{"id":"3bfbfcef-6687-4d4a-9e6b-0c86c37ed31b","arxiv_id":"2412.15673","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A tactic-conditioned diffusion model with Banzhaf-style interaction scores improves multi-player trajectory and tactic prediction on NBA SportVU and TeamTrack data.","lead":"This paper predicts where basketball and soccer players move next by adding team tactic labels and a game-theory style cooperation score as extra conditions to a diffusion model. It also labels NBA plays with team tactics and reports better trajectory and tactic forecasts than several existing methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Tactic-condition gains may be an artifact of oracle/leaky labels: past tactic labels are human annotations from full clips, and label-free baselines receive no tactic condition, so Table 1's improvement is not yet attributable to the method.","rationale":"The reader's weakest assumption identifies label availability and pseudo-label bias, which is close to my concern. I sharpen it: the issue is not only deployment availability but also potential future leakage in the annotation protocol. If the annotators viewed the full clip, the past-tactic labels can carry information about future trajectories, making the comparison against label-free baselines unfair and potentially invalidating the central SOTA claim. This is the most load-bearing risk because Table 1 is the paper's main empirical evidence. The Banzhaf interaction mismatch (Eq. 24 vs. Eq. 5) is a real correctness flaw, but it mainly undermines the interpretability/novelty claims about semantic intentions; it does not directly explain trajectory gains. The proposed concrete test would separate genuine tactic-conditioning from oracle-label leakage and from clustering artifacts. Because the concern is testable and the reader's conditional verdict already requires additional evidence, I do not change the verdict. I mark agreement as partial because the leakage angle is a distinct, stronger version of the reader's label-availability concern.","tokens_in":20428,"tokens_out":5450,"duration_ms":51697,"concrete_test":"Permute the tactic labels across training scenes (breaking label-trajectory correspondence but preserving label distribution) and retrain the reported model; compare Table 1 minADE/minFDE. Also train a tactic recognizer on observed frames only, freeze it, and feed its output as the condition at test time. If the human-label variant is statistically indistinguishable from the random-label variant, or if the recognized-label variant degrades to the random-label level, the trajectory gains are not caused by genuine tactic semantics and may leak future information. As a complementary check, re-annotate a held-out subset using only the first 10 frames and rerun the evaluation; any drop in performance relative to the full-clip annotations indicates leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim (Table 1) rests on conditioning the diffusion model on past team-tactic labels L^{-T_obs+1:0} via Eq. (8). Those labels are not sensor-observable: on NBA SportVU they come from human annotation (Sec. 5.2), and the annotation protocol says annotators label 'the previous 10 frames and the following 20 frames' and select a tactic for each time period. The paper never states that past-tactic labels were assigned using only the 10 observed frames. If annotators watched the full 30-frame clip, the condition can encode future trajectory information, directly inflating minADE/minFDE relative to label-free baselines (LED, MID, SingularTrajectory) that do not receive this signal. On TeamTrack, the condition is a K-means pseudo-label (Sec. 5.2) with no external tactic ground truth, so the model may simply exploit cluster IDs of observed motion; no experiment separates the contribution of 'true tactic semantics' from the clustering artifact. The Banzhaf module does not rescue the trajectory claim: Eq. (24) is an average of two masked softmax logits, not the Banzhaf interaction of Eq. (5), and its supervision I_B is generated by the same model. Therefore the reported SOTA trajectory numbers are evidence only for a tactic-conditioned diffusion model with oracle-like labels, not for the paper's semantic-intention mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a diffusion-based trajectory prediction framework for team sports that conditions trajectory generation on observed group-level tactic labels, enhances agent embeddings with global/local attention, and introduces an