{"id":"54f97614-8c52-4549-8e15-dc95d7198b7c","arxiv_id":"2504.18866","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A dual-space (Euclidean and hyperbolic) framework with hyperbolic Dirichlet energy gating and LLM-generated ambiguous text supervision reports 88.82% AP on XD-Violence, a 2.48-point gain over prior hyperbolic methods.","lead":"Dual-space video violence detection combining Euclidean and hyperbolic feature spaces with LLM-generated ambiguous text supervision achieves state-of-the-art scores on the XD-Violence benchmark. Readers outside the niche may care because the method targets a practical failure mode: distinguishing visually similar but semantically different events, such as a sports collision versus an assault.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Core fusion metric is undefined as written: Eq. 18 and Eq. 24 compute Lorentzian distance between a Euclidean query and a hyperbolic key without projecting the Euclidean input onto the hyperboloid, so the reported cross-space gains rest on unspecified implementation.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the Lorentzian similarity in Eq. 18 and in HVLGL is applied to a Euclidean-space feature that has not been projected onto the hyperboloid, even though Eq. 6 only defines the Lorentzian distance for points on L^n_K. I agree this is the central soft spot. The paper's headline claim is that dual-space modeling plus ambiguous text supervision improves violence detection; the DSI and HVLGL modules are exactly the new mechanisms that realize this claim, and both rely on an undefined metric operation. The issue is fixable, because the authors may have an implicit projection in their code, but the paper alone is not reproducible on this point. I do not move to reject: the architecture may work in practice, the authors disclose that DSRL is prior work, and the remaining concerns (no error bars, grid-searched hyperparameters, a 13-video self-curated subset) are secondary but reinforce the same conditional verdict. The right outcome is to require a corrected derivation or code release before the SOTA claim is accepted.","tokens_in":25741,"tokens_out":4644,"duration_ms":51175,"concrete_test":"Ask the authors to release code or pseudocode for the exact computation used in Eq. 18 and Eq. 24. Then re-run the Table 6 DSI ablation and Table 7 HVLGL ablation under two implementations: (i) V_q mapped to the hyperboloid via exp_o([0,V_q]) before computing Ls; (ii) an explicit tangent-space Euclidean distance. If both variants reproduce the reported 88.82/83.74 AP within 0.2 AP, the ambiguity is benign; if not, the Lorentzian-attention advantage depends on an unstated projection choice and the central claim is not supported as written.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. 6 defines d^K_L(x,y) only for points x,y in L^n_K, with d = arcosh(K <x,y>_L). In Section 4.2.2, DSI linearly transforms V_E into the query V_q and V_H into the key V_k, then sets A^{E->H} = softmax(Ls(V_q,V_k)) in Eq. 18. No exponential map, logarithmic map, or any projection of V_q onto L^n_K is specified; this is in contrast to Eq. 7 and Eq. 23, where the mappings into hyperbolic space are explicit. The same problem appears in HVLGL: Eq. 24 computes Ls(V,T^{+,L}) with the visual feature V taken from the Euclidean stream while only the text feature is exponentiated in Eq. 23. As written, the argument to arcosh need not lie in the domain of the Lorentzian distance, so the attention weights are not well-defined. This is not a cosmetic notation issue: Table 6 attributes 2.03/2.01 AP improvements to using the Lorentzian metric instead of cosine similarity in DSI, and HVLGL adds 0.83/1.16 AP over the hyperbolic loss. If the actual implementation silently projects V_q through the tangent space or uses a different distance, then those ablations are not testing the stated metric, and the 'dual-space interaction' mechanism is underspecified. Because DSI and HVLGL are precisely the components that distinguish PiercingEye from DSRL, the central claim of state-of-the-art dual-space improvement depends on this operation being defined and reproducible.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents PiercingEye, a weakly supervised video violence detection framework that combines Euclidean and hyperbolic representations. The method consists of four modules: a hyperbolic energy-constrained GCN (HE-GCN) with layer-sensitive hyperbolic association degrees, a dual-space interaction (DSI) module using Lorentzian cross-space attention, an ambiguous event text generation (AETG) module that uses a VLM and LLM to produce scene- or action-edited text descriptions, and a hyperbolic vision-language guided loss (HVLGL) with text-similarity-based weighting of hard negatives. Experiments on XD-Violence report state-of-the-art AP of 88.82% (multimodal) and 83.74% (unimodal), and on a self-curated 13-video ambiguous subset of UCF-Crime the method reports 83.21% AUC versus 82.34% for DSRL. The ablation studies isolate each component, and the paper explicitly