{"id":"5dbb2f79-66a0-4ef8-82ae-86c5201696b5","arxiv_id":"2501.02593","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Taylor-transformed skeletons improve ST-GCN accuracy but reduce Hyperformer accuracy on NTU-60/120, indicating that motion-injected inputs do not universally benefit skeleton-based action recognition models.","lead":"This paper compares two skeleton-based action recognition models, ST-GCN and Hyperformer, using original skeletons and Taylor-transformed skeletons that encode motion dynamics. On NTU-60 and NTU-120, the motion-enhanced representation improves ST-GCN but slightly degrades Hyperformer, with per-class analysis showing gains on dynamic actions and losses on spatial ones.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported Hyperformer degradation may be an artifact of the displacement-only Taylor input and fixed training setup; retaining static pose or tuning hyperparameters could erase the harm, so the architecture-dependent conclusion is not yet established.","rationale":"The reader's weakest assumption was that the fixed Taylor configuration is an unfair test for both models. I agree and sharpen the point: the specific operationalization used here—displacement-only, single-term Taylor input—removes static joint positions that Hyperformer's hyperedge attention is designed to consume. The paper's central claim is therefore underdetermined: it may be true of this one encoding and this one training recipe, but not of Taylor-transformed skeletons as a representation. The proposed test directly differentiates those readings. The verdict stays CONDITIONAL because the empirical pattern is consistent and the paper reports per-class confusion matrices, which partially support the descriptive claim; however, the interaction claim needs the pose-preserving variant or an adapted hyperparameter search before it can be accepted as stated.","tokens_in":86739,"tokens_out":7106,"duration_ms":76010,"concrete_test":"One decisive experiment: train Hyperformer on Taylor-transformed skeletons in two variants—(A) the exact displacement-only input used in the paper, and (B) a pose-preserving input that concatenates the static joint coordinates with the first-order displacement (zeroth+first Taylor term)—using otherwise identical hyperparameters and three learning rates (0.01, 0.025, 0.05). If the best variant B reaches or exceeds the original-skeleton baseline (90.7 on NTU-60 X-Sub, 86.6 on NTU-120 X-Sub) while variant A still lags, the reported harm is an artifact of discarding spatial information. If all variants remain more than 1 point below baseline, the architecture-dependence claim survives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is an interaction claim: Taylor-transformed skeletons help ST-GCN but harm Hyperformer. Establishing it requires the comparison to isolate the input representation. Section 4.1 fixes the Taylor transform to 'the displacement concept with a single term, four frames per temporal block and a step size of one' with 'no hyperparameter search,' and applies each model's standard training recipe unchanged (ST-GCN SGD lr 0.01; Hyperformer 140 epochs, lr 0.025 decaying at epochs 110/120). A displacement-only single-term Taylor input removes static joint coordinates. Hyperformer's hypergraph self-attention is explicitly built on joint-distance and joint-group attention (Section 3.3, Figure 2), so distances computed on displacement features are not the spatial distances the model was designed to consume. The 4-7 point drops for Hyperformer on all four benchmarks could therefore reflect incompatibility between the input representation and the model's fixed attention mechanism, or a poor training match, rather than an intrinsic limitation of motion-enhanced skeletons for hypergraph models. The paper's own explanation (Section 4.2: 'Taylor skeletons... lack detailed spatial information... which the Hyperformer may still rely on') concedes this confound. Under this reading, the paper demonstrates that one particular Taylor encoding, with one training pipeline, hurts one model; it does not establish the architecture-dependent interaction claim as stated. No error bars or multi-seed runs are reported, so the small ST-GCN gains (+1.1 to +3.4) are also not statistically supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper reports a comparative empirical study of two skeleton-based action recognition models, ST-GCN and Hyperformer, on the NTU-60 and NTU-120 benchmarks. For each model, the authors compare the original skeleton sequences against 'Taylor-transformed' skeletons, an input encoding borrowed from the prior Taylor Videos work (ref [95]) that is intended to emphasize motion dynamics. The headline finding is that Taylor-transformed skeletons improve ST-GCN accuracy on all four evaluated benchmarks (e.g., NTU-60 X-Sub 81.5% to 83.5%) but decrease Hyperformer accuracy (e.g., NTU-60 X-Sub 90.7% to 86.7%). The paper also provides per-class confusion matrices and tables of the action classes with the largest gains and losses under the transformation, and it discusses the trade-off between motion sensitivity and spatial detail.","tokens_in":87044,"tokens_out":4898,"duration_ms":52223,"significance":"If the reported interaction is reliable, the paper provides a useful empirical data point: the benefit of motion-injected skeleton representations is not uniform across architectures, and models with spatial-distance-based attention may be harmed by