auxiliary tactic-prediction task supervised by a learned 'Banzhaf Interaction' module. The authors extend the NBA SportVU dataset with human annotations of 16 team tactics and generate K-means pseudo-labels for the TeamTrack datasets. Experiments report state-of-the-art minADE/minFDE on NBA SportVU and TeamTrack, as well as improved Top-k tactic prediction accuracy over Transformer, pooling, and LLM baselines. The central claim is that group-level semantic context, modeled via game-theoretic interaction, improves both trajectory and tactic forecasting.","tokens_in":20749,"tokens_out":4261,"duration_ms":33657,"significance":"If validated, the framework would be a meaningful step toward incorporating group-level semantic knowledge into generative trajectory prediction, and the human-annotated NBA SportVU tactic dataset is a useful community resource. The paper ships code and data, reproduces LED and other baselines, and evaluates against LLM baselines, which is commendable. However, the significance is contingent on resolving several load-bearing concerns: the annotation protocol may permit future information leakage into the conditioning signal; the 'Banzhaf Interaction' computation in Eq. (24) does not implement the game-theoretic definition in Eq. (5) and its supervision is generated from the model's own predictions; the TeamTrack pseudo-labels lack external validation; and the reported numeric gains are presented without error bars or a condition-ablation that isolates the contribution of tactic labels.","major_comments":[{"comment":"The paper never states whether human annotators assigned past tactic labels L^{-T_obs+1:0} using only the observed 10 frames or the full 30-frame clip. If annotators watched the future 20 frames, the conditioning variable in Eq. (8) encodes future trajectory information, directly inflating the Table 1 gains over label-free baselines (LED, MID, SingularTrajectory). The authors must clarify the annotation protocol and, if necessary, re-annotate or demonstrate that labels are assigned from observed history alone; otherwise the central trajectory claim is not attributable to the method.","section":"Sec. 5.2 / Eq. (8)"},{"comment":"Eq. (24) defines I^j_B as the average of two masked softmax logits (Sa2t and St2a), which is not the Banzhaf interaction defined in Eq. (5): there is no summation over coalitions C with p(C)=1/2^{n-2} and no characteristic-function payoff differences. Additionally, the supervision I^j_B is computed from the model's own predicted Top-k tactic embeddings c_hat_j (Sec. 4.5.2), making the 'ground truth' self-generated rather than an externally anchored game-theoretic value. This undermines the paper's claim to model semantic intentions via Banzhaf interaction; either implement the actual Banzhaf interaction or rename the module and reframe the claims.","section":"Sec. 4.5.3, Eq. (24) vs Eq. (5)"},{"comment":"No error bars, standard deviations, or statistical significance tests are reported for any of the trajectory or tactic prediction results. The claimed improvements over LED (e.g., 0.19 vs 0.21 minADE at 1s in Table 1) may be within run-to-run variance. The authors should report means and standard deviations over multiple random seeds and, if possible, perform significance testing to support the state-of-the-art claim.","section":"Tables 1, 2, 3"},{"comment":"The TeamTrack pseudo-labels are produced by K-means clustering of trajectories into 16 clusters, with no external validation that these clusters correspond to meaningful team tactics. The t-SNE visualization in Fig. 3 and the referenced generalizability experiments do not establish such correspondence. Without a comparison to human-annotated tactics or other external grounding, the Table 2 results may reflect the model exploiting cluster-ID artifacts rather than semantic intentions. Provide quantitative validation of the pseudo-labels or remove/qualify the TeamTrack claims.","section":"Sec. 5.2 / Table 2"},{"comment":"The main trajectory improvement is attributed to conditioning on group-level tactic labels L, but no ablation isolates this factor. Table 4 varies the generation backbone (MLP, query, diffusion) while keeping the interaction encoder fixed, and the ablation in Sec. 6.7 does not compare against a diffusion model trained without the tactic condition or with random tactic labels. Without such an ablation, the reported gains cannot be attributed to the semantic intention mechanism rather than to architectural choices.","section":"Sec. 