{"id":"d3f3e638-fafb-485f-9516-7a08bb4c804f","arxiv_id":"2608.13546","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Evoke combines an external camera-indexed geometry memory with a long-horizon teacher to produce a three-step world model that generates hour-scale video with bounded per-step cost.","lead":"This paper describes a video world model that keeps generating a scene for hours while the memory and compute used for each new chunk stay flat. It stores scene geometry in an external, camera-indexed memory and uses a long-horizon teacher to train a fast three-step generator.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The world-state bank's coordinate frame is built from commanded camera poses and monocular depth on generated frames; neither is validated on model outputs, and the quantitative recall test uses identical poses that can be satisfied by copying the stored frame.","rationale":"The paper is internally consistent and unusually candid about limitations, and the conditional verdict is appropriate. The reader identified depth bias on generated imagery as the weakest assumption; I agree with that concern but sharpen it: the load-bearing issue is the joint assumption that both the monocular depth and the commanded camera trajectory are accurate enough to define a consistent world-space coordinate frame on generated rollouts. The paper's own recall evaluation is weakened by using identical camera poses for the revisit PSNR comparison, which can be satisfied by copying the stored source frame even if the underlying geometry is wrong. This does not make the paper fraudulent or the architecture invalid, and the system may still work in practice; but the central external-memory benefit is not yet tested at the viewpoints where geometric alignment actually matters. Other limitations noted by the reader, such as unmatched benchmark protocols and n=8 stability sessions, are real but affect confidence in the quantitative claims rather than the conceptual viability of the approach. The proposed test would settle whether the bank is geometrically grounded on generated outputs or whether its apparent recall comes from view-indexed copying.","tokens_in":23629,"tokens_out":8694,"duration_ms":105135,"concrete_test":"Run 20 leave-and-return sessions on the released student where the camera returns to a near-but-not-identical pose (e.g., 5-10 cm translation or 2-3 degrees yaw offset from the stored source views), and compare bank-rendered conditioning against the actual generated frame. Simultaneously estimate actual camera poses from the generated RGB using a relative-pose or SLAM method and compute cross-view ICP residuals after unprojecting stored frames under the commanded P_k. If the median angular drift over the 90 s retention window exceeds about 3 degrees, if median ICP residual exceeds about 5% of scene scale, or if recall PSNR at the offset poses drops to the no-bank floor, then the geometric-memory claim needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. 3.4 constructs M_k by unprojecting monocular depth under the commanded camera trajectory, so the bank's world coordinates inherit any depth bias on generated imagery and any deviation of the actual generated motion from P_k. Both quantities are unvalidated on model outputs. The paper's only quantitative recall test (Sec. 4.4) measures revisit PSNR \"between two 12 s windows rendered at identical camera poses\"; at an identical pose the stored source frame can be copied back almost regardless of geometric correctness, so this test does not exercise depth/pose alignment at novel or nearby viewpoints. The qualitative 45-degree-turn example returns to the opening view, the same favorable case. Meanwhile the WBench Navigation and Perspective scores (78.63 and 69.74 in Table 1) indicate imperfect camera adherence, so commanded-versus-actual pose drift is not merely hypothetical. If the bank's coordinate frame is systematically misaligned, Read(M_k, P_k) returns wrong geometry exactly where the visibility mask assigns low noise, and the external-memory consistency benefit central to the paper's claim is unsupported in the general case.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"Evoke is an interactive video world model that decouples persistent scene state from the denoiser context. A camera-indexed external world state bank stores geometry obtained by unprojecting monocular depth estimates from generated chunks and renders view-relevant stored content back into the current view, while the teacher is redesigned with chunk-wise sparse attention and per-chunk text conditioning to provide roughly 30-second supervision via distribution matching under self-forced rollouts; this supervision is transferred to a three-step, CFG-free student. The paper claims that extending a session increases only the number of recurrent calls, that each 1.5 s chunk is generated in 2.11 s on one H200 at 384x640, that the method reaches state-of-the-art performance on WBench while remaining competitive on VBench-Long and VBench-2.0, and that hour-scale rollouts remain stable.","tokens_in":23852,"tokens_out":7796,"duration_ms":76264,"significance":"If