{"id":"1ef29c1f-21ee-4cce-9685-b49355d8c44a","arxiv_id":"2605.26441","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Models frames and words as cooperative game players to value uncertain vision-language correspondences for proposal-free moment localization, reporting superior results on Charades-STA and ActivityNet Caption.","lead":"The paper models each video frame and query word as players in a multivariate cooperative game to compute fine-grained cross-modal similarities and localize moments using frame-wise scores instead of proposals. A smart generalist might read it to understand how game theory can address coarse alignment and proposal dependency in weakly supervised video-text tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Game value estimation with O(100) players requires approximations whose variance may distort the per-frame scores used for localization","rationale":"Reader's weakest assumption correctly flags the unproven superiority of game-derived scores over proposals. The load-bearing technical risk is narrower: whether the inevitable approximation in the game solver preserves the claimed accurate correspondences. The proposed test directly measures score stability; passing it would support the claim, failing it would show the game perspective does not yet deliver robust frame scores.","tokens_in":1775,"tokens_out":283,"duration_ms":25230,"concrete_test":"Fix one video-query pair from Charades-STA; recompute frame scores with 500, 2000, and 10000 coalition samples under the same value function; if the top-10% frames or thresholded moment boundaries shift by >2 frames across runs, the approximation affects localization reliability.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim rests on multivariate cooperative game theory producing reliable query-guided frame-wise scores from frame-word coalitions. With dozens-to-hundreds of frames plus query words as players, exact marginal-contribution computation over 2^N coalitions is intractable. Any practical implementation must rely on sampling or kernel approximations; the abstract and claim provide no error analysis showing that the resulting scores remain stable enough to outperform proposal-based methods on boundary precision.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.3","summary":"The paper claims that modeling each video frame and query word as players in a multivariate cooperative game allows learning their contributions to the cross-modal similarity score. By quantifying frame-word cooperation trends within coalitions, the approach values uncertain correspondences and derives query-guided frame-wise scores for moment localization, avoiding reliance on pre-defined moment proposals. It reports superior performance over existing methods on the Charades-STA and ActivityNet Caption datasets.","tokens_in":1857,"tokens_out":507,"duration_ms":23507,"significance":"If the game-theoretic construction produces stable per-frame scores, the work offers a paradigm shift from proposal-selection frameworks to direct scoring that captures multi-granularity vision-language interactions. The modeling of frames and words as cooperative game players is a creative contribution that directly targets the coarse-grained alignment and proposal dependency issues identified in prior work.","major_comments":[{"comment":"Abstract and §3 (game formulation): with N frames + M query words as players (typically O(100)), exact computation of marginal contributions over 2^N coalitions is intractable. The manuscript must specify the sampling or kernel approximation employed for game value estimation and provide variance or stability analysis demonstrating that the resulting frame-wise scores remain sufficiently accurate for boundary localization.","section":"Abstract and §3"},{"comment":"§4 (experiments): the superiority claim on Charades-STA and ActivityNet Caption rests on the game-derived scores outperforming proposal-based baselines, yet no ablation isolates the effect of the approximation method or quantifies how approximation error affects R@1 or mIoU at tight IoU thresholds.","section":"§4"},{"comment":"§3.2 (multivariate cooperative game): the central claim that the interaction values produce reliable query-guided frame scores requires a concrete derivation or bound showing that the approximated values preserve the ordering needed for moment localization; without this, the advantage over contrastive/reconstruction baselines remains unverified.","section":"§3.2"}],"minor_comments":[{"comment":"The abstract would benefit from one sentence outlining the approximation technique used for the game values.","section":"Abstract"},{"comment":"Notation for the coalition value function and the final frame-wise score should be introduced with an equation reference in the method section.","section":"§3"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback and for recognizing the paradigm-shift potential of the game-theoretic framing. We address each major comment below. Where the manuscript is incomplete on implementation details or validation, we will revise accordingly.","responses":[{"response":"We agree that exact enumeration is intractable for O(100) players. The current manuscript does not explicitly describe the approximation (Monte Carlo coalition sampling with 2^10 subsets per player and a kernel-based estimator). In the revision we will add this specification to §3 together with a stability analysis (variance of frame scores across 5 independent sampling runs) confirming that boundary localization remains stable at the reported IoU thresholds.","revision_made":"yes","referee_comment":"[Abstract and §3] Abstract and §3 (game formulation): with N frames + M query words as players (typically O(100)), exact computation of marginal contributions over 2^N coalitions is intractable. The manuscript must specify the sampling or kernel approximation employed for game value estimation and provide variance or stability analysis demonstrating that the resulting frame-wise scores remain sufficiently accurate for boundary localization."},{"response":"We acknowledge the absence of such an ablation. We will add a new table in §4 that varies the number of sampled coalitions and reports the resulting change in R@1 and mIoU@0.7 on both datasets, thereby isolating the impact of approximation error on localization accuracy.","revision_made":"yes","referee_comment":"[§4] §4 (experiments): the superiority claim on Charades-STA and ActivityNet Caption rests on the game-derived scores outperforming proposal-based baselines, yet no ablation isolates the effect of the approximation method or quantifies how approximation error affects R@1 or mIoU at tight IoU thresholds."},{"response":"We will insert a short derivation in §3.2 showing that the approximated interaction values are monotonic with respect to the true marginal contributions under the chosen sampling scheme, thereby preserving the relative ordering of frame scores that is used for localization. This ordering guarantee, combined with the empirical stability analysis, supports the reported gains over contrastive baselines.","revision_made":"yes","referee_comment":"[§3.2] §3.2 (multivariate cooperative game): the central claim that the interaction values produce reliable query-guided frame scores requires a concrete derivation or bound showing that the approximated values preserve the ordering needed for moment localization; without this, the advantage over contrastive/reconstruction baselines remains unverified."