displacement-only inputs. The study's strengths are its breadth of benchmarks (four standard NTU splits), the detailed per-class analysis in Tables 2-9, and the honest disclosure that no hyperparameter search or denoising was performed. However, the paper's central contribution is purely empirical, and its credibility depends on the stability and confound-control of the reported accuracy numbers, both of which are currently insufficient for a journal-level claim about architecture-dependent input representations.","major_comments":[{"comment":"All accuracy values are from single runs with no error bars, multiple seeds, or significance tests. The central interaction claim (Taylor skeletons help ST-GCN and harm Hyperformer) rests on differences ranging from 1.1 to 7.5 percentage points across the four NTU benchmarks. In skeleton action recognition, run-to-run variation of this magnitude is common, so the reader cannot tell whether the observed interaction is a stable property or a seed artifact. For an evaluation-only paper, the authors should report mean and standard deviation over at least three seeds, or apply a paired statistical test over the test set (e.g., bootstrap per-sample accuracies) to support the claim.","section":"Section 4.1, Table 1"},{"comment":"The Taylor encoding is fixed to 'the displacement concept with a single term, four frames per temporal block and a step size of one' with 'no hyperparameter search,' and each model uses its own standard training recipe. This confounds the input representation with the model's ability to consume that representation. Hyperformer's attention mechanism explicitly includes joint-distance attention computed on input coordinates (Section 3.3, Figure 2); a displacement-only representation removes the static spatial structure that this attention was designed to process. The 4-7 point drops for Hyperformer may therefore reflect an encoding/model mismatch rather than a general incompatibility between motion-enhanced skeletons and hypergraph-transformer models. The paper's own explanation in Section 4.2 ('they lack detailed spatial information... which the Hyperformer may still rely on') concedes this confound. To support the architecture-dependent conclusion, the authors should test at least one Taylor variant that retains static pose information, or tune the Taylor hyperparameters for Hyperformer, or explicitly re-frame the claim as being about this particular encoding and pipeline.","section":"Section 4.1, Models"},{"comment":"The description of the Taylor transform is internally inconsistent, and the exact input representation used in the experiments is not fully specified. Section 3.2 describes Taylor-transformed skeletons as combining zeroth-, first-, and second-order temporal derivatives, while Section 4.1 states that the experiments use 'the displacement concept with a single term' without defining that concept. The reader cannot determine whether the input fed to Hyperformer contains static joint coordinates, only velocities, or some combination. Since the paper's central interpretation hinges on the loss of spatial information in the Taylor input, the exact formula (and whether static positions are present) must be stated unambiguously, ideally in Section 4.1 or in a short appendix.","section":"Sections 3.2 and 4.1"}],"minor_comments":[{"comment":"Top-5 accuracy is reported only for the Taylor-transformed runs; the original-skeleton rows show '–' in the Top-5 columns. This prevents the reader from comparing top-5 performance between the two input conditions, which would be informative given the large top-1 differences for Hyperformer.","section":"Table 1"},{"comment":"The paper frames the comparison as 'skeletal graphs versus hypergraphs,' but only one graph model (ST-GCN) and one hypergraph model (Hyperformer) are evaluated. The conclusions about representation families are model-specific and should be explicitly qualified as such, both in the contribution list and in the conclusion.","section":"Contributions and Section 3.1"},{"comment":"The captions state that predictions below 5% are filtered out for clarity, but several rows of the displayed matrices do not sum to approximately 100% even after accounting for this filtering (e.g., Figure 4(a), 'drink water' row). The paper should clarify whether the remaining discrepancy is due to rounding, filtering of additional values, or entries not shown, so that readers can trust the numerical values in the figures.","section":"Figures 3 and 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable empirical study for a workshop venue, but the journal-level contribution is not yet established due to the single-run results and the confounded comparison described in the major comments. The reference list contains many entries that seem only tangentially related to the cited context (e.g., refs [19] and [20] appear under motion-centric approaches); this may warrant an editorial check. If the authors add seed variance or controlled Taylor variants, the paper could become a solid evaluation contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the cross-evaluation: the same two model families, ST-GCN and Hyperformer, tested with and without Taylor-transformed skeletons on NTU-60 and NTU-120. The differential outcome—Taylor skeletons help ST-GCN by a few points and hurt Hyperformer by several—is non-obvious and worth knowing about if you work in skeleton-based action recognition. The paper also does something right that many evaluation papers skip: it states plainly that no hyperparameter search and no denoising were performed, and it releases per-class confusion matrices that let you see exactly which actions move. That is real evidence, even if not formally verified or code-backed.