's central 99.6%-retention claim rests on the assumption (Sec. 3.2.1) that a frozen BLIP-ITM scorer's r_t and M_t are faithful proxies for what the target LVLM actually needs. These signals drive the frame filter, the WTA token budget (Eq. 16), and the WSM spatial gate (Eq. 13-14). No experiment measures this alignment. A concrete failure mode: a frame with low global image-text relevance can still contain the decisive evidence (e.g., a transient object or action). At ρ=0.1, with 64 frames × 196 tokens, B≈1254 tokens, so the average budget is ~20 tokens/frame; an evidence-bearing frame with low r_t can be filtered out or assigned near-zero tokens. Appendix E is consistent with this: at ρ=0.2 the largest losses are Action Recognition (−3.83), Object Reasoning (−2.86), Temporal Reasoning (−2.26), and OCR (−2.16), categories where BLIP-ITM global matching is likely misaligned with the downstream task. Table 4 only swaps one cross-modal scorer for another (BLIP vs BLIP2 vs similarity); it does not compare against the target LVLM's own attention or an oracle. The concern is not that BLIP is 'wrong' in general, but that the paper never verifies the proxy relation on which the entire compression mechanism depends.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"WaveZip is a training-free video token condensation method for LVLMs. It uses a frozen BLIP-ITM cross-modal scorer to derive frame-level relevance scores r_t and spatial saliency maps M_t; a Visual Change Estimator filters frames below a threshold τ; a wavelet temporal allocator rectifies r_t via 1D DWT with visual-change gating; a wavelet spatial modulator reweights high-frequency 2D DWT subbands by saliency; and a token assigner distributes a global budget B = ρN_v among retained frames. Experiments on EgoSchema-500, LongVideoBench, VideoMME, and LVBench across LLaVA-OneVision-7B, LLaVA-Video-7B, and Qwen2.5-VL report that WaveZip retains 99.6% of full-token average accuracy at ρ=0.1 and claims consistent average gains over state-of-the-art baselines. The paper includes controlled module ablations, wavelet-basis ablations, an efficiency analysis, and a query-type breakdown.","tokens_in":15644,"tokens_out":5073,"duration_ms":54196,"significance":"If the central empirical claim holds, the method is practically significant: it is training-free, plug-and-play, and shows consistent average gains across three backbones and several benchmarks, including an additional LVBench generalization test. The controlled ablations in Appendix B (Fig. 8) are a genuine strength, as are the honest query-type breakdown in Appendix E and the component-level runtime breakdown in Appendix F. However, the headline result is an average over a small benchmark suite, per-benchmark gains are not uniform, all numbers are single-run with no error bars, and the central cross-modal proxy assumption is not directly validated. These gaps currently under-support the strength of the claims.","major_comments":[{"comment":"The headline 'retains 99.6%' and the claimed consistent improvement rest on single-run accuracy differences as small as 0.1–1.1 points. For a 500-sample subset, a 1-point difference is well within binomial sampling error; even on the full benchmarks, a 0.9-point average gap between WaveZip (58.8) and VFlowOpt (57.9) at ρ=0.1 is not interpretable without variance. Please report standard errors or paired bootstrap intervals over benchmark items (ideally with multiple seeds). Also, the abstract's 'consistently outperforming' is only true for the average: at ρ=0.1, WaveZip is below DyCoke on EgoSchema (62.3 vs 63.0) and below FastVID on LongVideoBench (54.7 vs 55.7); at ρ=0.1 on Qwen2.5-VL it is 3.6 points below VisionZip on EgoSchema (Table 2). The average claim and the per-benchmark claim should be reported separately and the wording adjusted.","section":"§4.2, Table 1"},{"comment":"The Frame Filter threshold τ is a free parameter that is never reported, ablated, or given a selection criterion. It determines the retained frame set M and directly changes both the temporal budget allocation (Eq. 16) and how the global budget B is spread across retained frames. Since λ is ablated and wavelet basis pairs are ablated, the omission of τ is conspicuous. If τ is nonzero, the effective number of frames can be much smaller than T and the actual per-frame token density changes; if τ=0, the VCE gate still affects WTA but the filter is vacuous. Please report the default τ, state whether it is fixed or tuned per benchmark, and provide a sensitivity sweep. Without this, the 10× compression claim is not reproducible.","section":"§3.2.2, Eq. (8); §4.1 setup"},{"comment":"The entire compression