{"id":"325893e8-ce06-46e2-b321-eac7064caa90","arxiv_id":"2508.12084","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A conditional diffusion model, DiffGEBD, generates diverse but plausible event boundary predictions for videos, with a new symmetric F1 and diversity score protocol for evaluating multi-prediction quality.","lead":"This paper applies diffusion models to generic event boundary detection, generating multiple plausible boundary sets for the same video instead of one deterministic answer. It also introduces a new evaluation metric combining diversity and fidelity, and reports strong results on Kinetics-GEBD and TAPOS.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Diversity-aware SOTA claim depends on comparing single-model diffusion samples against five-seed deterministic retrains; the baseline protocol may understate deterministic diversity, so the F1sym and F1g2p margins may not be a real advantage.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: the diversity-aware evaluation compares a single-model diffusion sampler against multi-seed retraining of deterministic baselines, conflating two different sources of variation. I agree with the reader's conditional verdict; the concern is addressable and does not by itself invalidate the generative formulation or the novel metric framework, but it does threaten the specific state-of-the-art claim on the new metrics. I would make the condition explicit: the authors should either (a) release code and rerun the diversity-aware comparison with a matched diversity-vs-fidelity calibration for baselines (e.g., threshold sweeps), or (b) report seed-variance error bars and confirm the margins in Table 1 exceed them. The paper has independent support in the conventional evaluation (Table 4: 78.4 on Kinetics-GEBD, 65.8 on TAPOS, competitive with DyBDet's 79.6 and 62.5) and in the ablation structure (Tables 2, 3, S1, S2, S3), which strengthens confidence that the method is technically sound and reproducible in its conventional regime. The flagged supplementary algorithm inconsistency (Alg. 1 calls h(y_0, E, t) on the clean label instead of the noisy y_crpt) is a genuine internal inconsistency that also needs correction and re-verification, but it is secondary to the evaluation-protocol concern for the headline claim. The test I propose is concrete and would settle whether the diversity-aware SOTA is an artifact: it changes only the baseline protocol, holding the DiffGEBD results and metrics fixed. If the margins collapse under the threshold-sweep baseline, the paper's central new-metric claim must be rephrased from 'state of the art' to 'competitive under a specific protocol'; if the margins survive, the concern is resolved and the conditional can be lifted.","tokens_in":19732,"tokens_out":3388,"duration_ms":27878,"concrete_test":"Recompute Table 1's F1sym, F1g2p, and diversity after replacing the five-seed deterministic baseline protocol with a within-model protocol: for each deterministic baseline, keep five boundary-prediction variants per video by sweeping the boundary threshold delta (e.g., 0.2, 0.3, 0.4, 0.5, 0.6) on a single trained model, mirroring DiffGEBD's single-model sampling with a controlled calibration knob. If DiffGEBD no longer holds the best F1sym or F1g2p, the claimed diversity-aware state of the art is a protocol artifact. Alternatively, keep the five-seed baseline but report the per-seed standard deviation of F1sym and diversity across the five seeds; if the best baseline's seed variance exceeds the DiffGEBD margin, the headline comparison is not established.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that DiffGEBD achieves state-of-the-art diversity-aware performance (Table 1: best F1sym, F1g2p, and diversity vs. reproduced deterministic baselines) depends on the reader accepting that five predictions from one DiffGEBD model are comparable to five models obtained by training each deterministic baseline with five random seeds. This is the weakest assumption because the two protocols sample from different distributions: seed variation of deterministic models spans optimization landscape variation and can produce near-duplicate or degenerate predictions, whereas diffusion sampling with CFG weight w=0.6 (Table S7: diversity 20.4 at w=0.6) draws from a deliberately regularized generative distribution. The F1g2p and F1sym gaps (72.9 vs. 71.6 for SC-Transformer; 74.0 vs. 72.9) could be artifacts of the baseline protocol rather than genuine coverage of human annotation variability. The supplementary Table S6 shows reproduced conventional F1@0.05 matching published numbers, which supports baseline reproduction quality, but it does not validate the five-seed diversity protocol. No uncertainty intervals, per-seed ranges, or distributional statistics are reported for the five-seed baselines, so we cannot tell whether the F1sym and diversity margins are larger than seed-induced variance. The paper's own ablation (Table S1) shows the non-diffusion model trained with five seeds underperforms DiffGEBD, which is consistent with the claim but does not isolate the protocol mismatch. Additionally, the annotation selection (Sec 5.1 uses at most 4 annotations selected by F1 consistency) means the GT distribution used for F1g2p evaluation is the same set used to filter training annotations, but the deterministic baselines were trained on the same filtered set; this asymmetry is not the main risk.