discloses that the HE-GCN and DSI modules originate from the authors' prior DSRL work.","tokens_in":26078,"tokens_out":8293,"duration_ms":78720,"significance":"If the technical details are made precise and the evaluation protocol strengthened, the paper's central idea—jointly modeling visual features in Euclidean space and event hierarchies in hyperbolic space, with LLM-generated ambiguous text as auxiliary supervision—is a meaningful contribution to weakly supervised video violence detection. The paper is honest about the provenance of the core dual-space modules and provides a thorough set of ablations, including comparisons of distance metrics, message-passing strategies, and contrastive losses. It also reports model complexity and speed. However, the current manuscript leaves a mathematically central operation undefined: the Lorentzian similarity in the cross-space attention is applied to a Euclidean query without a specified projection onto the hyperboloid. This, together with test-set hyperparameter selection and a very small self-curated benchmark, means the state-of-the-art claim is not yet fully substantiated.","major_comments":[{"comment":"Lorentzian similarity requires both arguments to lie on the hyperboloid L^n_K, but in Eq. (18) the query V_q is obtained by a linear layer on the Euclidean features V_E while only the key V_k is derived from hyperbolic features; no exponential or logarithmic map is applied to V_q. Similarly, in Eq. (24) the visual feature V is used directly in Ls(V, T^{+,L}) while only the text features are exponentiated in Eq. (23). Because Table 6 attributes 2.03/2.01 AP improvements to replacing cosine similarity with the Lorentzian metric in DSI, this is not a notation issue: the ablations are testing an operation that is undefined as written. The same concern applies to Eq. (28), where VF is a concatenation of Euclidean and hyperbolic branch outputs but is fed into a Lorentzian inner product. Please specify the projection of Euclidean features onto the hyperboloid for DSI, HVLGL, and the classifier, and confirm that the implementation used for Tables 6 and 7 matches the corrected definition.","section":"Sec. 4.2.2, Eq. (18); Sec. 4.4, Eq. (24)"},{"comment":"The simplification of the hyperbolic Dirichlet energy is mathematically incorrect. With d_i = d_j = n-1, the arguments of the distance in Eq. (10) become exp_o(c log_o x_i) with c = 1/sqrt(n), which are not equal to x_i; because exp_o is nonlinear, the distance between these scaled geodesic points is not d^K_L(x_i, x_j). The simplified form in Eq. (11) therefore does not follow. Since HDE directly determines the LSHAD threshold in Eq. (12), the layer-sensitive node selection claimed in the HE-GCN ablation (Table 5) relies on an ill-defined quantity. Please provide a correct derivation or state the actual HDE variant used in the experiments.","section":"Sec. 4.2.1, Eqs. (10)-(11)"},{"comment":"All hyperparameter ablations (β, γ in Fig. 6; λ, α in Fig. 7; τ, θ in Table 8; ψ in Fig. 8) are performed on the XD-Violence test set, and no validation split is described in Section 5.1. Selecting hyperparameters on the test set and then reporting the selected configuration's test performance overstates generalization and makes the SOTA numbers difficult to interpret. Please either introduce a validation split for model selection or report results across multiple seeds with mean and standard deviation so the reader can assess variance. At minimum, state explicitly how the final hyperparameter values were chosen.","section":"Sec. 5.3.2-5.3.4"},{"comment":"The curated ambiguous-event subset of UCF-Crime contains only 13 videos (32,154 frames) and is constructed by first selecting videos that PEL4VAD scores poorly, then asking annotators to confirm ambiguity. On this small, selection-biased set, the reported advantage over DSRL is 0.87 AUC points (83.21 vs. 82.34). As reported, this difference is not shown to be statistically reliable, and the selection procedure is not described in enough detail to rule out bias toward any particular method family. Please provide per-video results, a confidence interval or significance test, and a more transparent selection protocol, or treat the result as a qualitative illustration rather than a benchmark claim.","section":"Sec. 5.2, Table 3"}],"minor_comments":[{"comment":"There are several typos: 'pratice' in Section 5.1, 'representated' in Section 4.3.2, 'Lorenzt' in Section 4.2.1, 'mutilmodal seetings' and 'PirecingEye' in the Fig. 6 caption, 'Mutilmodal' in the Fig. 8 caption, and 'T exts' in the Section 5.4.3 heading.","section":"Throughout"},{"comment":"The LSHAD definition writes f(E_K^H(x), k) but the right-hand side is explicit; please clarify the role of f and the admissible ranges of β and γ so that the threshold is always in the intended interval.","section":"Sec. 4.2.1, Eq. (12)"},{"comment":"The notation CSAE→H(VH,VE) in Eq. (19) appears inconsistent with the text describing a transfer 'from VE to VH'; please align the argument order and the textual description.","section":"Sec. 4.2.2, Eq. (19)"},{"comment":"Figure 2 is very crowded, and the text labels for the loss module are hard to read; consider enlarging the figure or splitting it into two panels.","section":"Fig. 2"},{"comment":"Reference [34] is cited as 'NeurIPS24' in Tables 1-3 but appears as an arXiv preprint in the bibliography; please update the citation and the reference entry to the published version.