{"id":"e22cfb36-9932-41ee-8a28-87b15de56bef","arxiv_id":"2608.08600","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Khora decouples world-state evolution from visual rendering through a shared STBoard and fixed-dimensional per-view renderers, enabling inference-time addition and removal of agents without retraining.","lead":"This paper introduces Khora, a multi-agent world model that keeps a shared world-state board and renders each agent's view independently, so new agents can join or leave a simulation without retraining. A generalist reader may care because it targets the population-scaling bottleneck in interactive simulation for games, robotics, and embodied AI, claiming near-linear cost growth with the number of requested views.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Even if the STBoard is state-sufficient, Eq. (9) describes independently sampled per-view rendering, so cross-view consistency at the pixel/appearance level is not actually anchored by the shared state.","rationale":"The reader's weakest assumption is state sufficiency: whether STBoard contains all information needed for the per-view renderer to produce consistent observations. I agree this is the right locus, but the sharper technical issue is that even granting state sufficiency, Eq. (9) uses a stochastic renderer with independent per-view queries, so the shared state anchors the distribution of observations but not individual samples. The paper's consistency evaluation is semantic and matches exactly the semantic channels of the rasterized condition map; it does not test whether the rendered texture/appearance of, say, a newly inserted agent is the same across two viewpoints. This does not invalidate the architecture or the scalability claims, which are plausible, but it makes the headline claim about cross-view consistency broader than the evidence. The reader's CONDITIONAL verdict is appropriate: the paper needs either a mechanism that shares stochastic degrees of freedom across views or a quantitative cross-view appearance consistency evaluation. My check would settle whether the concern lands; until then, the conditionality should remain.","tokens_in":9254,"tokens_out":6700,"duration_ms":83028,"concrete_test":"Run a held-out multi-view trajectory through Khora under two sampling protocols: independent per-view diffusion noise (as implied by \"independent queries\") versus a single shared noise/latent across views. Using the STBoard's predicted poses and depth maps, warp one generated view into another and compute LPIPS/SSIM in the overlapping regions that observe the same 3D surface. Compare the two protocols and also compare against Khora w/o Cross-Agent State. If independent per-view noise is significantly worse than shared noise, then Eq. (9) does not anchor cross-view appearance consistency and the central claim must be weakened to semantic consistency; if the two protocols match, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim turns on Eq. (9): o_{t+1}^i = R(S_{t+1}, C_{t+1}^i, p_{t+1}^i; G), with the text stating that \"additional views are generated by issuing additional independent queries\" (Section 4.3). R is described as a latent diffusion renderer, i.e., a stochastic map, and nothing in Eq. (9) shares the sampled noise or latent across the N_v queries. Each view receives the same S and G, but the stochastic realization of appearance is not coupled across views. The rasterized conditioning map C_i carries only occupancy, depth, orientation, identity, and task-dependent state, which are exactly the semantic quantities the paper's consistency evaluation targets. The shared STBoard therefore anchors geometry and entity identity, but it does not determine the rendered texture or fine appearance; two synchronized views of the same surface can receive different diffusion samples. The paper's own consistency evidence, Table 3, is a semantic user study of identity, relative position, and persistence, which is precisely what C_i encodes; it would not detect texture drift or appearance divergence across views. The text promises 95% confidence intervals for the user study but none are reported. Thus the advertised \"maintaining cross-view consistency\" is overclaimed: the demonstrated consistency is at the level of the rasterized conditioning, not necessarily of the rendered observations. The same issue applies to agents inserted at inference time: their appearance is initialized once, but then each requesting view independently samples its realization, so even a sufficient STBoard does not guarantee consistent appearance across viewers.