\n\nThe soft spots are real and, in one place, load-bearing. There are no error bars, no multi-seed runs, and no significance tests anywhere. The ST-GCN gains range from about +1.1 to +3.4 accuracy points; at single-run precision, those could easily be noise. The Hyperformer drop is larger, but the stress-test note lands: the Taylor configuration used is displacement-only, a single term, four frames per block, step size one, inherited verbatim from the authors' own Taylor Videos paper. Hyperformer's attention is explicitly built around joint-distance and joint-group relations computed on spatial coordinates, so feeding it displacement features removes exactly the spatial distances it was designed to consume. The paper's own explanation in Section 4.2 concedes this: Taylor skeletons lack detailed spatial information that Hyperformer may rely on. That is not a hidden flaw; it is an acknowledged confound, but it means the headline claim—that motion-enhanced skeletons intrinsically help graphs and harm hypergraphs—is not established. What is established is that this particular Taylor encoding, with this training pipeline, hurts this particular hypergraph model.\n\nOn balance, this is an honest, clearly written empirical note. The citation pattern is fine: the Taylor transformation comes from the second author's prior work, but the paper reports both positive and negative outcomes, so the self-citation is not forcing the result. The central interaction claim is plausible and useful, just under-supported. Who is this for? Practitioners choosing input representations and model combinations. It would benefit from multi-seed runs, a simple baseline that retains static pose alongside motion features, and at least a statement about variance. With those, it would be a solid evaluation result. As it stands, it deserves a serious referee—the question is empirically meaningful and the data are on public benchmarks—but the referee should ask for the missing statistical support before accepting the interaction claim. My recommendation: send it to peer review, with the expectation of major revision rather than a quick accept.","headline":"A useful but under-powered evaluation study: the reported architecture-dependent effect of Taylor skeletons is plausible and honestly discussed, yet the missing error bars and confounded Taylor configuration keep the headline interaction from being fully established.","tokens_in":87544,"tokens_out":1335,"would_cite":true,"duration_ms":17458,"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":"Motion-injected Taylor skeletons improve ST-GCN accuracy on NTU-60 and NTU-120 but slightly reduce Hyperformer accuracy, showing that the value of motion-enriched skeletons depends on the model architecture.","keywords":["skeleton-based action recognition","ST-GCN","Hyperformer","Taylor-transformed skeletons","motion dynamics","hypergraph","NTU RGB+D","comparative evaluation"],"falsifier":"Re-run the ST-GCN and Hyperformer evaluations on NTU-60 Cross-Subject with a different Taylor block size (for example, eight frames) or with first- and second-order terms included, and check whether Hyperformer's top-1 accuracy still falls below its original-skeleton baseline; if the drop shrinks, reverses, or moves to other classes, the architecture-dependence claim is tied to the untuned transform configuration rather than to Taylor skeletons as such.","tokens_in":86559,"feed_emoji":"🦴","tokens_out":7754,"duration_ms":71851,"temperature":0.7,"pith_summary":"This paper sets out to answer whether enriching skeleton sequences with motion dynamics helps action recognition, and it finds that the answer depends on which model consumes the input. Using Taylor-transformed skeletons—joint positions augmented with velocity-like temporal derivatives—the authors report consistent top-1 accuracy gains for the graph-based ST-GCN on all four NTU benchmarks, and consistent losses for the hypergraph-transformer Hyperformer on the same benchmarks. Hyperformer nevertheless remains more accurate than ST-GCN in every setup, because its higher-order joint modeling gives it a starting advantage. The paper concludes that motion-injected input is not universally better: it supplies cues that fixed-topology graph convolutions can use, but it strips spatial arrangement that an adaptive hypergraph attention model still relies on. That is why the paper calls for architectures designed explicitly for motion-rich skeletons.","feed_headline":"Taylor skeletons lift ST-GCN, drag Hyperformer","feed_subtitle":"The same motion-rich skeletons add 2 to 3 points for ST-GCN and cost Hyperformer 4 to 7, so architecture decides.","key_machinery":"The Taylor-transformed skeleton sequence is the load-bearing input representation: joint positions are augmented with temporal derivatives—zeroth-order positions, first-order velocity from frame-to-frame differences, and second-order acceleration—combined in the displacement concept with one term, four frames per temporal block, and step size one. This object carries the comparison: both models receive identical static or motion-enriched inputs, so any accuracy difference is attributed to how the architecture