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is a journal extension of the authors' own prior DSRL work, and the new material (AETG and HVLGL) is incremental but potentially acceptable if the technical gaps are closed. I am particularly concerned that the undefined Lorentzian operation and the test-set hyperparameter tuning will not survive a careful reproduction, and the 13-video subset is too weak to support the 'ambiguous events' headline. The paper is otherwise clearly written and well organized."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honest headline: this is a real but incremental extension of the authors' own DSRL work, and the XD-Violence numbers look plausible. The thing to check before buying the story is Eq. 18: the cross-space attention computes a Lorentzian similarity between a Euclidean query and a hyperbolic key without any projection into the hyperboloid, so the core fusion mechanism is not reproducible as written.\n\nWhat's actually new: AETG and HVLGL. The HE-GCN and DSI backbone is explicitly disclosed as prior work [34], and that is the right kind of honesty. The new pieces are well motivated—generate texts that are visually similar but semantically different by swapping scene or action, then use a hyperbolic vision-language contrastive loss with similarity-based weighting on negatives. The ablations are consistent: each component adds a bit of AP, and the gains over DSRL are modest, which matches the claim of incremental improvement. The 13-video ambiguous subset is clearly a sanity check, not a benchmark, and it is labeled as such.\n\nSoft spots, in order of severity. First, the undefined metric in Eq. 18 and Eq. 24. Lorentzian distance in Eq. 6 is defined for two points on the hyperboloid. In DSI, V_q comes from the Euclidean stream and V_k from the hyperbolic stream; in HVLGL, the visual feature V is Euclidean while only the text is exponentiated. No exponential map or projection is specified for the Euclidean input. This is not a notation nitpick: two ablations credit the Lorentzian metric for about 2 AP points in DSI and 1 point in HVLGL, so the gains rest on an operation that, as written, does not exist. It is probably fixable by adding the missing mapping, but the paper is not reproducible without it.\n\nSecond, Eq. 11. The simplified hyperbolic Dirichlet energy drops the degree normalization from Eq. 10 and treats the scaled tangent vectors as the original points. That simplification is not valid as stated. Minor, but it should be corrected.\n\nThird, experimental reporting. Single runs, no error bars, and a lot of grid-searched hyperparameters (beta, gamma, lambda, alpha, tau, theta, psi) without a described validation split. The robustness curves help, but they don't show the chosen values weren't selected on the test set. No code or prompts are released, which matters for AETG because the generated texts depend on exact prompts.\n\nBottom line: this paper deserves a serious referee. I would send it out, with the expectation of a major revision to define the metric properly and add multi-seed results. If those land, the empirical claims could hold. It is honest, clearly written, and the engineering is clean enough to be worth the referees' time.","headline":"Real but incremental dual-space violence detection; the headline numbers look plausible, but the core cross-space Lorentzian attention is undefined as written.","tokens_in":26667,"tokens_out":3634,"would_cite":false,"duration_ms":37999,"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":"The paper claims that dual-space representation with LLM-generated ambiguous text supervision reaches state-of-the-art violence detection on XD-Violence and a curated ambiguous subset of UCF-Crime.","keywords":["video violence detection","weakly supervised learning","hyperbolic representation learning","dual-space learning","ambiguous event detection","vision-language alignment","large language models","Lorentzian metric"],"falsifier":"Run the same training and evaluation with the cross-space attention changed so that the Euclidean query is first projected onto the hyperboloid with the paper's own exponential map before the Lorentzian similarity is computed; if the XD-Violence AP stays at the reported 88.82% multimodal and 83.74% unimodal and the gap over cosine similarity from Table 6 disappears, then the claimed Lorentzian advantage is not what resolves ambiguity.","tokens_in":25509,"feed_emoji":"👁","tokens_out":12967,"duration_ms":113878,"temperature":0.7,"pith_summary":"PiercingEye tries to show that weakly supervised video violence detection fails mainly on ambiguous events—visually similar moments that mean