{"id":"bddf84d4-e521-41f3-9c5c-8fa84416ebf2","arxiv_id":"2504.12021","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A new action anticipation benchmark for football broadcasts, with a dataset, metrics, and a baseline model that predicts ball-related actions up to ten seconds ahead.","lead":"This paper introduces the first benchmark for anticipating football actions, such as passes or shots, before they happen in broadcast video. The authors release a new dataset built from SoccerNet, along with a transformer-based baseline model and new evaluation metrics.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Joint-training experiments may leak SN-BAS test matches through SN-AS; the headline 24.08 mAP is unverified unless the splits are shown disjoint.","rationale":"The reader's conditional verdict focuses on error bars, inference decision rule, clipping, and class exclusion. Those are addressable. I looked for the condition that would have to be true for the central claim to hold: the reported mAP values must be valid generalization numbers. The weakest link is the joint-training protocol with SN-AS. SN-BAS is described as an adaptation of SoccerNet, and SN-AS as the original SoccerNet Action Spotting dataset. Whether the 9 SN-BAS games are a subset of the 500 SN-AS games is not addressed. If they are, the test videos are in the training set. This would inflate both FAANTRA and T-DEED results and undermine the 'feasible yet challenging' claim. The check is straightforward: compare official match IDs. This is more load-bearing than the clipping concern because the clipping issue, while under-specified, does not obviously change the reported numbers, whereas split leakage would invalidate the headline number. The paper is otherwise solid: the task definition is clear, the mAP@delta adaptation is principled, ablations are extensive, and artifacts are promised. The verdict should remain conditional, with the additional condition that split disjointness be verified or the joint-training results be reproduced on non-overlapping data.","tokens_in":19615,"tokens_out":14310,"duration_ms":148574,"concrete_test":"Extract the match identifiers for the SN-BAS train/val/test/challenge splits and compare them with the SN-AS train/val split (e.g., via the official SoccerNet match IDs). If any SN-BAS test or challenge match ID appears in SN-AS train/val, retrain the best FAANTRA (400MF, SN-AS+SN-BAA) and T-DEED with those matches removed from SN-AS, and recompute Table 2. If the average mAP drops by more than approximately 2 points, the joint-training claim is not a valid generalization result and the manuscript must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim (Table 2) is the 24.08 average mAP of FAANTRA-400MF trained jointly on SN-BAA and SN-AS. The paper says SN-AS is 'the original SoccerNet Action Spotting dataset, which includes 500 additional games,' but SN-BAA is itself adapted from SoccerNet (SN-BAS, 9 games). It is never stated that the 9 SN-BAS matches are disjoint from the 500-game SN-AS split. If the two SN-BAS test matches (and challenge matches) appear in SN-AS train/validation, then joint training has already seen the test videos, and the 19.08 to 24.08 jump from adding SN-AS could reflect video-level memorization rather than transferable anticipation. The same issue affects the T-DEED upper bound trained on SN-AS+SN-BAA, so the reported gap is not a clean generalization measure. Because the paper's framing as a benchmark depends on the reported numbers being honest generalization results, split disjointness is a load-bearing unverified assumption.","agreement_with_reader":"disagree"},"referee_report":null,"author_rebuttal":null,"desk_editor":null,"rs_alignment":null,"lean_confirmation":null,"pith_extraction":null,"created_at":"2026-08-16T12:38:57.181017+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":null,"supporting_citations":[],"review_version":1}