{"id":"dca2bb67-1892-44b2-8b4a-c8b1c6013b64","arxiv_id":"2608.03269","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A construction-based video distillation pipeline selects teacher-confident clips, allocates slots to feature-space clusters, and blends prototype-anchor pairs with matched soft labels, avoiding gradient updates of stored videos.","lead":"ProtoBlend builds a tiny video training set by keeping the most recognizable clip from each source video, grouping clips by a teacher model's features, and blending pairs of compatible clips into one stored video without optimizing the stored videos directly. The result is competitive accuracy on several action-recognition benchmarks at much lower construction cost than optimization-based video distillation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Efficiency claim rests on unmeasured teacher-preparation and candidate-inference costs; total cost could reverse the Fig. 1 trade-off.","rationale":"The reader flagged teacher preparation in the rationale but selected teacher quality as the weakest assumption. I see teacher quality as a known, self-declared limitation (the Limitations section states the dependence). The more load-bearing issue is the unmeasured total cost, because the paper's headline contribution is 'Efficient' and the only efficiency evidence is a post-teacher, MiniUCF-only time plot. This is a missing measurement rather than a demonstrated false claim, so it does not overturn the CONDITIONAL verdict; it strengthens the condition that the authors must report fair end-to-end cost. The proposed check is concrete and would settle the concern: either the total cost is still lower than optimization-based baselines, or the efficiency claim must be substantially softened.","tokens_in":14046,"tokens_out":7309,"duration_ms":76931,"concrete_test":"Compute and report an end-to-end cost table for all four datasets: GPU-hours for (a) VideoMAE teacher fine-tuning (specify epochs, resolution, batch size, hardware), (b) teacher forward passes over R=5 candidates per source video, and (c) clustering/blending. Compare these totals with the full wall-clock/GPU-hour cost of DM+VDSD and PRISM to construct the same VPC distilled sets on the same GPU. If ProtoBlend's total exceeds the baseline on MiniUCF or on a K400 subset, the efficiency claim is unsupported; if it is below, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the accuracy-efficiency trade-off of construction vs iterative optimization (abstract; Fig. 1). The only time measurements are 'post-teacher construction time' in Fig. 3, and the conclusion explicitly narrows the claim to 'lower post-teacher construction cost.' Two costs are excluded: (i) fine-tuning a dataset-specific VideoMAE teacher on each full training set, and (ii) forward inference over R=5 temporal candidates for every source video (Eqs. 3-4). No wall-clock or GPU-hour figure is given for either component, and no efficiency data is reported for HMDB51, Kinetics-400, or SSv2. The optimization-based baselines (DM+VDSD, PRISM) do not require this teacher preparation, yet their reported 'distillation time' includes the synthesis optimization. If teacher fine-tuning plus candidate inference dominates baseline optimization cost, the Fig. 1 trade-off reverses and the method is efficient only in a narrow post-teacher sense. This is not an internal inconsistency, but it makes the unqualified 'Efficient' title and abstract claim rest on an unmeasured quantity.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"ProtoBlend proposes a construction-based video dataset distillation method that avoids iterative gradient-based optimization of the stored videos. For each source video, a frozen VideoMAE teacher (fine-tuned on the source dataset) scores R=5 temporal candidates and keeps the one with the lowest target-class cross-entropy (TGS, Eq. 3–4). The selected clips of each class are clustered with K-means in teacher feature space, with one distilled slot per cluster (CGA). Within each cluster, the lowest- and second-lowest-loss clips are chosen as prototype and anchor, then blended via pixel-level interpolation with coefficient α (Eq. 8); the teacher posteriors of the two clips are mixed with the same coefficient to form the soft label (Eq. 9). A student is trained from random initialization on the resulting set using a combined hard/soft loss (Eq. 11–13). The paper reports experiments on MiniUCF, HMDB51, Kinetics-400, and Something-Something V2, comparing with coreset baselines and optimization-based distillation methods. The central claims are that ProtoBlend achieves competitive accuracy and a favorable accuracy–efficiency trade-off without optimizing the distilled videos.","tokens_in":14304,"tokens_out":6077,"duration_ms":74891,"significance":"If the efficiency claim holds, ProtoBlend offers a simple, non-iterative alternative to video dataset distillation, whose optimization cost is a recognized bottleneck. The paper has several strengths: the ablations isolate the contributions of temporal selection, allocation, blending, and soft-label supervision; results are repeated over three student runs with standard deviations; hyperparameter sensitivity is reported; cross-architecture transfer is examined; and the limitations are stated explicitly. The method is falsifiable and easy to reproduce