the bounded-recurrence claim holds, this is a valuable systems contribution: it cleanly separates persistent geometric state from transient denoiser context, and the per-call cost bound follows by construction from fixed budgets (19-frame history, 90 s geometry retention, at most 8 retrieved sources, local positional indices). The paper is unusually explicit about the scope of its evidence: Figure 4 is labeled n=1, the long-session content descriptor is a self-consistency measure rather than an identity metric, VBench protocol deviations are declared, and the WBench row is labeled as the authors' own evaluation. The controlled short- vs long-horizon teacher ablation and the wall-clock decomposition (Fig. 10) are well designed and genuinely informative. However, the geometric-memory benefit is not exercised at novel viewpoints, the headline SOTA claim rests on a 0.1-point margin on a self-run evaluation, and the demonstrated teacher-horizon benefit is photometric only.","major_comments":[{"comment":"The quantitative recall test in Sec. 4.4 measures \"revisit PSNR between two 12 s windows rendered at identical camera poses.\" At an identical pose, the stored source frame itself can satisfy the query, so the test does not exercise alignment at nearby or novel viewpoints, which is exactly where a systematically biased monocular depth estimate (the bank is built by unprojecting depth from a monocular model, Sec. 3.4) or drift between the commanded trajectory P_k and the motion the model actually generates would return wrong geometry. No calibration of depth error on generated rollouts is provided, and the Navigation (78.63) and Perspective (69.74) scores in Table 1 indicate that camera adherence is imperfect, so pose drift is not merely hypothetical. I recommend an additional recall evaluation at nearby non-identical poses and a depth-error measurement on generated frames before the general claim of geometric recall is accepted.","section":"Sec. 3.4, Sec. 4.4"},{"comment":"The abstract's \"state-of-the-art performance on WBench\" is based on Table 7, whose footnote states \"Our own evaluation of the released student, not a leaderboard submission,\" and the margin over the second-ranked system is 0.1 in Average. The paper itself says this margin is \"of the same order as the ones Sec. 4.1 declines to read as wins.\" The SOTA claim should either be softened or accompanied by the margin and the self-run status in the abstract, since a 0.1-point lead on a self-evaluated row does not support the unqualified wording used in the headline.","section":"Abstract, Sec. 4.1, Table 7"},{"comment":"Section 3.2 motivates long-horizon supervision primarily by content drift (\"scene identity, object appearance, or spatial layout may evolve gradually while every short window remains individually plausible\"), yet the matched short- vs long-teacher comparison shows no significant separation on the content descriptor and no improvement on sharpness; the demonstrated benefit is restricted to photometric stability (Sec. 4.3, Appendix C). The paper is candid about this in the body, but the abstract's \"improving resistance to long-term drift\" and contribution (iii)'s \"strong long-horizon consistency\" are broader than the evidence. Either qualify these claims to photometric stability or provide additional evidence of content-level consistency transfer.","section":"Sec. 3.2, Sec. 4.3, Appendix C"},{"comment":"The VBench-2.0 row (66.77, leading Veo 3 by 0.05) and VBench-Long row (85.11) use one sample per prompt instead of the official five, different clip lengths, and prompt augmentation; the paper itself states that \"Single-sampling inflates variance but not the expectation, and the 0.05 margins reported in Table 2 are of that order.\" Because the rank-1 VBench-2.0 result and the \"competitive\" claim depend on margins at or below the declared protocol-induced variance, the main text should carry the same caveat that the appendix expresses, rather than presenting the leaderboard positions without it.","section":"Sec. 4.1, Appendix E, Table 2"}],"minor_comments":[{"comment":"The caption's phrase \"shading marks assertion\" is unclear; please specify what the shaded region denotes (for example, the time interval during which the clause is in force).","section":"Fig. 8"},{"comment":"The role of the detach and the normalization constant nu is described in prose, but the equation alone does not make clear which term is detached; a one-line clarification in the notation would help reproducibility.","section":"Eq. (3)"},{"comment":"The sentence \"because every third frame is ingested\" is ambiguous given the 12 depth-estimated frames per chunk and the 9 latent / 36 pixel frames; please clarify the mapping between these counts and the 2160-frame retention budget.","section":"Sec. 3.4"},{"comment":"The anchored-floor result (4% realization for anchored clauses) means the world state bank actively resists text-driven modification of stored content; this is presented as a property, but it is also an interaction limitation that should be stated explicitly in the conclusions alongside the future-work items.","section":"Sec. 4.4, Fig. 8"},{"comment":"The caption sentence \"The geometric path costs 38% of the denoiser and 93% of a recurrent step that skips it\" is confusing; rephrasing to state that geometry adds 1.84 s to an approximately 2.08 s denoiser step would make the decomposition easier to read.","section":"Fig. 10"},{"comment":"References [Wan et al., 2025a] and [Wan et al., 2025b] appear to be the same paper and should be deduplicated.