{"id":"10dfbc72-56a3-496c-bd6a-b02d47ff530f","arxiv_id":"2607.12592","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A distilled Wan-1.3B video diffusion system generates real-time two-player KOF '97 combat from dual keyboard input at 30 FPS on one RTX 5090.","lead":"WanToFight generates real-time two-player King of Fighters '97 matches from dual keyboard input using a distilled video diffusion model. It is of interest as a systems attempt to make multi-player adversarial gameplay run as generative video at interactive frame rates.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Abstract-only review leaves the multi-player binding claim untestable; no stronger load-bearing flaw can be isolated beyond the reader's already-flagged assumption.","rationale":"The paper is an abstract-only systems claim. The reader's weakest_assumption correctly isolates the only premise that must hold for the joint multi-player / real-time / adversarial claim to be true, and correctly notes that the abstract supplies no evidence for it. No additional internal inconsistency, circularity, or hidden mathematical assumption can be diagnosed without methods, equations, or results. Therefore the stress-test adds no new load-bearing concern and leaves the UNVERDICTED / LOW-confidence verdict unchanged. The concrete test simply operationalizes the check that would settle the shared concern once artifacts become available.","tokens_in":2072,"tokens_out":405,"duration_ms":3485,"concrete_test":"When the full paper or code appears, measure identity-swap and control-crosstalk rates on held-out full matches under simultaneous dual keyboard input (e.g., fraction of frames where the wrong character responds or both characters execute the same action). If either rate exceeds a few percent, or if removing Player Association collapses dual-control coherence, the multi-player adversarial claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that Player Association plus gated keyboard injection (trained under the single-player-to-full-gameplay curriculum) reliably bind each keyboard stream to the correct character and sustain coherent adversarial physics under streaming block-causal generation. With only the abstract available, that binding is asserted rather than evidenced: no architecture details, no identity-swap / control-crosstalk rates, no ablations of the association module or curriculum, and no dual-control metrics over full matches. This is the same load-bearing premise the reader already identified; the absence of the full text simply prevents any deeper technical objection from being formulated or refuted.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript introduces WanToFight, a generative game engine that aims to simulate real-time two-player The King of Fighters '97 gameplay from dual keyboard streams. Built on a Wan-1.3B video diffusion transformer, the system comprises (i) a streaming autoregressive generator with block-causal attention and a rolling KV cache, (ii) a visually grounded Player Association module that binds each keyboard stream to a character identity, and (iii) a gated, locally causal keyboard injection path trained under a single-player-to-full-gameplay curriculum. A four-step DMD-distilled student with a pruned VAE decoder is claimed to sustain 30 FPS at 512×384 on one RTX 5090 for a complete match. The central claim is that this is the first generative game engine to jointly deliver multi-player control, real-time inference, complex physical interaction, and adversarial gameplay.","tokens_in":2246,"tokens_out":967,"duration_ms":14044,"significance":"If the system works as stated, the contribution is substantial for generative interactive media: prior generative game engines have largely been limited to single-player first-person or non-real-time cooperative settings, and jointly solving dual-control binding, streaming real-time generation, and adversarial physical combat would close a clear capability gap. The engineering stack (block-causal streaming, association module, gated injection, DMD distillation, pruned VAE) is a concrete systems package that, if validated, would be useful to the community. The abstract-only record, however, does not yet let a referee confirm that those capabilities hold under dual control over full matches.","major_comments":[{"comment":"The load-bearing multi-player claim rests on the Player Association module plus gated keyboard injection (trained under the single-player-to-full-gameplay curriculum) reliably binding each concurrent keyboard stream to the correct on-screen fighter and sustaining coherent adversarial physical interaction under streaming block-causal generation. The abstract asserts this binding but reports no identity-swap / control-crosstalk rates, dual-control accuracy over full matches, ablations of association or curriculum, or failure-mode characterization. Without those measurements the multi-player and adversarial claims cannot be assessed.","section":null},{"comment":"The real-time performance claim—30 FPS at 512×384 on a single RTX 5090 for a complete match via a four-step DMD-distilled student and pruned VAE—is stated without measurement protocol (wall-clock vs. model-only, cache warm-up, resolution/cache policy, match length distribution) and without baselines or latency breakdowns. For a systems paper whose novelty includes real-time dual-control