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DiffGEBD, a diffusion-based model for generic event boundary detection (GEBD). The model uses a temporal self-similarity encoder to extract conditioning features and a denoising decoder that iteratively refines random noise into boundary predictions, with classifier-free guidance (CFG) to control the diversity of the generated boundary sets. The authors also introduce a diversity-aware evaluation protocol with two metrics: symmetric F1, which combines Pred-to-GT (F1p2g) and GT-to-Pred (F1g2p) alignment scores, and a diversity score measuring average pairwise dissimilarity among predictions. Experiments on Kinetics-GEBD and TAPOS show that DiffGEBD achieves the best F1sym, F1g2p, and diversity among compared methods on Kinetics-GEBD in the diversity-aware protocol, and the best F1@0.05 on TAPOS in the conventional protocol. The paper includes ablations on CFG weight, number of annotations, inference steps, conditioning features, and diffusion samplers, together with supplementary results including generalized energy distance and threshold robustness.","tokens_in":20005,"tokens_out":6519,"duration_ms":65656,"significance":"The generative formulation of GEBD is novel and addresses a real limitation of prior deterministic methods, and the proposed diversity-aware metrics (symmetric F1 and diversity score) fill a gap in how stochastic predictors are evaluated for tasks with subjective annotations. The paper also provides reproduced baselines, an efficiency comparison, and a broad set of ablations, which are useful for the community. If the evaluation protocol is accepted, the reported results would demonstrate a meaningful advance. However, the headline comparisons rest on an asymmetric protocol that may bias the diversity-aware claims, and the absence of uncertainty estimates weakens the support for state-of-the-art statements.","major_comments":[{"comment":"The diversity-aware comparison in Table 1 compares five predictions sampled from a single DiffGEBD model with five predictions obtained by training each deterministic baseline five times with different random seeds. These two protocols draw from different distributions: seed-varied training spans optimization uncertainty, which can be arbitrarily small if training is stable, whereas diffusion sampling with w=0.6 draws from a deliberately regularized generative distribution. Consequently, the claimed state-of-the-art margins in F1sym (74.0 vs 72.9) and diversity (20.4 vs 18.9) may be artifacts of the baseline protocol rather than evidence that DiffGEBD better covers the space of human annotations. The authors should justify this protocol explicitly, add an alternative way of obtaining multiple predictions from deterministic baselines (e.g., test-time dropout, varied thresholds, or multiple checkpoints), or temper the state-of-the-art claim.","section":"Sec. 5.4, Table 1"},{"comment":"No uncertainty intervals are reported anywhere in the paper: all tables give single F1 or diversity values, and the five-seed baselines are summarized only by means. Without per-seed ranges or confidence intervals, it is impossible to tell whether the reported margins (e.g., F1sym 74.0 vs 72.9, diversity 20.4 vs 18.9) exceed the seed-induced or sampling-induced variance. For a paper whose main contribution is an evaluation protocol, reporting the spread of the metrics (across the five predictions for DiffGEBD and across the five seeds for the baselines) should be a requirement.","section":"All results tables (Tables 1, 4, S6-S8)"},{"comment":"Algorithm 1 in the supplementary material computes a corrupted label y_crpt but then calls the decoder as h(y_0, E, t), i.e., it passes the clean ground-truth label instead of the noisy label y_t. This contradicts Sec. 4.2 and Eq. (9), which state that the decoder input is y_t. As written, the diffusion time-step conditioning is vacuous and the training procedure cannot be reproduced from the pseudo-code. The algorithm should be corrected to pass y_crpt to the decoder, or the main-text equations should be amended accordingly.","section":"Supplementary Alg. 1"},{"comment":"The forward process definition in Eq. (1), q(xt|xt-1) = N(xt; sqrt(alpha_t) x_{t-1}, alpha_t I), uses a variance of alpha_t I, which is inconsistent with the standard DDPM variance of (1 - alpha_t) I. In Eq. (3), the subscript 0 is used instead of t in the cumulative product coefficient, and the expression