4.2 and Sec. 6.7"}],"minor_comments":[{"comment":"The implementation details state 'we set the game factor to 0.001 and the γ in Eq. (21) to 4.0. The tactic prediction rate is set at 1000.' It is unclear which symbols correspond to α, β, and η in Eq. (25), and what 'tactic prediction rate' means; please clarify.","section":"Sec. 6.3"},{"comment":"The BI module improves Top-1 accuracy by only about 1.8 percentage points (from 58.44 to 60.23); given the absence of error bars, it is unclear whether this gain is statistically significant.","section":"Table 5"},{"comment":"The proportionality '∝' in the classifier-free guidance update is loose; the exact update is fϵ(y, G) = fϵ(y, A) + sg·(fϵ(y, G) − fϵ(y, A)). Please state the update directly.","section":"Eq. (9)"},{"comment":"The notation 'arg max_k (p_j)' is ambiguous; it should be defined as selecting the indices of the k largest probabilities, not as a standard argmax over a single value.","section":"Sec. 4.5.2, Eq. (18)"},{"comment":"The t-SNE plot alone does not validate that K-means clusters align with semantic tactic categories; please add a quantitative measure (e.g., cluster–label agreement on a human-annotated subset) or move the figure to supplementary with explicit caveats.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The tactic-label leakage concern in Sec. 5.2 is the most serious issue: if annotators saw the future frames when labeling the past tactic condition, the Table 1 results are not a valid comparison against label-free baselines. This should be verified with the authors before any acceptance decision. The Banzhaf module's mismatch with the stated game-theoretic definition is also a conceptual defect that needs either a correct implementation or a re-framing. The paper has merit in its dataset contribution and diffusion framework, but these load-bearing issues must be resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful paper to engage, but the strongest result is not the semantic-intention story; it is the tactic-conditioned diffusion recipe plus the annotated NBA data. The trajectory gains over LED are consistent across horizons, the ablations show diffusion beats MLP and query decoders, and the tactic prediction numbers beat sensible baselines. Those parts are worth a referee's time.\n\nWhat's genuinely new: conditioning a diffusion trajectory model on team-level tactic embeddings with classifier-free guidance, a dual tactic forecast head, and a human-annotated NBA SportVU tactic extension. The dataset annotation effort (16 tactics, 8,238 entries) is a concrete contribution if released.\n\nSoft spots, in order of severity. First, the input tactic labels are the load-bearing condition in Eq (8), and the paper never shows they are available or recognizable at inference from observed frames alone. The annotation protocol says annotators label tactics for \"the previous 10 frames and the following 20 frames\"; if the past-tactic label was assigned while watching the full clip, Table 1's margin over label-free baselines is partly leakage, not method. This needs an automatic recognizer or a careful protocol statement. On TeamTrack, the K-means pseudo-labels are even weaker; results there show only that cluster IDs of observed motion help, not that semantic tactics help.\n\nSecond, the Banzhaf module doesn't implement Eq (5). Eq (24) averages two masked softmax logits; it is not a sum over coalitions of characteristic-function differences. That is a naming overclaim, and the same model produces the \"ground-truth\" I_B from its own predicted top-k tactic embeddings. The tactic head does have external human labels, so this is circular only in the auxiliary interaction target, not the whole system, but the interpretability story is not supported.\n\nThird, no error bars anywhere, and Tables 1-3 report single runs. Minor relative to the other two, but with margins of 5-15% it matters.