mechanism assumes the frozen BLIP-ITM cross-modal scores r_t and M_t are faithful proxies for what the target LVLM needs. Table 4 only replaces the scorer with another off-the-shelf scorer or with base-LVLM similarity; it does not measure agreement with the target model's own per-token attention, gradients, or an oracle relevance. The category breakdown in Fig. 10 is telling: at ρ=0.2 the largest losses are Action Recognition (−3.83), Object Reasoning (−2.86), Temporal Reasoning (−2.26), and OCR (−2.16), exactly the categories where a global image-text matching model should be most suspect. To support the claim that WaveZip preserves query-relevant evidence, I would expect at minimum a comparison of the frames/tokens selected by WaveZip against those selected by the target LVLM's own attention, or an oracle-ablation showing the ceiling. Without this, the central as","section":"§3.2.1, Table 4, Appendix E"}],"minor_comments":[{"comment":"Typo: 'ues' should be 'uses'.","section":"§3.1"},{"comment":"The text says 'ρ=0.2%' when referring to a retained ratio of 20%; this should be 'ρ=0.2' or '20%'.","section":"§4.2"},{"comment":"The retained ratio ρ is not stated in the table or its caption, making the cross-modal scorer comparison impossible to interpret. Please specify the ρ value and evaluation subset.","section":"Table 4"},{"comment":"The abstract says '10× compression ratio' while §4.1 defines the retained ratio ρ. Please reconcile: does 10× mean ρ=0.1? Clarify the direction of the ratio to avoid ambiguity.","section":"Abstract and §4.1"},{"comment":"These ablation figures report single-run accuracies with no error bars or significance information. Given the small observed differences, adding error bars or at least stating the number of samples would substantially improve interpretability.","section":"Figs. 3, 5, 8"}],"recommendation":"major_revision","confidential_remarks":"This is a promising training-free method with a clear formulation and several honest ablations. The main issues are statistical: the headline average differences are small and single-run, the threshold τ is unreported, and the cross-modal proxy is not validated against the target LVLM. These are fixable within the manuscript's scope; a revised version addressing them could be acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: WaveZip is a genuine new combination—1D DWT for temporal relevance rectification, 2D DWT for saliency-guided spatial detail preservation—and it's training-free, which makes it practically attractive. The average results are consistent across three benchmarks and three backbones, and the controlled ablations in Appendix B actually isolate the mechanism. The efficiency breakdown is honest about the cross-modal scorer overhead, which is a nice change of pace.\n\nSoft spots, in proportion: First, the abstract's \"consistently outperforming\" is not true per-benchmark. On LongVideoBench at ρ=0.1 and 0.15, FastVID beats WaveZip. The advantage is average-level and VideoMME-specific. That wording should be softened. Second, the frame filter threshold τ is never reported. It's a free parameter that could affect results, and the paper should state it. Third, no error bars, single runs, and no code in the manuscript—so the 99.6% retention figure carries less weight than it looks, especially for EgoSchema-500 where a couple of points is a few samples.\n\nThe stress-test concern about BLIP-ITM alignment is fair. The method's temporal budget and spatial modulation both depend on r_t and M_t being faithful to what the target LVLM needs, and the paper never checks that against the LVLM's own attention or an oracle. Appendix E is actually evidence in the stress-test's favor: the largest drops are in Action Recognition (−3.83), Object Reasoning (−2.86), Temporal Reasoning (−2.26), and OCR (−2.16), exactly where a global image-text matcher might misalign. That doesn't sink the paper—average retention is still strong—but it qualifies the headline.\n\nBottom line: This deserves a serious referee. I'd send it to review, with requests for τ, variance or multiple seeds, and a more measured abstract. The core mechanism is sound and likely useful to the video-LVLM efficiency community.","headline":"A real training-free token condensation method with a solid empirical core, but the abstract overclaims and the scorer-alignment assumption deserves scrutiny.","tokens_in":16115,"tokens_out":2681,"would_cite":true,"duration_ms":42447,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":null,"created_at":"2026-08-04T03:28:28.202339+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":null,"supporting_citations":[],"review_version":2}