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is self-aware and generally honest about its limitations, which is a strength. The main risk is that the abstract and contribution statements overclaim relative to the body: the WBench SOTA is a self-run evaluation with a 0.1-point margin, and the teacher-horizon benefit is photometric only. The identical-pose recall test is the most significant evidential gap, and it is fixable within scope by evaluating recall at nearby non-identical poses and by calibrating depth on generated rollouts. The self-referential elements (DMD from the authors' prior work, student backbone from the same group) are not by themselves disqualifying given the external benchmarks, but the editor may wish to verify the provenance of the WBench row and the VBench protocol deviations if leaderboard claims are material to the journal's acceptance criteria."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The central formulation is sound and the combination is new. Evoke's bounded recurrent process keeps per-call context independent of session length, and that claim holds by construction; the wall-clock measurements support it. The second new piece, redesigning the teacher for long-horizon supervision with per-chunk conditioning and transferring it via a 30-second distribution-matching objective, is also well motivated. The controlled teacher-horizon comparison, where the long-horizon teacher improves photometric stability in 7 of 8 clips (Wilcoxon p=0.016), is a real result.\n\nThe authors also deserve credit for disciplined limitations. They distinguish degradation drift from content drift, restrict the long-session claim to \"no runaway degradation\" rather than identity preservation, disclose protocol deviations, and run ablations on supervision horizon and anchoring. The cost decomposition is genuinely useful.\n\nThe weakest point is the world state bank's coordinate frame. Section 3.4 builds it by unprojecting monocular depth under commanded camera poses on generated frames, and neither the depth estimator's behavior on generated imagery nor the actual generated camera motion is validated. The WBench Navigation and Perspective scores (78.63 and 69.74) show imperfect camera adherence, so commanded-versus-actual pose drift is not hypothetical. The quantitative recall test measures PSNR at identical camera poses, where a stored frame can be copied back almost regardless of geometric correctness; it does not exercise novel or nearby viewpoints. So the general external-memory consistency benefit is not as well supported as the bounded-cost claim. This is a genuine gap, but not a load-bearing error, because the paper's main claims are about bounded cost and photometric stability, and the geometry bank is one component.\n\nThe long-session evidence is also limited: eight rollouts, mostly photometric, with the paper itself noting that the content descriptor does not significantly separate the teacher-horizon variants. The benchmark comparisons are step-unmatched, include self-run leaderboard rows, and the margins are of order 0.05, which the authors explicitly say not to over-read. No code commit hash or full training artifacts are provided, so independent reproduction is not yet possible.\n\nThis paper is for researchers working on interactive world models, long-video generation, and distillation. It deserves a serious referee. I would accept it for review and ask for depth calibration on generated rollouts, a novel-view or nearby-pose recall test, and more long-session runs if feasible. I would cite it for the bounded recurrent formulation and the teacher-side supervision-horizon result.","headline":"A well-built bounded-recurrent world model with a genuinely new external-geometry design, but the memory-consistency evidence is thinner than the architecture claims and the depth/pose calibration on generated rollouts is the load-bearing soft spot.","tokens_in":24380,"tokens_out":2131,"would_cite":true,"duration_ms":25018,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Evoke recasts interactive video generation as a bounded recurrent process: persistent scene geometry lives in a camera-indexed world state bank, and a 31.4-second teacher supervision budget transfers long-horizon stability and timed text…","keywords":["interactive world models","world state bank","long-horizon supervision","distribution matching distillation","sparse attention","few-step generation","camera control","external geometric memory"],"falsifier":"Take a synthetic scene with ground-truth depth, run a leave-and-return camera trajectory, and compare the bank-rendered warp with the true geometry under the same camera pose. If revisit PSNR or geometric alignment does not rise once the retention