generation, this evidence is required to support the central claim.","section":null},{"comment":"The priority claim that WanToFight is the first system to jointly combine multi-player control, real-time inference, complex physical interaction, and adversarial gameplay is not supported in the abstract by a comparative table or explicit positioning against named prior generative engines. A referee cannot verify novelty or scope of the gap without that comparison and the corresponding dual-control metrics.","section":null}],"minor_comments":[{"comment":"The abstract names free design choices (DMD step count, spatial/temporal operating resolution and cache policy, curriculum schedule) without even order-of-magnitude values; a short quantitative teaser would help readers gauge the operating point.","section":null},{"comment":"Notation for the three modules (Player Association, gated locally causal keyboard injection, streaming block-causal generator) is introduced only by name; a one-line functional description of inputs/outputs for each would improve abstract clarity.","section":null},{"comment":"No dataset scale, character/move coverage, or training compute is mentioned; even a brief clause would situate the empirical scope.","section":null}],"recommendation":"uncertain","confidential_remarks":"This is an abstract-only review: the full manuscript was not available. The central multi-player binding premise is asserted rather than evidenced in the abstract, so I cannot responsibly recommend accept, minor_revision, major_revision, or reject on technical grounds. Please route the full paper (methods, ablations, dual-control metrics, failure cases, baselines) for a complete review. If the full text lacks identity-swap/crosstalk rates and dual-control full-match evaluation, a major_revision or reject would then be appropriate; if those results are strong, the systems contribution looks journal-relevant for cs.CV / generative interactive media."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is an abstract-only systems paper claiming the first generative game engine that jointly does multi-player keyboard control, real-time inference, complex physical interaction, and adversarial combat (KOF '97) at 30 FPS 512×384 on one RTX 5090 for a full match. That joint capability is the punchline. If it holds, it is a real within-subfield advance over prior single-player or non-real-time generative engines.\n\nWhat looks new is the integration, not the individual pieces. They start from Wan-1.3B, add streaming block-causal attention with rolling KV cache, a four-step DMD-distilled student plus pruned VAE for the speed claim, a visually grounded Player Association module, and gated locally causal keyboard injection trained under a single-player-to-full-gameplay curriculum. The engineering story is coherent on paper: distillation and pruning for latency, association plus gated injection for dual control, curriculum to stage the hard multi-agent interaction. Credit where due—the abstract is clear about the gap it targets and names concrete modules and a concrete operating point.\n\nThe soft spot is exactly the one the reader and stress-test flag, and it is load-bearing: we have no evidence that Player Association plus gated injection actually binds each keyboard stream to the right character and sustains coherent adversarial physics under streaming generation. No architecture details, no identity-swap or crosstalk rates, no ablations of association or curriculum, no dual-control metrics over full matches, no failure modes. The 30 FPS full-match claim is also just stated. Free parameters (DMD steps, cache policy, curriculum schedule) and the invented modules are uninspectable from the abstract alone. Circularity is low—this is an empirical systems claim—but soundness is currently low because the central multi-player result is asserted rather than shown.\n\nWho it is for: people building interactive generative video / game engines who care about dual control and real-time adversarial settings. A serious referee should see the full paper (methods, numbers, ablations, videos). I would not desk-reject on the abstract; the claimed joint capability is important enough to evaluate properly. I would not cite or bring to reading group until there is evidence the binding works. Send it to review if the full text materializes with real measurements; otherwise it stays a claim.","headline":"Abstract-only systems claim of first real-time dual-control generative KOF engine; interesting if true, but multi-player binding is asserted without evidence.","tokens_in":2862,"tokens_out":581,"would_cite":false,"duration_ms":5438,"reading_group":"no","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"WanToFight generates real-time two-player King of Fighters '97 combat from keyboard input, uniting multi-player control, physical interaction, and adversarial play in one generative engine.","keywords":["generative game engine","multi-player combat","video diffusion transformer","real-time inference","King of Fighters","keyboard control","player association","model distillation"],"falsifier":"Run a full two-player match and check whether each fighter consistently responds only to its assigned keyboard stream across punches, combos, jumps, and knockdowns; any sustained identity swap, frozen character, or crosstalk where one player's keys move the wrong