has misplaced parentheses around the mean term. These appear to be typographical errors, but because the forward process is foundational to the method, they should be corrected for clarity and precision.","section":"Sec. 3, Eqs. (1) and (3)"}],"minor_comments":[{"comment":"The sentence 'We employ the BasicGEBD-L4 encoder [60] and a 6-layer Transformer [47] for our encoder f and decoder h, .' contains a stray comma and incomplete clause; the sentence should end cleanly.","section":"Sec. 5.1"},{"comment":"Algorithm 2 uses 'linespace' instead of 'linspace' in the time-step scheduling line.","section":"Supplementary Alg. 2"},{"comment":"The entry 'cV AE' should be formatted consistently, e.g., as 'cVAE' or 'VQ-VAE', and the model should be named completely the first time it appears.","section":"Table S1"},{"comment":"The name 'EfficentGEBD' is misspelled in the discussion paragraph after Table 1; it should be 'EfficientGEBD'.","section":"Sec. 5.4"},{"comment":"Please clarify whether the selection of the top-NG annotations by F1 consistency score is performed per video or globally across the training set, and whether the evaluation at inference time uses all five ground-truth annotations.","section":"Sec. 5.1 and Fig. 5 caption"}],"recommendation":"major_revision","confidential_remarks":"The core idea is sound and the paper addresses a genuine gap in evaluating stochastic GEBD predictors, but the evaluation protocol fairness and the lack of uncertainty quantification are substantial concerns. The inconsistency in the training pseudo-code is a reproducibility red flag that must be fixed. I would support publication after these points are convincingly addressed and the claims are appropriately qualified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is the first diffusion-based generative formulation of GEBD, and the new evaluation protocol (symmetric F1 plus diversity score) is a sensible response to the multi-annotator nature of the task. Second, the headline diversity-aware SOTA claim is plausible but not nailed down: the comparison pairs one model's five samples against five seed-retrained deterministic baselines, and the paper gives no per-seed ranges or error bars, so the margins could be an artifact of protocol.\n\nWhat is genuinely new: framing boundary detection as conditional denoising over binary boundary maps, using classifier-free guidance to trade diversity against fidelity, and the F1sym/diversity metrics for many-to-many evaluation. The paper is honestly executed: they reproduce baselines from official code, publish full threshold sweeps, test different samplers, and ablate conditioning. The trend in CFG weight (diversity rises as w falls) is clear and the qualitative examples support the story.\n\nThe main soft spot is the baseline protocol in Table 1. Five predictions from a single diffusion model live in a smooth, deliberately regularized distribution, while five retrained deterministic models vary over optimization landscapes and can collapse to near-duplicate or degenerate outputs. Those are different sampling distributions. Without the per-seed numbers, we cannot tell whether an F1sym margin of 1.1 (74.0 vs 72.9) is bigger than seed-induced variance. It is worth saying this explicitly to the authors, because it is the only load-bearing weakness in the diversity-aware claim. Secondary issues: no code release, no error bars on any number, the CFG weight and threshold are tuned per evaluation setting, and the supplementary Algorithm 1 passes y_0 to h instead of the corrupted y_t (a typo, but it must be fixed).\n\nThe conventional GEBD results (78.4 Kinetics, 65.8 TAPOS) are competitive even under the old protocol, so the paper has value independent of the new metrics. Overall the central claim—that a conditional diffusion model can generate diverse and plausible boundaries—is supported. The paper is for video-event researchers and anyone designing evaluation for multi-annotator prediction. It deserves a proper peer review; I would send it out and ask for code, seed-level statistics, and a cleaner baseline protocol in revision.","headline":"A genuinely new generative angle on GEBD with a sensible two-metric evaluation, but the diversity-aware SOTA claim rests on a baseline protocol that may not be apples-to-apples.","tokens_in":20627,"tokens_out":2815,"would_cite":true,"duration_ms":27805,"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":"DiffGEBD claims that generic event boundary detection should be modeled as sampling from a distribution of plausible boundaries, and that one diffusion model can generate diverse, controllable boundary sets for the same video.","keywords":["generic event boundary detection","diffusion models","classifier-free guidance","temporal self-similarity","diversity-aware evaluation","symmetric F1","video segmentation"],"falsifier":"Train the strongest deterministic baseline from Table 1 five times with different random seeds, feed its