\n\nMy bottom line: send it to review, but as a major-revision candidate. The empirical core—tactic-conditioned diffusion plus the annotated benchmark—is worth publishing if the authors either fix the Banzhaf math/name or drop the claim, and if they address the label-leakage question with an ablation or automatic recognition. I wouldn't cite the Banzhaf number until then, but I would cite the dataset once released.","headline":"Tactic-conditioned diffusion for sports trajectory prediction is a real empirical idea with a usable new benchmark, but the semantic-intention module overclaims Banzhaf and the headline gains may rest on oracle-like tactic labels.","tokens_in":21297,"tokens_out":2476,"would_cite":false,"duration_ms":22319,"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 argues that conditioning a diffusion-based trajectory predictor on team-level tactic labels and game-theoretic Banzhaf interaction scores improves multi-agent trajectory and tactic forecasting in team sports.","keywords":["trajectory prediction","multi-agent forecasting","diffusion model","classifier-free guidance","group interaction","Banzhaf interaction","tactic prediction","sports analytics"],"falsifier":"Permute or randomly drop the observed tactic labels at inference on NBA SportVU and re-measure minADE/minFDE: if scores stay near 0.19/0.29 rather than reverting toward the LED baseline's 0.21/0.31, the tactic condition is not the cause of the gain. Re-running TeamTrack evaluation with independent tactic ground truth instead of the K-means pseudo-labels would test whether those labels carry the improvement.","tokens_in":20161,"feed_emoji":"🏀","tokens_out":8962,"duration_ms":73290,"temperature":0.7,"pith_summary":"The paper tries to establish that group-level semantic context improves multi-agent trajectory prediction in team sports. It builds a diffusion-based forecaster whose denoising process is conditioned not only on observed positions but on each team's current tactic label, and it adds a game-theoretic scoring of how strongly agents and tactics cooperate. On the NBA SportVU dataset the method reports the best accuracy among the compared methods, improving on the reproduced LED baseline; it also reports gains on TeamTrack soccer and basketball. If the claim holds, tactic-conditioned diffusion is a workable route to more realistic and interpretable sports trajectory forecasts.","feed_headline":"Team tactics as diffusion guidance cut NBA trajectory error by 9.5%","feed_subtitle":"Tactic-conditioned paths beat prior baselines and forecast the next tactic 60% of the time.","key_machinery":"The carrying object is the interaction-condition vector $G = f_\\phi(X, L^{-T_{obs}+1:0})$, where each agent's token is the concatenation of a learned trajectory embedding and a tactic embedding drawn from a vocabulary; the denoising model then follows the classifier-free guidance rule $f_\\epsilon(y, G) = f_\\epsilon(y, A) + s_g (f_\\epsilon(y, G) - f_\\epsilon(y, A))$. The second mechanism is Banzhaf Interaction, a cooperative-game score for how much a coalition's joint contribution exceeds its members' separate contributions, computed between agents and top-$k$ candidate tactics; its predicted values are fused with globally and locally attended agent tokens, and a graph-attention head outputs tactic probabilities.","core_discovery":"On the paper's own terms, the central claim is that group-level semantic context is a usable condition for trajectory generation. The authors construct the condition vector $G = f_\\phi(X, L^{-T_{obs}+1:0})$ by concatenating each agent's trajectory embedding with an embedding of its team's current tactic, feed it into the denoising network through classifier-free guidance, and supervise an auxiliary tactic predictor with Banzhaf Interaction scores computed between agents and candidate tactics. They report minADE/minFDE of 0.19/0.29 at 1s and 0.84/1.17 at 4s on NBA SportVU, down from the reproduced LED baseline's 0.21/0.31 and 0.89/1.24, and Top-1 tactic accuracy of 60.23%; on TeamTrack soccer and basketball the same recipe also improves on the compared baselines.","pith_inferences":["Beyond the paper, a learned tactic recognizer could replace the observed labels at inference, since the current formulation consumes manual or pseudo labels as input rather than detecting them from motion alone.","Beyond the paper, the computed Banzhaf Interaction values could serve as a per-player importance map accompanying each forecast, extending the paper's visualization that the ball handler and nearest defenders receive the highest scores.","Beyond the paper, corrupting tactic labels with controlled noise and measuring minADE/minFDE would map how accurate an automatic labeler must be for the reported gains to survive."],"forward_implications":["Tactic-conditioned diffusion gives a template for group-behavior forecasting wherever a team-level semantic label is available.","One forward pass produces both future trajectories and future tactic probabilities, so a downstream system could anticipate an offensive or defensive switch before it is visible in player