different things, like a hockey fight versus legal body checking—and that this failure can be fixed by learning in two geometries at once. It combines a Euclidean stream that preserves visual appearance and motion with a hyperbolic stream that captures the hierarchy of event categories and temporal context, then fuses the two with Lorentzian cross-space attention. To supply scarce supervision for ambiguity, it uses a vision-language model and a large language model to generate text descriptions that change the scene or the action while staying visually plausible, and trains with a hyperbolic contrastive loss that weights the hardest confusing negatives most heavily. The paper reports state-of-the-art frame-level average precision on XD-Violence (88.82% multimodal, 83.74% unimodal), a comparable 86.64% AUC on UCF-Crime, and a gain over the authors' earlier dual-space model on a 13-video ambiguous subset. If true, this amounts to a recipe for detecting violence where appearance alone cannot decide.","feed_headline":"Dual-space model tops violence detection at 88.8% AP","feed_subtitle":"Euclidean and hyperbolic features plus generated ambiguous text separate fights from look-alike sports plays.","key_machinery":"The load-bearing object is the Lorentz model of hyperbolic space, a constant-negative-curvature manifold on which hierarchical event structure can be encoded in distances. Three mechanisms use it. HE-GCN maps snippet features onto the hyperboloid, computes Lorentzian similarities $L_s(x_i,x_j)=\\exp(-d^K_L(x_i,x_j))$ between nodes, and prunes the message-passing graph with a layer-sensitive threshold built from the hyperbolic Dirichlet energy. DSI treats Euclidean and hyperbolic features as query/key/value sets and applies the same Lorentzian similarity in cross-space attention, so appearance and motion from Euclidean space can interact with event hierarchy from hyperbolic space. AETG modifies VLM-generated captions by changing the scene or the action to create visually similar but semantically distinct text, and HVLGL projects those texts onto the hyperboloid and applies a contrastive loss with negative weights $w_j=\\exp(-\\theta L_s(T^+,T^-_j))$ that emphasize confusing cases. The ablations tie the reported gains to these components: removing HE-GCN, DSI, or HVLGL lowers AP, and replacing the Lorentzian metric with cosine similarity costs about 2% AP.","core_discovery":"The central claim is that Euclidean and hyperbolic representations are complementary for event understanding, and that language-generated ambiguity can be turned into a trainable signal. In concrete terms, PiercingEye asserts that modeling events on a hyperboloid with a layer-sensitive graph, exchanging information across spaces through Lorentzian attention, and aligning video features to LLM-generated ambiguous text descriptions improves fine-grained discrimination. The reported numbers are 88.82% AP on XD-Violence with audio-visual input and 83.74% with video-only input, both above the listed previous methods; 86.64% AUC on UCF-Crime; and 83.21% AUC versus 82.34% for DSRL on a manually curated subset of 13 visually ambiguous UCF-Crime videos. The paper also argues that the hyperbolic text guidance is what pushes the ambiguous-event performance up: removing it drops multimodal AP by 1.21% and unimodal AP by 1.73% in the ablations.","pith_inferences":["A natural extension the paper does not test: the AETG scene/action perturbation recipe could generate hard negatives for other fine-grained video tasks such as action recognition or anomaly explanation, effectively using LLM text as data augmentation for semantic ambiguity.","Because the hyperbolic contrastive loss outperforms a Euclidean InfoNCE-style loss even before the text-weighting is added, one testable prediction is that the same dual-space plus hyperbolic-text recipe will help open-vocabulary anomaly detection, where the text vocabulary at test time is unknown.","The ambiguous-event subset contains only 13 videos, so the 0.87-point AUC gap over DSRL, while consistent with the method's motivation, should be checked on a larger curated set before treating it as a robust property; the authors' own future-work section acknowledges the scarcity of ambiguous samples.","One could also combine the dynamic text-similarity weighting with a curriculum schedule in Euclidean CLIP-style losses; if the gain comes from weighting rather than hyperbolic geometry, the same weighting should reproduce part of the HVLGL improvement."],"forward_implications":["If the reported XD-Violence numbers hold, audio-visual violence detection can be improved without bigger datasets or extra inference parameters, because the text generation and the HVLGL loss are used only during training.","If the ambiguous-subset result holds, LLM/VLM-generated hard-negative text can stand in for scarce real ambiguous footage in weakly supervised video tasks, which matters for other safety-critical recognition problems.","If the geometry claim is right, Euclidean-only or hyperbolic-only feature learning is a real