five outputs into the paper's symmetric-F1 and diversity metrics, and check whether those outputs already reach DiffGEBD's reported diversity of 20.4 at F1sym 74.0; if they do, the claimed generative advantage in diversity is not established.","tokens_in":19512,"feed_emoji":"🎬","tokens_out":9377,"duration_ms":93771,"temperature":0.7,"pith_summary":"The paper tries to establish that generic event boundary detection—finding the class-agnostic frames where one meaningful event ends and another begins—is better treated as a generative sampling problem than as a deterministic prediction problem. Because event boundaries are subjective, a single video admits several plausible boundary sets, and the paper argues that a denoising diffusion model can learn that distribution and sample from it. The proposed DiffGEBD conditions a denoising decoder on temporal self-similarity features and uses classifier-free guidance so that one weight controls how diverse the sampled boundaries are. To evaluate such a model fairly, the paper introduces symmetric F1 and a diversity score that measure many-to-many alignment between a set of predictions and a set of human annotations, together with spread among the predictions. If the claim is right, video understanding systems could offer multiple human-like segmentations of the same video instead of committing to a single boundary map.","feed_headline":"Diffusion model lets one video yield many event boundaries","feed_subtitle":"A single DiffGEBD model samples plausible, controllable cuts; the diversity setting shifts with one guidance weight.","key_machinery":"The load-bearing mechanism is a conditional denoising decoder $h(y_t, t, E)$ that predicts clean boundary labels $y_0$ from noisy labels $y_t$, conditioned on temporal self-similarity embeddings $E$. The encoder builds $E$ through a sliding-window temporal self-similarity module, which highlights changes between adjacent frames; the decoder is a Transformer encoder layer with a scale-and-shift injection of the diffusion time step. At inference the decoder follows DDIM steps from random Gaussian noise $\\hat{y}_T$ to $\\hat{y}_0$, so different starting noises yield different candidate boundary sets. Diversity control comes from classifier-free guidance: the model is trained both with and without $E$, and inference combines the two predictions by $\\hat{y}_t = (1+w)\\hat{y}_t^c - w\\hat{y}_t^u$, so a small $w$ produces spread-out samples and a large $w$ produces near-deterministic ones. The paper's evaluation machinery is equally central: the symmetric F1 score is the harmonic mean of $\\mathrm{F1}_{p2g}$ and $\\mathrm{F1}_{g2p}$, measuring how well predictions cover annotations and annotations cover predictions, while the diversity score averages pairwise $1-\\mathrm{F1}$ over the generated set.","core_discovery":"The central claim is that the mapping from a video to its event boundaries is a distribution, not a single answer, and that this distribution can be learned and sampled by a conditional denoising diffusion model. From the video, a backbone extracts visual features; a temporal self-similarity encoder turns them into conditioning embeddings that highlight changes between adjacent frames; a Transformer-based decoder iteratively denoises random Gaussian noise into a binary boundary label sequence. Different initial noises produce different boundary sets from the same trained model, and classifier-free guidance with weight $w$ interpolates between diverse samples and strongly conditioned, faithful ones. The paper also claims that the standard F1 protocol is inadequate for such a generative model, and proposes symmetric F1—the harmonic mean of prediction-to-ground-truth F1 and ground-truth-to-prediction F1—plus a pairwise diversity score. On Kinetics-GEBD and TAPOS it reports the best or comparable fidelity among compared methods while producing measurably more diverse predictions.","pith_inferences":["A fairer reading of the Table 1 comparison would test whether seed-varied deterministic retraining is the right reference distribution for a generative model's within-model samples; the paper does not justify that equivalence.","The same recipe could transfer to other structured video outputs with subjective annotations, such as temporal action segmentation, if the conditioning carried class information and the diversity metric were made class-aware.","A practical extension not explored here is per-video guidance: a user could request one confident segmentation (high $w$) or many exploratory segmentations (low $w$) from the same checkpoint without retraining.","The annotation-reliability experiment suggests that filtering or reweighting human annotations by consistency during diffusion conditioning could improve fidelity further, since including all five annotators lowers F1sym even as diversity keeps rising."],"forward_implications":["A single DiffGEBD model can replace multiple training runs of a