fighter falsifies the multi-player claim.","tokens_in":2928,"feed_emoji":"🎮","tokens_out":935,"duration_ms":17957,"temperature":0.7,"pith_summary":"This paper presents WanToFight, a generative game engine that simulates complete two-player matches of The King of Fighters '97 in real time from ordinary keyboard input. Earlier generative game systems covered only single-player first-person scenes or non-real-time cooperative play; none jointly delivered multi-player control, real-time speed, complex physical interaction, and adversarial combat. WanToFight builds on a 1.3-billion-parameter video diffusion transformer, adding a streaming generator with rolling memory, a module that visually links each keyboard stream to the correct fighter, and a gated keyboard injection path trained by gradually moving from single-player to full adversarial gameplay. After distillation and decoder pruning, the system sustains 30 frames per second at 512×384 resolution on one high-end GPU for an entire match. If the approach holds, modern video generators can act as interactive multi-player game engines rather than only as passive video synthesizers.","feed_headline":"Generative engine runs two-player KOF '97 at 30 FPS from keys","feed_subtitle":"First system to combine multi-player control, real-time video, and adversarial fighting in one model.","key_machinery":"A streaming autoregressive video generator based on the Wan-1.3B diffusion transformer with block-causal attention and a rolling KV cache; a visually grounded Player Association module that binds each player's keyboard signal to a character identity; and a gated, locally causal keyboard injection module trained under a single-player-to-full-gameplay curriculum, then distilled with a pruned VAE decoder for real-time speed.","core_discovery":"WanToFight is the first generative game engine to jointly deliver multi-player control, real-time inference, complex physical interaction, and adversarial gameplay, by simulating two-player KOF '97 from keyboard input at 30 FPS on a single GPU for a complete match.","pith_inferences":["The Player Association idea may generalize to any multi-agent video task where multiple control streams must map to distinct visual entities without identity swaps.","Success at adversarial combat would suggest generative models can internalize reaction and counter-play dynamics, not merely replay cooperative motion.","If the single-player-to-full-gameplay curriculum is the decisive training step, similar staged curricula could unlock other complex multi-agent settings such as team sports or racing.","Long-match consistency of the rolling KV cache will determine whether the approach scales beyond short fighting rounds."],"forward_implications":["Full two-player fighting-game matches can be driven end-to-end by a generative video model without a conventional game engine.","Separate keyboard streams can be bound to distinct on-screen characters and produce coherent adversarial physics in real time.","Distilled diffusion models can sustain 30 FPS interactive video generation at modest resolution on a single high-end GPU for multi-minute sessions.","The same association-plus-injection recipe could support other multi-player interactive titles once trained for them.","Generative engines become candidates for interactive entertainment and training simulators that previously required hand-authored physics and animation systems."],"fun_headline_variants":["WanToFight streams two-player KOF '97 combat at 30 FPS from keys","First real-time generative engine for multiplayer adversarial KOF '97","Keyboard-driven two-player KOF '97 simulated at 30 FPS full match","Generative KOF '97 engine binds dual keys to real-time fight physics","Multi-player real-time KOF '97 generative model hits 30 FPS on one GPU"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"That the visual Player Association module and the gated keyboard injection reliably keep each player's controls attached to the correct fighter throughout a match, without identity swaps or control crosstalk that would break adversarial combat.","fun_headline_variants_meta":{"raw":{"variants":["WanToFight streams two-player KOF '97 combat at 30 FPS from keys","First real-time generative engine for multiplayer adversarial KOF '97","Keyboard-driven two-player KOF '97 simulated at 30 FPS full match","Generative KOF '97 engine binds dual keys to real-time fight physics","Multi-player real-time KOF '97 generative model hits 30 FPS on one GPU"]},"model":"grok-4.5","effort":"low","cost_usd":0.004372,"raw_usage":{"total_tokens":1262,"prompt_tokens":752,"num_sources_used":0,"completion_tokens":112,"cost_in_usd_ticks":43720000,"prompt_tokens_details":{"text_tokens":752,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":398,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":752,"tokens_out":112,"duration_ms":3265,"temperature":1.0,"reasoning_tokens":398,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-15T04:53:35.332151+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run a full two-player match and check whether each fighter consistently responds only to its assigned keyboard stream across punches, combos, jumps, and knockdowns; any sustained identity swap, frozen character, or crosstalk where one player's keys move the wrong fighter falsifies the multi-player claim.","supporting_citations":[],"review_version":1}