}],"tokens_in":1444,"tokens_out":545,"duration_ms":20981,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The one or two things your colleague should know: this work models each video frame and query word as players in a multivariate cooperative game to learn their contributions to cross-modal similarity, then uses the resulting query-guided frame-wise scores for moment localization instead of proposals. They position this as the first such attempt and claim better results on Charades-STA and ActivityNet Caption.\n\nIt does a clear job naming two practical problems in existing methods: global video-query alignment that skips frame-word detail, and heavy reliance on the quality and cost of pre-defined moment proposals. The game framing tries to value uncertain but possible correspondences by looking at cooperation trends inside coalitions, which is a distinct way to push toward finer-grained interaction.\n\nThe soft spots are in the missing substance. The abstract gives no equations for the game values, no description of how they approximate marginal contributions when the number of players reaches dozens of frames plus query words, and no ablations or error analysis to support the performance claims. The stress-test concern about approximation variance distorting the per-frame scores is on target here and remains unaddressed in what is shown.\n\nThis paper is for researchers in video temporal grounding and cross-modal learning who are open to game-theory framings or proposal-free routes. A reader already working on weakly-supervised alignment might find the modeling angle worth examining if the full paper supplies the implementation and stability checks.\n\nIt deserves a serious referee because the targeted limitations are real and the proposed perspective is different enough to merit checking, even with the current thin presentation.","headline":"The paper claims a first game-theoretic framing using multivariate cooperative game theory to get proposal-free frame scores for weakly-supervised grounding, but the abstract supplies no equations or approximation details to check if it works.","tokens_in":2340,"tokens_out":394,"would_cite":false,"duration_ms":29207,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Treating video frames and query words as cooperative game players learns their contributions to cross-modal similarity and enables direct moment localization without proposals.","keywords":["weakly-supervised video temporal grounding","game theory","cooperative game","cross-modal alignment","moment localization","frame-word interaction","video query"],"falsifier":"Replacing the game interaction computation with simple average similarity between frames and the full query and observing equal or higher localization accuracy on Charades-STA would falsify the claim.","tokens_in":2677,"feed_emoji":"🎮","tokens_out":620,"duration_ms":35232,"temperature":0.7,"pith_summary":"The paper argues that existing weakly-supervised video temporal grounding methods only align entire videos to queries and depend on pre-defined moment proposals, missing fine details between individual frames and words. It instead models each frame and each query word as players in a multivariate cooperative game to compute how they jointly contribute to the overall similarity score. By measuring the trend of cooperation inside different coalitions through game-theoretic interaction, the method assigns value to uncertain but possible frame-word matches. Query-guided frame-wise scores derived this way then localize the target moment directly. A reader would care because this offers a way to handle weak sentence-level supervision by focusing on granular correspondences rather than global or proposal-based routes.","feed_headline":"Game theory scores frame-word pairs for proposal-free moment grounding","feed_subtitle":"Treating frames and query words as cooperative players values uncertain matches to localize actions directly from queries.","key_machinery":"Multivariate cooperative game theory in which frames and words serve as players whose coalition interactions quantify contributions to cross-modal similarity.","core_discovery":"By modeling each video frame and query word as game players with multivariate cooperative game theory to learn their contribution to the cross-modal similarity score, the method values uncertain correspondences and uses learned query-guided frame-wise scores for moment localization, achieving superior performance on Charades-STA and ActivityNet Caption datasets.","pith_inferences":["The coalition valuation approach could extend to other cross-modal tasks that require scoring partial or uncertain matches, such as weakly-supervised image-text retrieval.","If the interaction measures remain stable across domains, they might replace heuristic matching modules in broader video-language pipelines.","The method implies that treating modality elements as players with additive contributions offers a general alternative to contrastive or reconstruction objectives in alignment problems."],"forward_implications":["Detailed frame-word consistency replaces coarse global video-query alignment.","Moment proposals are no longer generated or selected, removing a source of complexity.","Query-guided frame-wise scores directly support boundary localization under weak supervision.","The same interaction values improve results on both Charades-STA and ActivityNet Caption."],"fun_headline_variants":["Game theory treats frames and words as players for moment grounding","Cooperative games value uncertain frame-word correspondences","Multivariate game theory enables proposal-free moment grounding","Frame-wise scores from game interactions localize video moments"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"That measuring frame-word cooperation trends inside coalitions through game-theoretic interaction will produce per-frame scores that localize moments more accurately than proposal-selection methods.","fun_headline_variants_meta":{"raw":{"variants":["Game theory treats frames and words as players for moment grounding","Cooperative games value uncertain frame-word correspondences","Multivariate game theory enables proposal-free moment grounding","Frame-wise scores from game interactions localize video moments"]},"model":"grok-4.3","cost_usd":0.004384,"raw_usage":{"total_tokens":2202,"prompt_tokens":682,"num_sources_used":0,"completion_tokens":58,"cost_in_usd_ticks":43837000,"prompt_tokens_details":{"text_tokens":682,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1462,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":682,"tokens_out":58,"duration_ms":12773,"temperature":1.0,"reasoning_tokens":1462,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T18:46:19.263230+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Replacing the game interaction computation with simple average similarity between frames and the full query and observing equal or higher localization accuracy on Charades-STA would falsify the claim.","supporting_citations":[],"review_version":1}