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"Khora is a multi-agent world model whose central design choice is to decouple world-state evolution from per-view rendering. A shared STBoard (static scene memory plus a dynamic entity table in world coordinates) is updated by an action-conditioned transition model that combines a kinematic proposal with a learned residual (Eqs. 4–6). Each requested observation is decoded by a single shared latent-diffusion renderer from the shared state, the target pose, and a rasterized fixed-dimensional condition map encoding occupancy, depth, orientation, identity, and task state (Eqs. 7–9). Because the rendering interface has fixed dimensionality, the authors claim inference-time population expansion without retraining and a practical cost of T ≈ N_v C_render + N_a N_v C_proj (Eq. 10). The experimental section reports an internal ablation (Khora versus Khora w/o Cross-Agent State) on PSNR/SSIM/LPIPS/FID/FVD, a five-point-scale user study of action and world consistency, runtime measurements up to 80 agents, and a figure-based demonstration of agents joining and leaving a rollout at inference time. The paper's own Limitations section concedes that the static prior is map-specific and that the evaluation is primarily qualitative.","tokens_in":9500,"tokens_out":9904,"duration_ms":97136,"significance":"The core formulation is genuinely useful if the claims hold: replacing dense cross-view interaction inside the generator with a fixed-dimensional query interface over a shared state is a clean architectural idea that makes the population-dependence of the model explicit and exposes a crisp, falsifiable cost model (Eq. 10). The insertion/removal protocol (initialize an entity, append it to the dynamic table, issue an extra query) is concrete and mechanically consistent with the architecture. The ablation that masks other agents' dynamic entries is an appropriate control for isolating the contribution of cross-agent state, and the authors avoid presenting fitted constants as predictions. Credit is also due for an unusually honest Limitations section. In its current form, however, the evidence supports the qualitative and architectural claims rather than the quantitative ones: there are no external baselines, the consistency study omits the promised confidence intervals, and the dynamic-population demonstration is illustrative only.","major_comments":[{"comment":"The paper's core consistency claim — that cross-view agreement is anchored by the shared STBoard — is established only at the level of the rasterized conditioning, not at the level of the rendered observation. Section 4.3 states that 'additional views are generated by issuing additional independent queries'; since R is a stochastic latent-diffusion renderer, independent queries draw independent noise/latent realizations, so two synchronized views of the same surface can produce different texture and fine appearance even with identical S, C, and p. The channels of C_i (occupancy, depth, orientation, identity, task state) are exactly the semantic quantities tested in the Table 3 user study, which therefore cannot detect appearance-level divergence. To support the advertised claim of maintained cross-view consistency, the authors should either couple the stochastic realization across the N_v queries (for example, shared noise or shared latent) or provide a pixel- or patch-level cross-view consistency measurement, such as perceptual similarity between corresponding image regions of synchronized rendered views.","section":"§4.3, Eq. (9)"},{"comment":"The text in Section 5.2 promises that 'we report the mean score together with a 95% confidence interval,' but Table 3 reports only means. Given that the Khora-versus-ablation differences are 1.495 points on action consistency and 0.822 points on overall score, the reader cannot tell whether these differences are meaningful without intervals, participant counts, the number of rated clips, and a description of the rater pool. Because Table 3 is the only quantitative evidence for the cross-view consistency claim, it must be reported completely and with the promised intervals.","section":"§5.2, Table 3"},{"comment":"All quantitative evaluation is against the authors' own ablation (Khora w/o Cross-Agent State); no existing multi-agent world model (Solaris, MultiWorld, Gamma-World, MultiGen, Agora-1) is measured. Table 1 assigns asymptotic scaling labels (O(N²), O(N), ≈O(N)) to other methods without citing a cost analysis or a measurement, so the claim that Khora's population-time scaling improves on prior work is asserted rather than demonstrated. Even a single measured runtime comparison, or a cited per-method complexity argument, would materially strengthen the scalability contribution.","section":"§5.1–§5.3, Tables 1–2"},{"comment":"The '≈O(N)' label for Khora in Table 1 is in tension with Eq. (10): when every agent requests a view (N_v = N_a), the projection term alone is O(N_a N_v) = O(N²), and the paper's own Limitations section acknowledges the worst-case O(N_a N_v) geometric cost. The linear-scaling claim is defensible only for renderer evaluations (O(N_v), with C_render ≫ C_proj) and in the practical regime where the number of views per agent is small. The abstract and Table 1 should state the regime in which 'approximately linear' holds; as written, the headline scaling claim is stronger