in principle. However, the central efficiency claim is not fully substantiated because the cost of preparing the dataset-specific teacher and running teacher inference over all candidates is excluded from the reported time. Furthermore, on the two larger benchmarks the method is at or below simple coreset baselines, weakening the broad claim of competitive performance. No code or supplementary material is provided, which limits verification.","major_comments":[{"comment":"The paper's title and abstract claim \"efficient\" and a favorable accuracy–efficiency trade-off, but the only time reported is \"post-teacher construction time\" in Fig. 3. This excludes two large costs: (i) fine-tuning a dataset-specific VideoMAE teacher on the full training set, and (ii) teacher forward inference over R=5 temporal candidates for every source video (Eqs. 3–4). The text even says ProtoBlend shifts cost to \"a single teacher-inference stage,\" yet the teacher is iteratively trained, not just run once. Since the optimization baselines (DM+VDSD, PRISM) do not require such a teacher, the total-cost comparison in Fig. 1 may reverse once teacher preparation is included. Please report wall-clock/GPU-hours for teacher fine-tuning and candidate inference on all datasets, or explicitly re-scope the efficiency claim to the post-teacher construction phase.","section":"Efficiency Analysis; Fig. 3; Eqs. (3)–(4)"},{"comment":"ProtoBlend uses a VideoMAE-Base teacher fine-tuned separately on each full source dataset. This teacher supplies clip scoring, feature clustering, and soft labels. The optimization-based baselines (DM, VDSD, PRISM) do not have access to such a strong, dataset-specific teacher. This is not circular, but it makes the accuracy comparison less apples-to-apples: the gains could largely reflect the teacher's pretrained and fine-tuned knowledge rather than the select-allocate-blend mechanism itself. To support the claimed advantage, please ablate the teacher by (a) using a fixed pretrained VideoMAE without fine-tuning, (b) giving the baselines the same teacher-derived soft labels or a similarly strong initialization, or at minimum (c) reporting the teacher's full-dataset accuracy and the cost of obtaining it. The Limitations section acknowledges teacher dependence but does not address the compa","section":"Implementation details; Limitations"},{"comment":"The abstract states ProtoBlend demonstrates \"competitive accuracy–efficiency trade-off\" on four benchmarks, but Table 2 shows that on SSv2 at VPC=5 ProtoBlend (4.3) is below the simple K-center coreset (4.5), and on Kinetics-400 at VPC=5 it is below DM (8.0 vs 9.1). These are not competitive large-scale results, and the paper itself notes limited returns. The abstract and conclusion should qualify the claim to small/medium benchmarks (MiniUCF, HMDB51) or provide additional evidence that the efficiency advantage on these larger datasets offsets the accuracy gap. No efficiency measurements are reported for HMDB51, Kinetics-400, or SSv2, so the trade-off claim is only demonstrated on MiniUCF.","section":"Table 2; Abstract"}],"minor_comments":[{"comment":"The y-axis \"Distillation Time (h)\" in Fig. 3 is described in the text as post-teacher time. Please make this explicit in the figure or caption (e.g., \"Post-teacher construction time\") to avoid misleading readers who compare it directly with the optimization baselines.","section":"Fig. 1 and Fig. 3"},{"comment":"The caption or surrounding text should note that the strongest competitor on SSv2 VPC=5 is K-center, a simple coreset method, and that ProtoBlend's advantage over optimization-based methods is within noise on some large-scale settings. This would help calibrate the conclusion.","section":"Table 2"},{"comment":"The number of temporal candidates R=5 is fixed and not ablated. Since TGS is the largest contributor in Table 4, a sensitivity analysis on R would clarify how robust the selection stage is to this hyperparameter.","section":"Experimental Setup; Table 6"},{"comment":"The paper says \"Complete implementation and training hyperparameters are provided in the supplementary material,\" but no supplementary file is included with the arXiv submission. Please provide code or a detailed hyperparameter table for all datasets and baselines to enable reproduction.","section":"Implementation details"}],"recommendation":"major_revision","confidential_remarks":"The efficiency question is the hinge of this paper. The current measurements cover only the post-teacher phase, and the teacher itself is an expensive, dataset-specific component that the baselines do not use. I would like to see either full-pipeline cost numbers or a careful re-scoping of the headline claim. The paper is otherwise technically sound and transparent about its limitations, but the abstract and title currently overstate the evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a genuine contribution, but the efficiency claim is narrower than the title suggests. The select–allocate–blend pipeline is new: teacher-guided temporal clip selection, cluster-guided prototype allocation, and pixel-level blending with mixture-source soft labels. The ablations are clean and isolate each component's contribution, including the non-additivity of blending and soft supervision. On MiniUCF and HMDB51, ProtoBlend beats the optimization-based baselines at VPC=5 and VPC=10, and it transfers to recurrent students, which is a nice bonus. The authors are also honest about limitations: near-parity on Kinetics-400 and SSv2, ghosting in motion-centric actions, and dependence on teacher quality.