window covers the time away, or if depth bias causes systematic misalignment on generated imagery, the claim that external geometric memory provides persistent recall fails. A second check is whether per-chunk latency on a single H200 stays at the reported 2.11 s after many thousands of chunks with the bank at full retention.","tokens_in":23442,"feed_emoji":"🎬","tokens_out":9803,"duration_ms":91972,"temperature":0.7,"pith_summary":"The paper aims to remove the two bottlenecks that keep interactive video world models from running for hours: the denoiser context must not grow with session history, and a few-step student must inherit capabilities its teacher never showed it. Evoke's first move is to externalize persistent state: generated frames are unprojected into a camera-indexed world state bank, and each recurrent step reads back only the geometry visible from the current camera pose, so per-chunk context, positional span, and compute stay fixed as the session lengthens. Its second move is to treat the teacher as part of the design: chunk-wise sparse attention makes long-window scoring affordable, per-chunk text conditions put prompt changes inside a single rollout, and a 30-second distribution-matching objective under self-forced rollouts transfers long-horizon photometric stability and timed text control to a three-step student that runs without classifier-free guidance. If the claims hold, interactive generation becomes a bounded recurrence: long sessions cost more calls, not bigger calls, and the reported 2.11 s per 1.5 s chunk at 384x640 on one H200 is a latency, not a limit.","feed_headline":"Hour-long interactive video worlds with no growing context","feed_subtitle":"A camera-indexed geometry bank and 30-second teacher supervision let a 3-step model generate 1.5 s chunks in 2.11 s.","key_machinery":"The load-bearing object is the bounded recurrent loop formed by the student and the external world state bank. The bank is built by taking monocular depth estimates for twelve frames of each generated chunk and unprojecting them with the known camera intrinsics and extrinsics; a read ranks stored views by co-visibility with the target camera, renders up to eight sources with z-buffering, and returns a view-aligned warp plus a per-pixel visibility mask. That mask decides how much geometry enters the denoiser: unsupported regions are assigned full noise and pruned from the token sequence, so the memory contributes exactly where stored coverage exists and nowhere else. Making long-horizon supervision affordable is the teacher's chunk-wise sparse attention, where each chunk sees a first-frame sink, local context with one-frame overlap, compressed nearby frames, a few selected distant frames, and a linear-attention global state, keeping attention cost approximately linear in sequence length. The 30-second distribution-matching objective is applied to the full self-forced rollout with per-chunk detached backward graphs, so the supervision horizon exceeds the gradient horizon and long-range signals reach the student without backpropagating through the whole trajectory.","core_discovery":"Evoke's central claim is that persistent memory and interactive latency do not have to be carried by the same mechanism. Scene geometry is held in an explicit, camera-indexed world state bank; at each step the bank is read by rendering stored surfaces into the current view, the render is injected at the coarsest stage of a three-step pyramid denoiser alongside a short local history, and the emitted chunk is written back to the bank with a bounded retention budget. The paper's formal statement is that extending a session increases only the number of recurrent calls, without increasing the context length, positional span, or computational footprint of an individual call. To give this bounded student the long-horizon behavior it needs, the teacher is redesigned for supervision: chunk-wise sparse attention with bounded local, distant, and linear-attention sources makes a 31.4-second full-window objective practical, per-chunk conditioning makes mid-session prompt changes expressible, and distribution matching over the complete self-forced rollout transfers those capabilities. The reported experimental outcome is a three-step, CFG-free model that leads WBench navigation, stays competitive on VBench-Long and VBench-2.0, keeps photometric statistics flat over 65.5-minute rollouts, and revisits stored geometry with 2.3-3.2 dB higher PSNR once retention covers the time away.","pith_inferences":["One implication the paper leaves implicit: if the 30-second supervision horizon saturates, the same teacher recipe could supervise much longer rollouts without recomputing full-trajectory objectives; the paper's own window-length sweep suggests diminishing returns once the relevant temporal scale is covered, and an explicit test would measure drift on hour-scale rollouts after training with differ","The world state bank's fidelity is only as good as the monocular depth used to build it; a testable extension is to feed the depth estimator's calibrated uncertainty into the visibility mask so unreliable geometry is suppressed rather than