than the stated cost model.","section":"Table 1; §4.3, Eq. (10)"},{"comment":"The headline capability — that agents can join and leave a single autoregressive rollout without retraining — is demonstrated only by illustrative frames. The text asserts that 'existing agents preserve their trajectories and observations' when new agents are inserted, but no quantitative evidence is given (for example, trajectory error of existing agents before and after insertion, or appearance and identity similarity of the inserted agent to its initialization). The authors' Limitations section concedes that the evaluation 'is primarily qualitative'; that concession applies directly to the dynamic-population experiment, which is the paper's most distinctive result. A small quantitative protocol, such as population-varying rollouts with measured identity persistence and trajectory drift, would substantially raise confidence in the central claim.","section":"§5.4, Fig. 6; §6"}],"minor_comments":[{"comment":"The section heading 'Priliminary' should read 'Preliminary.'","section":"§3"},{"comment":"Equation (3) presents the rendering interface as R(S_{t+1}, {p^i}_{i∈I_t}, G) without the target-view condition map C^i that appears in Eq. (9); the two formulations should be reconciled so the preliminaries match the method.","section":"§3, Eq. (3); §4.3, Eq. (9)"},{"comment":"The 8-view row reports only 159 cases (1,272 videos) versus 10,000 cases for the 2- and 4-view rows, and the reason for this drop is unexplained. FVD on roughly 159 clips is statistically unstable; the authors should report the full 8-view evaluation or state the constraint that produced the smaller sample.","section":"Table 2"},{"comment":"The relationship between 'views' in Fig. 3 and 'agents' in Fig. 5 is not defined; please clarify whether one view per agent is assumed and how the dashed 'projected' curves are extrapolated.","section":"Fig. 3, §5.3"},{"comment":"Several 2026 citations (e.g., Gamma-World, MultiGen, Solaris) appear only as arXiv URLs with broken line-wrapping in the reference list; please unify the citation format. There are also spacing typos in the author list ('Y ong-Lu Li', 'V olhejn').","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is best understood as a system/position paper with a strong architectural claim; its evaluation depth is below what the abstract's language ('demonstrate', 'generalizes') implies, and the authors themselves concede that the evaluation is primarily qualitative. I do not see grounds for rejection: the ideas are coherent, the fixed-dimensional query interface is a real contribution, and the identified gaps (appearance-level consistency or coupling, external baselines, confidence intervals, and a quantitative dynamic-population protocol) are all addressable within the paper's scope. My main residual concern is that the novelty and advantage claims over prior work rest on Table 1's unmeasured labels rather than on empirical comparison; the authors should be asked to verify their characterization of the other methods' interfaces and scaling behavior against the cited papers directly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is a real architectural idea, not just a rehash of MultiGen or Gamma-World. The STBoard shared-state plus fixed-dimensional per-view rendering interface genuinely lets you add and remove agents mid-rollout without retraining, and the cost model in Eq. 10 is plausible. It deserves a serious read.\n\nThe genuinely new piece is the decoupling: world-state evolution is population-agnostic, agents are entries in a dynamic table rather than fixed view slots, and every view is an independent query against the same state. The dynamic-population rollout (2→4→2) is a useful proof-of-concept, and the limitations section is honest about the map-specific prior and qualitative evaluation. The writing is clear and the conceptual framing is sound.\n\nThe soft spots are real, and they are mostly evidential. Table 2 compares only against an internal ablation, not against Solaris, MultiGen, MultiWorld, or Gamma-World. Table 3 is a user study that promises 95% confidence intervals but never reports them. The dynamic-population result is figure-based, not quantified. No code or data release. These are fixable, but they are not minor.\n\nThe stress-test about cross-view consistency is correct, and it lands on reading the paper. Eq. (9) renders each view with an independent stochastic query conditioned on the shared state and the rasterized condition map. The shared state anchors geometry, identity, and semantic state—exactly what C_i encodes—but it does not couple the diffusion sample across views. Fine texture and appearance can therefore drift between two synchronized views of the same surface. The user study tests precisely what C_i encodes (identity, relative position, persistence) and would not catch that drift. So the advertised \"cross-view consistency\" should be tempered to \"semantic and geometric consistency,\" or the renderer should share noise/latent across views. This is a genuine overclaim, not a nitpick.