\n\nThe soft spot is the efficiency claim. Fig. 1 and Fig. 3 report 'post-teacher construction time,' which excludes two real cost centers: fine-tuning a dataset-specific VideoMAE teacher on each full training set, and running R=5 forward passes on every source video for candidate scoring. The conclusion carefully says 'lower post-teacher construction cost,' but the title and abstract say 'efficient' without that qualifier. If teacher fine-tuning plus candidate inference costs more than the baseline optimization, the trade-off in Fig. 1 could reverse. This is not an internal inconsistency; the authors are accurate about what they measured. But it means the headline claim rests on an unmeasured quantity. I'd want to see a wall-clock or GPU-hour number for the full pipeline, or at least an explicit statement that the teacher is prepaid.\n\nOther concerns are minor. No code is available, and hyperparameters (alpha, lambda, tau, R) are tuned on MiniUCF, which is also the main evaluation benchmark; that's standard practice but worth flagging. The large-scale results are near-baseline, and the paper says so. The central accuracy claim holds up on the small benchmarks.\n\nWho should read it: anyone working on dataset distillation for video, and people interested in construction-based alternatives to synthetic optimization. It deserves a serious referee, but with a request to report end-to-end teacher cost and, ideally, release code.","headline":"Solid construction-based video distillation method with honest ablations, but the headline efficiency claim excludes teacher preparation and candidate inference, so read Fig. 1 as post-teacher construction time only.","tokens_in":14793,"tokens_out":1600,"would_cite":false,"duration_ms":19033,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Blending real clips distills video data without costly optimization.","keywords":["video dataset distillation","action recognition","prototype blending","teacher-guided selection","cluster-guided allocation","soft labels","construction-based distillation","VideoMAE"],"falsifier":"Replace the teacher-guided temporal clip selection (Eqs. 3–4) with uniform random clip selection while keeping clustering and blending unchanged, and measure distilled-set accuracy on MiniUCF. If accuracy does not fall to the level of the Random or Herding coreset baselines, the teacher's cross-entropy signal is not the load-bearing component the paper claims; if it does fall, that confirms the teacher's scoring is necessary.","tokens_in":13954,"feed_emoji":"🎬","tokens_out":4812,"duration_ms":45471,"temperature":0.7,"pith_summary":"ProtoBlend claims that video dataset distillation does not require iteratively optimizing synthetic videos. Instead, it constructs a compact surrogate set by selecting the most teacher-confident temporal clip from each source video, clustering those clips in teacher feature space, and blending in-cluster prototype–anchor pairs into single stored videos. Those stored videos are trained with mixture-source soft labels derived from the same blending coefficient, coupling input and supervision. On four action-recognition benchmarks, the paper reports accuracy comparable to or better than optimization-based baselines while cutting construction time to under an hour. A sympathetic reader should see this as evidence that useful video surrogates can be built by guided selection and interpolation rather than repeated gradient updates.","feed_headline":"Blending real clips distills video data without costly optimization","feed_subtitle":"Teacher-picked clips are clustered and blended into a compact training set in under an hour.","key_machinery":"The select–allocate–blend pipeline. Teacher-Guided Temporal Clip Selection retains the clip with the lowest target-class cross-entropy per source video; Cluster-Guided Prototype Allocation runs class-wise K-means on teacher embeddings and assigns one distilled slot per cluster, choosing the lowest-loss clip as prototype and second-lowest as anchor; Blended Sample Construction forms each distilled video as $\\tilde{x}_{c,k}=\\alpha x^{\\star}_{p_{c,k}}+(1-\\alpha)x^{\\star}_{a_{c,k}}$ and constructs its soft label by mixing the teacher posteriors of the same two clips with the same coefficient $\\alpha$. This shared coefficient is the mechanism that keeps the input blend and the supervision consist","core_discovery":"The central claim is that a frozen, dataset-specific VideoMAE teacher can carry the entire video distillation pipeline. Its target-class cross-entropy scores select the most discriminative temporal segment from each source video; its embeddings define intra-class modes that allocate the limited videos-per-class budget; and its prediction posteriors, mixed with the same coefficient used for pixel-level blending, supply soft labels aligned with the constructed inputs. Because the distilled set is built from forward inference, class-wise K-means, and direct interpolation, no gradient update is ever applied to the stored videos. The paper reports that this construction-based approach