deterministic detector when multiple plausible boundary sets are needed: at CFG weight $0.6$ it reports F1sym $74.0$, F1g2p $72.9$, and diversity $20.4$ on Kinetics-GEBD, the best among the compared methods.","The same model can be switched to a high-fidelity mode by raising the guidance weight: at $w=4.0$ it reaches F1@0.05 of $78.4$ on Kinetics-GEBD and $65.8$ on TAPOS, the best reported on TAPOS.","Temporal self-similarity conditioning carries the boundary signal: replacing $E$ with raw visual features drops F1sym from $74.0$ to $68.5$.","Diffusion steps behave as an accuracy knob: F1sym and diversity both improve up to 32 DDIM steps and then plateau, so the generative gain does not require full 50-step sampling.","The new metrics offer a template for other tasks with multiple human annotations, where a set of predictions must be compared with a set of ground truths rather than with a single label."],"supporting_citations":[{"why":"defines the GEBD task and supplies the Kinetics-GEBD benchmark with multiple human annotations per video.","marker":"[37]"},{"why":"provides the temporal self-similarity encoder used as the diffusion condition and the deterministic baselines that the diversity-aware comparison must beat.","marker":"[60]"},{"why":"contributes classifier-free guidance, the mechanism that makes the diversity-versus-fidelity trade-off controllable by one weight.","marker":"[15]"},{"why":"supplies the denoising diffusion probabilistic model formulation that the decoder is trained with.","marker":"[16]"},{"why":"provides the DDIM sampling procedure used to turn random noise into boundary predictions with few steps.","marker":"[40]"},{"why":"gives a Transformer-based deterministic boundary detector used as a comparison in both the diversity-aware and conventional evaluations.","marker":"[24]"},{"why":"gives another deterministic boundary detector used as a comparison baseline.","marker":"[43]"},{"why":"is the source of the pairwise dissimilarity formula adopted for the diversity score.","marker":"[58]"}],"fun_headline_variants":["Diffusion samples multiple event boundaries per video","One video, many plausible cuts via diffusion","Diffusion turns GEBD into a generative sampling task","Controllable diversity for event boundary detection","Denoising diffusion redefines event boundary detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that five boundary sets sampled from one DiffGEBD model are the same kind of quantity as five boundary sets obtained by training a deterministic baseline five times with different random seeds, so that the reported diversity and symmetric-F1 comparison is fair.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion samples multiple event boundaries per video","One video, many plausible cuts via diffusion","Diffusion turns GEBD into a generative sampling task","Controllable diversity for event boundary detection","Denoising diffusion redefines event boundary detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00071,"raw_usage":{"total_tokens":3169,"prompt_tokens":890,"completion_tokens":2279,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":2209}},"tokens_in":506,"tokens_out":2279,"duration_ms":17304,"temperature":1.0,"reasoning_tokens":2209,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:25:01.717787+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the strongest deterministic baseline from Table 1 five times with different random seeds, feed its five outputs into the paper's symmetric-F1 and diversity metrics, and check whether those outputs already reach DiffGEBD's reported diversity of 20.4 at F1sym 74.0; if they do, the claimed generative advantage in diversity is not established.","supporting_citations":[{"cited_title":"Generic event boundary de- tection: A benchmark for event segmentation","cited_arxiv_id":null,"evidence_quote":"defines the GEBD task and supplies the Kinetics-GEBD benchmark with multiple human annotations per video."},{"cited_title":"Rethinking the architecture design for efficient generic event boundary detection","cited_arxiv_id":null,"evidence_quote":"provides the temporal self-similarity encoder used as the diffusion condition and the deterministic baselines that the diversity-aware comparison must beat."},{"cited_title":"Classifier-free diffusion guidance","cited_arxiv_id":null,"evidence_quote":"contributes classifier-free guidance, the mechanism that makes the diversity-versus-fidelity trade-off controllable by one weight."},{"cited_title":"Temporal perceiver: A general architecture for arbitrary boundary detection","cited_arxiv_id":null,"evidence_quote":"gives another deterministic boundary detector used as a comparison baseline."},{"cited_title":"A probabilistic model for controlling diversity and accuracy of ambiguous medical image segmentation","cited_arxiv_id":null,"evidence_quote":"is the source of the pairwise dissimilarity formula adopted for the diversity score."}],"review_version":1}