\n\nWho this is for: people working on multi-agent simulation, interactive video generation, or embodied AI platforms. It is a systems/architecture paper with a clean idea, not a benchmark-topping empirical study. I would bring it to a reading group if you are in that area.\n\nRecommendation: send it to peer review. The architecture deserves referee time, but the authors need external baselines, the promised confidence intervals, a quantitative dynamic-population evaluation, and a sharper statement of what consistency they actually guarantee. I would not accept it as-is.","headline":"Genuinely new architecture for population-scalable multi-agent world models, but the cross-view consistency claim is stronger than the stochastic per-view renderer and the evidence actually support.","tokens_in":10092,"tokens_out":1721,"would_cite":true,"duration_ms":19539,"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 proposes Khora, a multi-agent world model that maintains cross-view consistency through a shared world state and thereby supports adding or removing agents at inference time without retraining or architectural changes.","keywords":["multi-agent world models","population scalability","shared world state","cross-view consistency","neural rendering","inference-time generalization","interactive simulation","STBoard"],"falsifier":"Take a trained Khora checkpoint that has only seen two or four agents, run rollouts with 8, 16, 32, and 64 agents in a scene with repeated occlusion, and measure whether two synchronized views agree on an occluded agent's identity and position after reappearance; growing cross-view disagreement would show the STBoard does not carry enough state. A second decisive check is to measure per-view latency while doubling population and compare it with Eq. (10); superlinear growth before 80 agents would mean the renderer still does population-dependent work.","tokens_in":9037,"feed_emoji":"🎮","tokens_out":4581,"duration_ms":50992,"temperature":0.7,"pith_summary":"The paper's central claim is that multi-agent world models can scale to arbitrary, changing agent populations at inference time if cross-view consistency is anchored in a shared world state rather than in interactions inside the video generator. Khora maintains a Spatio-temporal Board (STBoard) that records static scene memory and dynamic entities in a common coordinate system, evolves it from agent actions, and renders each requested view independently from a fixed-dimensional rasterized condition map. Because the renderer's input size does not depend on the number of agents, adding an agent is just adding a state entry and issuing one more rendering query. The authors report that this preserves per-view fidelity and cross-view consistency while keeping the dominant rendering cost approximately linear in the number of requested views.","feed_headline":"World model adds agents at runtime without retraining","feed_subtitle":"Each view renders independently from one shared board, so growing the population is just one more query.","key_machinery":"The central object is the STBoard, a world-centric shared state consisting of a static scene memory and a dynamically sized entity table in a common coordinate system. It carries the argument because both the transition model and every view query read from this same board: the renderer conditions each observation only on the updated state $S_{t+1}$, a fixed-shape rasterized condition map $C^i_{t+1}$, the target pose $p^i_{t+1}$, and the static prior $G$, as expressed in Eq. (9). This fixed-dimensional rendering interface is what makes population growth an inference-time operation rather than an architectural change, and it yields the cost expression $T_{\\text{Khora}} \\approx N_v C_{\\text{render}} + N_a N_v C_{\\text{proj}}$ in Eq. (10), where the neural rendering term dominates and scales only with the number of views $N_v$.","core_discovery":"Khora's central discovery is a decoupling: world-state evolution and visual rendering are treated as separate stages, with consistency coming from a shared STBoard rather than from dense communication among observation streams inside the expensive video generator. The STBoard holds a static scene memory plus a dynamic entity table in world coordinates; an action-conditioned transition model updates agent poses and entity states, and a geometry-guided renderer projects every possibly visible entity into the target camera frame and rasterizes it into a condition map of fixed spatial shape. The same renderer is invoked once per requested view, so its architecture has no dependence on agent count. The paper reports that this supports changing the active population within one rollout from two agents to four to eight and back down, and that measured latency rises only modestly from 107.16 ms to 116.73 ms per rollout step as the population grows from 1 to 80 agents, with aggregate throughput