bottleneck for fine-grained event recognition, and dual-space fusion is a generally applicable design rather than a violence-specific trick.","A direct corollary of the ablations is that the Lorentzian metric itself, not just the extra text data, carries much of the gain: replacing it with cosine similarity in DSI drops AP by about 2% in both unimodal and multimodal settings."],"supporting_citations":[{"why":"Supplies the XD-Violence dataset, the audio-visual weak-supervision protocol, and the I3D/VGGish features used for evaluation.","marker":"[2]"},{"why":"The authors' earlier dual-space model (DSRL) that PiercingEye extends and must beat; provides the HE-GCN/DSI backbone and the main comparison numbers.","marker":"[34]"},{"why":"HyperVD, the hyperbolic-space VVD baseline whose hard node selection and implementation details HE-GCN improves on.","marker":"[28]"},{"why":"CFA-HLGAtt, the strongest hyperbolic-graph VVD baseline (86.34% AP) that PiercingEye compares against.","marker":"[29]"},{"why":"Supplies the UCF-Crime dataset, its video-level training protocol, and the MIL baseline for anomaly detection.","marker":"[1]"},{"why":"CLIP, the pretrained text encoder used in HVLGL and the visual encoder used for UCF-Crime features.","marker":"[54]"},{"why":"InternVL2, the vision-language model used in AETG to generate initial scene-and-action captions.","marker":"[33]"},{"why":"Provides the Lorentz linear layer formalism (Eqs. 8-9) used for hyperbolic feature transformation.","marker":"[47]"},{"why":"Defines the Lorentz model and the distance metric that the paper's hyperbolic modules are built on.","marker":"[63]"},{"why":"The large language model used in AETG to produce the ambiguous scene- and action-modified text descriptions.","marker":"[31]"}],"fun_headline_variants":["Dual-space, LLM-guided framework tops violence detection benchmarks","Hyperbolic+Euclidean features plus text ambiguity sharpen violence spotting","PiercingEye: language-driven ambiguity training boosts violent event classification","Euclidean and hyperbolic spaces combine to beat prior violence detectors","Ambiguous text from LLMs helps dual-space model ace violence detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that a Lorentzian distance, which is only defined for points on the hyperboloid, is applied to an ordinary Euclidean-space vector and a hyperbolic-space vector without the paper specifying how the Euclidean vector is moved onto the hyperboloid first, so the fusion mechanism's correctness depends on an operation that is not defined in the text.","fun_headline_variants_meta":{"raw":{"variants":["Dual-space, LLM-guided framework tops violence detection benchmarks","Hyperbolic+Euclidean features plus text ambiguity sharpen violence spotting","PiercingEye: language-driven ambiguity training boosts violent event classification","Euclidean and hyperbolic spaces combine to beat prior violence detectors","Ambiguous text from LLMs helps dual-space model ace violence detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000441,"raw_usage":{"total_tokens":2226,"prompt_tokens":924,"completion_tokens":1302,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":1214}},"tokens_in":540,"tokens_out":1302,"duration_ms":12438,"temperature":1.0,"reasoning_tokens":1214,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:08:09.104011+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same training and evaluation with the cross-space attention changed so that the Euclidean query is first projected onto the hyperboloid with the paper's own exponential map before the Lorentzian similarity is computed; if the XD-Violence AP stays at the reported 88.82% multimodal and 83.74% unimodal and the gap over cosine similarity from Table 6 disappears, then the claimed Lorentzian advantage is not what resolves ambiguity.","supporting_citations":[{"cited_title":"Not only look, but also listen: Learn- ing multimodal violence detection under weak supervision","cited_arxiv_id":null,"evidence_quote":"Supplies the XD-Violence dataset, the audio-visual weak-supervision protocol, and the I3D/VGGish features used for evaluation."},{"cited_title":"Cross-modal fusion and attention mechanism for weakly supervised video anomaly detection","cited_arxiv_id":null,"evidence_quote":"CFA-HLGAtt, the strongest hyperbolic-graph VVD baseline (86.34% AP) that PiercingEye compares against."},{"cited_title":"Real-world anomaly detection in surveillance videos","cited_arxiv_id":null,"evidence_quote":"Supplies the UCF-Crime dataset, its video-level training protocol, and the MIL baseline for anomaly detection."},{"cited_title":"Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks","cited_arxiv_id":null,"evidence_quote":"InternVL2, the vision-language model used in AETG to generate initial scene-and-action captions."},{"cited_title":"Learning continuous hierar- chies in the lorentz model of hyperbolic geometry","cited_arxiv_id":null,"evidence_quote":"Defines the Lorentz model and the distance metric that the paper's hyperbolic modules are built on."}],"review_version":1}