outperforms","pith_inferences":["Because construction cost weakly depends on the budget, the cached teacher outputs could be reused to generate multiple compression levels (VPC=1, 5, 10) at near-zero marginal cost, making ProtoBlend practical for rapid hyperparameter search and cross-architecture studies on surrogate sets.","The observed plateau on Something-Something V2 and Kinetics-400 suggests that frame-aligned pixel blending is the bottleneck for motion-centric classes; a motion-phase-aware warp or a class-adaptive allocation scheme could recover larger-budget gains where this paper sees diminishing returns.","The reliance on a dataset-specific teacher implies ProtoBlend inherits whatever biases the teacher has; testing with a single generalist video model as teacher could make the method dataset-agnostic and cheaper to deploy on new domains.","One could probe how sensitive the allocation stage is to the choice of feature space: replacing teacher embeddings with optical-flow or action-tag embeddings for K-means would reveal whether the teacher's appearance-biased representation is the right geometry for defining intra-class modes."],"forward_implications":["ProtoBlend's post-teacher construction cost is driven by a single teacher-inference pass; increasing VPC from 1 to 5 barely changes runtime, while optimization-based methods' cost grows with the number of distilled videos.","With VPC of 5 or more, cluster-guided allocation adds clear gains over selecting only the most confident clips, indicating that intra-class diversity coverage becomes more valuable than per-sample quality as the budget grows.","The accuracy gain transfers across student architectures: the same MiniUCF distilled sets rank first under ConvNet3D, CNN+GRU, and CNN+LSTM, suggesting the constructed videos retain general temporal cues rather than overfitting the evaluation model.","Blending and mixture-source soft labels are coupled: with hard labels blending hurts at VPC=5, but soft labels aligned to the blend recover and improve accuracy at both VPC=1 and VPC=5.","Random same-class pairing substantially degrades accuracy relative to in-cluster pairing, so the benefit of blending depends on the anchor being from the same teacher-feature cluster as the prototype."],"supporting_citations":[{"why":"Defines the static–dynamic disentanglement video distillation baseline and the MiniUCF evaluation protocol that ProtoBlend extends and compares against.","marker":"Wang et al. 2024b"},{"why":"Provides PRISM, the optimization-based video distillation method that ProtoBlend contrasts with and outperforms in several settings.","marker":"Choi et al. 2026"},{"why":"Supplies the VideoMAE teacher used for temporal clip scoring, feature extraction, and soft-label generation.","marker":"Tong et al. 2022"},{"why":"Provides the DM distribution-matching baseline that ProtoBlend compares against in accuracy and construction time.","marker":"Zhao and Bilen 2023"},{"why":"Provides the FRePo feature-regression baseline that ProtoBlend outperforms on MiniUCF at VPC=1 and VPC=5.","marker":"Zhou, Nezhadarya, and Ba 2022"},{"why":"Introduces the dataset distillation task that ProtoBlend addresses and establishes the surrogate-set training paradigm.","marker":"Wang et al. 2018"}],"fun_headline_variants":["No optimization needed: Blend teacher-picked clips to distill video data","Cluster-guided blending distills video data in under an hour","Teacher-picked clips blended into distilled video set without gradients","Video distillation without optimization: cluster and blend teacher-selected clips","From full videos to compact set by blending teacher picks, no optimization"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that a single frozen teacher's cross-entropy scores, feature embeddings, and prediction posteriors are all reliable enough to guide selection, clustering, and soft-label supervision; if teacher signal degrades, all three stages degrade together and the accuracy advantage disappears.","fun_headline_variants_meta":{"raw":{"variants":["No optimization needed: Blend teacher-picked clips to distill video data","Cluster-guided blending distills video data in under an hour","Teacher-picked clips blended into distilled video set without gradients","Video distillation without optimization: cluster and blend teacher-selected clips","From full videos to compact set by blending teacher picks, no optimization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000491,"raw_usage":{"total_tokens":2231,"prompt_tokens":703,"completion_tokens":1528,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":447,"completion_tokens_details":{"reasoning_tokens":1442}},"tokens_in":447,"tokens_out":1528,"duration_ms":11453,"temperature":1.0,"reasoning_tokens":1442,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:05:45.388199+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the teacher-guided temporal clip selection (Eqs. 3–4) with uniform random clip selection while keeping clustering and blending unchanged, and measure distilled-set accuracy on MiniUCF. If accuracy does not fall to the level of the Random or Herding coreset baselines, the teacher's cross-entropy signal is not the load-bearing component the paper claims; if it does fall, that confirms the teacher's scoring is necessary.","supporting_citations":[],"review_version":1}