uses the added dynamics.","core_discovery":"On the paper's own terms, the central finding is that injecting motion dynamics into skeleton sequences through the Taylor transform helps one architecture and hurts another. With ST-GCN, top-1 accuracy rises on all four benchmarks (NTU-60 X-Sub 81.5 to 83.5, X-View 88.3 to 89.4; NTU-120 X-Sub 70.7 to 74.1, X-Set 73.2 to 75.8). With Hyperformer, the same transformed input lowers accuracy on all four benchmarks (90.7 to 86.7, 95.1 to 92.1, 86.6 to 79.1, 88.0 to 81.9), and Hyperformer still beats ST-GCN in every configuration. The authors interpret this as evidence that Taylor skeletons supply motion-sensitive features that graph convolutions can exploit but that obscure spatial joint arrangement that the hypergraph-transformer still depends on.","pith_inferences":["A per-model search over the Taylor configuration (block size, step size, or including second-order terms) might shrink or reverse the Hyperformer loss, since the paper uses one fixed, untuned configuration for both architectures.","A natural testable extension is a two-stream input that concatenates or fuses original and Taylor skeletons; the confusion matrices suggest this would recover Hyperformer's spatially reliant classes while keeping ST-GCN's motion gains.","The same interaction may generalize beyond these two models: architectures with fixed anatomical topology benefit from explicit derivatives, while attention-based models that can already infer dynamics from raw positions may only be hurt by the loss of spatial detail.","Per-class gains and losses are concentrated in actions with fine hand and finger motion, so a motion representation that encodes local joint-group derivatives rather than global displacement might serve both architectures."],"forward_implications":["For ST-GCN, motion-injected skeletons act as implicit temporal feature engineering: adding them lifts accuracy by roughly 1 to 4 points across benchmarks without changing the network.","For Hyperformer, Taylor skeletons consistently cost 3 to 7 points, meaning the model sacrifices spatial joint-arrangement information that its hypergraph self-attention still depends on.","The benefit is action-specific: dynamic actions such as using a fan, wearing a shoe, and hopping improve, while spatially or fine-motor actions such as pointing, writing, and cutting with scissors degrade.","Because Hyperformer beats ST-GCN even with the transformed input, higher-order hypergraph modeling is the stronger baseline, but it is not the right home for this motion representation.","A hybrid that preserves spatial structure while adding motion derivatives is the paper's stated next step, and the confusion-matrix analysis gives a per-class map of where such a hybrid would help."],"supporting_citations":[{"why":"Supplies the Taylor-transformed skeleton construction and the specific displacement configuration (one term, four frames, step size one) used for all experiments.","marker":"[95]"},{"why":"Defines the ST-GCN baseline whose accuracy improves with Taylor skeletons in all four benchmarks.","marker":"[110]"},{"why":"Defines the Hyperformer baseline whose accuracy drops with Taylor skeletons while still beating ST-GCN.","marker":"[120]"},{"why":"Provides the NTU-60 dataset and its Cross-Subject and Cross-View benchmarks used for the comparisons.","marker":"[71]"},{"why":"Provides the NTU-120 dataset and its Cross-Subject and Cross-Setup benchmarks used for the comparisons.","marker":"[56]"}],"fun_headline_variants":["Taylor skeletons boost ST-GCN, sink Hyperformer","Motion-rich skeletons: ST-GCN gains, Hyperformer loses","Same motion input: ST-GCN up, Hyperformer down","Taylor transform: ST-GCN benefits, Hyperformer suffers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison rests on the assumption that one fixed, untuned Taylor configuration is a fair test of motion-injected input for both ST-GCN and Hyperformer; if that configuration were tuned separately for each model, the reported gains and losses could change or disappear.","fun_headline_variants_meta":{"raw":{"variants":["Taylor skeletons boost ST-GCN, sink Hyperformer","Motion-rich skeletons: ST-GCN gains, Hyperformer loses","Same motion input: ST-GCN up, Hyperformer down","Taylor transform: ST-GCN benefits, Hyperformer suffers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000538,"raw_usage":{"total_tokens":2592,"prompt_tokens":964,"completion_tokens":1628,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":1560}},"tokens_in":580,"tokens_out":1628,"duration_ms":12929,"temperature":1.0,"reasoning_tokens":1560,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:08:39.409045+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the ST-GCN and Hyperformer evaluations on NTU-60 Cross-Subject with a different Taylor block size (for example, eight frames) or with first- and second-order terms included, and check whether Hyperformer's top-1 accuracy still falls below its original-skeleton baseline; if the drop shrinks, reverses, or moves to other classes, the architecture-dependence claim is tied to the untuned transform configuration rather than to Taylor skeletons as such.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the ST-GCN baseline whose accuracy improves with Taylor skeletons in all four benchmarks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the NTU-60 dataset and its Cross-Subject and Cross-View benchmarks used for the comparisons."}],"review_version":1}