positions.","Because the method uses classifier-free guidance, tactic information can be randomly dropped during training, letting users tune guidance strength at inference to balance diversity and fidelity.","The manually annotated NBA tactic benchmark provides a supervised target for future tactic-conditioned trajectory models and for studying label noise."],"supporting_citations":[{"why":"Supplies the diffusion trajectory prediction formulation and the noise-prediction loss the method builds on.","marker":"[13]"},{"why":"Provides the leapfrog initializer used to accelerate denoising and is the main trajectory baseline the experiments improve on.","marker":"[14]"},{"why":"Gives the classifier-free guidance recipe used to inject tactic embeddings into the denoising process.","marker":"[17]"},{"why":"Supplies the Banzhaf Interaction formulation and the learnable approximation used to model semantic intentions.","marker":"[18]"},{"why":"Provides the TeamTrack soccer and basketball datasets used to test generalization with pseudo-labels.","marker":"[68]"},{"why":"Supplies the variety loss that defines the best-of-20 minADE/minFDE evaluation.","marker":"[10]"},{"why":"Supplies the focal loss used to train tactic prediction under imbalanced tactic frequencies.","marker":"[66]"}],"fun_headline_variants":["Tactic-guided diffusion cuts NBA trajectory error by 9.5%","Group-aware diffusion outperforms baselines on multi-object path prediction","Team intent as Banzhaf interaction conditions trajectory diffusion","Diffusion with group interactions beats state-of-the-art on three sports datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that tactic labels for the observed frames are available and reliable at inference, since on the TeamTrack datasets the labels come from K-means clustering without external ground truth and any noise or bias in those labels could erase the reported trajectory gains.","fun_headline_variants_meta":{"raw":{"variants":["Tactic-guided diffusion cuts NBA trajectory error by 9.5%","Group-aware diffusion outperforms baselines on multi-object path prediction","Team intent as Banzhaf interaction conditions trajectory diffusion","Diffusion with group interactions beats state-of-the-art on three sports datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000674,"raw_usage":{"total_tokens":3047,"prompt_tokens":902,"completion_tokens":2145,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":2071}},"tokens_in":518,"tokens_out":2145,"duration_ms":15909,"temperature":1.0,"reasoning_tokens":2071,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:11:54.015862+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Permute or randomly drop the observed tactic labels at inference on NBA SportVU and re-measure minADE/minFDE: if scores stay near 0.19/0.29 rather than reverting toward the LED baseline's 0.21/0.31, the tactic condition is not the cause of the gain. Re-running TeamTrack evaluation with independent tactic ground truth instead of the K-means pseudo-labels would test whether those labels carry the improvement.","supporting_citations":[{"cited_title":"Stochastic trajectory prediction via motion indeterminacy diffusion,","cited_arxiv_id":null,"evidence_quote":"Supplies the diffusion trajectory prediction formulation and the noise-prediction loss the method builds on."},{"cited_title":"Leapfrog diffusion model for stochastic trajectory prediction,","cited_arxiv_id":null,"evidence_quote":"Provides the leapfrog initializer used to accelerate denoising and is the main trajectory baseline the experiments improve on."},{"cited_title":"Video-text as game players: Hierarchical banzhaf interac- tion for cross-modal representation learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the Banzhaf Interaction formulation and the learnable approximation used to model semantic intentions."},{"cited_title":"TeamTrack: A Dataset for Multi-Sport Multi-Object Tracking in Full-pitch Videos","cited_arxiv_id":"2404.13868","evidence_quote":"Provides the TeamTrack soccer and basketball datasets used to test generalization with pseudo-labels."},{"cited_title":"Social gan: Socially acceptable trajectories with generative adversarial networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the variety loss that defines the best-of-20 minADE/minFDE evaluation."},{"cited_title":"Focal loss for dense object detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the focal loss used to train tactic prediction under imbalanced tactic frequencies."}],"review_version":1}