reaching 2741.3 view-fps when views are distributed across GPUs.","pith_inferences":["Beyond the paper, the shared-state decomposition suggests that consistency could be measured directly: pairwise view disagreement should be independent of the number of active agents, a property the current qualitative evaluation does not fully quantify.","The paper leaves implicit that state sufficiency is the real testable hypothesis: if any visual detail needed for a correct view lives outside the STBoard's static and dynamic entries, the renderer cannot recover it, so a future version could augment the board with learned appearance tokens while keeping the fixed per-view interface.","A testable extension would be to insert agents whose initial observation is partial or occluded and check whether the board still produces consistent identity and positioning; the paper initializes agents from complete first observations, so this harder case remains open.","The architecture implies a benchmark for population-time cost: measuring per-view latency while doubling population should follow Eq. (10) closely, and deviation would reveal hidden population-dependent work in the renderer."],"forward_implications":["New agents can enter or leave a live rollout by appending or removing an entity row in the STBoard and issuing an additional rendering query, with no retraining and no change to the transition or rendering architecture.","The neural renderer's cost scales approximately linearly with the number of requested views rather than quadratically with pairwise agent interactions, so the expensive part of generation does not grow with population.","Distribution of views across GPUs keeps per-GPU memory roughly constant, so aggregate rendering throughput increases as the agent population grows.","Cross-view agreement is anchored in a shared evolving state, so synchronized views describe the same world even when agents are added or removed mid-rollout; the paper demonstrates this with dynamic-population experiments.","The same shared-state and rendering interface is usable across different environments by swapping the scene-specific static memory, making the architecture map-adaptive rather than population-adaptive."],"supporting_citations":[{"why":"Establishes the two-player cross-attention approach whose dense observation-stream interactions Khora is designed to avoid.","marker":"[Savva et al., 2026]"},{"why":"Introduces external memory for editable multiplayer worlds while retaining predefined view layouts, serving as a fixed-population baseline.","marker":"[Po et al., 2026]"},{"why":"Provides a scalable multi-view synthesis baseline with a global state encoder that Khora compares against on multi-agent consistency.","marker":"[Wu et al., 2026]"},{"why":"Uses permutation-symmetric encodings and sparse hub attention to reduce pairwise cost, the closest scalable multi-player baseline.","marker":"[Liu et al., 2026]"},{"why":"Demonstrates real-time multi-agent interaction in a shared generated world, providing the interactive-system comparison point.","marker":"[Odyssey, 2026]"}],"fun_headline_variants":["Add agents on the fly, no retraining needed","Shared world state enables infinite agent scaling","Khora scales to any agent count at inference time","Runtime population expansion without retraining","One shared board, unlimited agent views"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"State sufficiency is the load-bearing premise: the STBoard's static scene memory plus dynamic entity table must contain every factor needed to render correct and mutually consistent observations, and if any visual or dynamic detail lives outside that structured state, synchronized views can diverge.","fun_headline_variants_meta":{"raw":{"variants":["Add agents on the fly, no retraining needed","Shared world state enables infinite agent scaling","Khora scales to any agent count at inference time","Runtime population expansion without retraining","One shared board, unlimited agent views"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000207,"raw_usage":{"total_tokens":1395,"prompt_tokens":937,"completion_tokens":458,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":392}},"tokens_in":553,"tokens_out":458,"duration_ms":4663,"temperature":1.0,"reasoning_tokens":392,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:29:25.206335+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a trained Khora checkpoint that has only seen two or four agents, run rollouts with 8, 16, 32, and 64 agents in a scene with repeated occlusion, and measure whether two synchronized views agree on an occluded agent's identity and position after reappearance; growing cross-view disagreement would show the STBoard does not carry enough state. A second decisive check is to measure per-view latency while doubling population and compare it with Eq. (10); superlinear growth before 80 agents would mean the renderer still does population-dependent work.","supporting_citations":[],"review_version":1}