injected, which could improve revisit consistency on synthetic or stylized content where depth priors are weakest.","The evocation asymmetry, where text can add unanchored content but not overwrite anchored geometry, suggests a practical authoring rule: introduce controllable elements before their anchoring surfaces are generated, or extend the bank with an explicit erase-and-rewrite channel that the paper does not claim.","Connecting to embodied navigation, camera-pose-indexed geometry could serve as a general episodic memory for agents that must re-identify locations; Evoke demonstrates the mechanism for generation, and the same read-write pattern could be evaluated on navigation metrics."],"forward_implications":["If the bounded-recurrence claim holds, session length can grow for hours without changing the per-chunk cost, so open-ended interactive generation becomes a repeated fixed-size call rather than a context-management problem.","The three-step, CFG-free student inherits long-horizon photometric stability and mid-session prompt responsiveness from the teacher, meaning low latency need not cap long-term consistency.","Because the world state bank is read through camera pose, revisiting a location restores stored geometry instead of relying on the denoiser to reinvent it; the paper reports 2.3-3.2 dB higher revisit PSNR once retention covers the time away.","Per-chunk text conditioning introduces a timed evocation capability: new elements can appear in unanchored space, with a 67% realization rate in the controlled test, while geometry-supported content resists overwriting at 4%, so users can stage events without destabilizing the scene."],"supporting_citations":[{"why":"Supplies the view-aligned warp-conditioning mechanism that Evoke extends from short-range frame correspondence into persistent camera-indexed geometric memory.","marker":"Wang and He, 2026"},{"why":"Provides the monocular depth estimates that are unprojected with camera intrinsics and extrinsics to build the world state bank.","marker":"Lin et al., 2025a"},{"why":"Introduces distribution-matching distillation, the objective Evoke extends to a 30-second full-window score over self-forced rollouts.","marker":"Yin et al., 2024b"},{"why":"Provides self-forced rollout training and the train-test gap framing used for long-horizon student supervision.","marker":"Huang et al., 2025a"},{"why":"Supplies chunk-wise sparse attention that keeps teacher scoring cost linear and makes 31.4-second supervision windows affordable.","marker":"Jia et al., 2025"},{"why":"Documents the windowing and eviction trade-offs and the long-video objectives that motivate externalizing persistent world state.","marker":"Cui et al., 2025"},{"why":"Analyzes the teacher-student attention mismatch that Evoke extends by redesigning supervision horizon and conditioning schedule.","marker":"Zhu et al., 2026"},{"why":"Provides the linear-attention global-state component of the teacher's sparse attention pattern.","marker":"Yang et al., 2023"},{"why":"Serves as the base 14B diffusion transformer from which the long-horizon Evoke Teacher is adapted.","marker":"Wan et al., 2025a"}],"fun_headline_variants":["Endless worlds from a bounded context denoiser","External geometry bank, fixed-cost video generation","3-step world model: endless sessions, no context growth","Unbounded video sessions at constant per-chunk cost","World model with external memory, no context bloat"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the monocular depth estimates used to build the world state bank are accurate enough that geometry rendered on a return visit aligns with what the session actually generated; if depth is systematically biased on synthetic or generated imagery, the external-memory consistency benefit fails.","fun_headline_variants_meta":{"raw":{"variants":["Endless worlds from a bounded context denoiser","External geometry bank, fixed-cost video generation","3-step world model: endless sessions, no context growth","Unbounded video sessions at constant per-chunk cost","World model with external memory, no context bloat"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000401,"raw_usage":{"total_tokens":2195,"prompt_tokens":1150,"completion_tokens":1045,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":766,"completion_tokens_details":{"reasoning_tokens":969}},"tokens_in":766,"tokens_out":1045,"duration_ms":9206,"temperature":1.0,"reasoning_tokens":969,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:24:53.656316+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a synthetic scene with ground-truth depth, run a leave-and-return camera trajectory, and compare the bank-rendered warp with the true geometry under the same camera pose. If revisit PSNR or geometric alignment does not rise once the retention window covers the time away, or if depth bias causes systematic misalignment on generated imagery, the claim that external geometric memory provides persistent recall fails. A second check is whether per-chunk latency on a single H200 stays at the reported 2.11 s after many thousands of chunks with the bank at full retention.","supporting_citations":[],"review_version":1}