{"id":"a0e2ae5d-83d5-444a-9da1-8931b38623a4","arxiv_id":"2508.12087","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"MAPF-World, an autoregressive action world model that predicts future states and actions, is claimed to beat state-of-the-art learnable MAPF solvers while using 96.5% fewer parameters and 92% less data.","lead":"An abstract describes MAPF-World, a learning-based system that predicts future states and actions to guide multi-agent path finding. A generalist might read it because better coordination with a far smaller model and less data could make robot fleets cheaper and more practical.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The supplied full text is a different paper (VimoRAG), so MAPF-World's central claims of superior generalization and 96.5%/92% efficiency gains cannot be verified.","rationale":"The reader's weakest_assumption focuses on whether MAPF-World's predicted future states are faithful to true environment dynamics. That is a scientifically substantive concern that would matter if the MAPF-World manuscript were available. My read, however, identifies a more fundamental issue: the provided full text is a different paper (VimoRAG), so the manuscript's own content explicitly shows that the MAPF-World claims are not accompanied by any supporting method, experiments, or results. I partially agree with the reader because the reader also flags the full-text mismatch in the red flags section, but the reader's formal weakest_assumption is about the world-model's faithfulness rather than the absence of the manuscript itself. I do not allege any misconduct; the concern is purely that the artifact under review does not contain the claimed contribution. No internal consistency check, baseline audit, or reproduction attempt can be run on MAPF-World from the supplied text. The paper may or may not be correct; it is simply unverifiable. Therefore the appropriate disposition remains the reader's UNVERDICTED verdict, and I recommend no change.","tokens_in":20545,"tokens_out":2807,"duration_ms":27064,"concrete_test":"Download the official arXiv PDF for identifier 2508.12087 directly from arxiv.org and check whether its title and content match the MAPF-World abstract. If it matches VimoRAG, the submission is a metadata mismatch and no scientific verdict on MAPF-World is possible until the correct manuscript is supplied. If it matches MAPF-World, then inspect the experimental section: verify that the learnable baselines are evaluated on the same MAPF benchmark suite, confirm the out-of-distribution test set is genuinely disjoint from the automatic map generator's training distribution, and recompute the 96.5% model-size reduction and 92% data reduction from the reported parameter counts and dataset sizes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that MAPF-World outperforms state-of-the-art learnable MAPF solvers with superior zero-shot generalization, a 96.5% smaller model, and 92% reduced data. For that claim to hold, the manuscript must describe the world-model architecture, the autoregressive future-state and action predictions, the automatic map generator, benchmark protocols, and baseline comparisons. None of that is present in the supplied full text. Instead, the full text is VimoRAG, an entirely different paper about video-based retrieval-augmented 3D motion generation, with different title, authors, and subject matter. This is not a subtle hidden assumption inside an otherwise complete argument; it is the complete absence of the artifact needed to evaluate the argument. The self-description of the manuscript as VimoRAG means the MAPF-World abstract's statements are supported by no derivations, experiments, equations, or tables in the provided material. Because the claimed method and evidence are absent, the central claim can be neither confirmed nor internally cross-checked. The correct scientific status is unverified rather than rejected on the merits.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript as submitted consists of an abstract announcing MAPF-World, an autoregressive action world model for multi-agent path finding, followed by a full text that is an unrelated paper titled \"VimoRAG: Video-based Retrieval-augmented 3D Motion Generation for Motion Language Models\" by a different set of authors. The abstract claims that MAPF-World outperforms state-of-the-art learnable MAPF solvers with superior zero-shot generalization, a 96.5% smaller model, and 92% less training data. However, the body contains no description of the MAPF-World architecture, its world-model prediction mechanism, the automatic map generator, training data, baselines, or evaluation protocol; all sections, equations, tables, and experiments belong to VimoRAG. The central claims are therefore entirely unsupported by the provided manuscript text.","tokens_in":20702,"tokens_out":4829,"duration_ms":47126,"significance":"If the MAPF-World results were actually established, the paper would be significant: it would demonstrate that world-model-based lookahead improves decentralized MAPF decision-making at dramatically lower model and data cost, and it would introduce a grounded map generator for more realistic benchmarks. The current submission, however, provides no method or evidence for any of these claims. There are no parameter-free derivations, machine-checked proofs, reproducible code, or falsifiable predictions pertaining to MAPF in the manuscript; the only concrete experimental content concerns 3D motion generation and is irrelevant to the stated topic. The claim of significance is therefore unverifiable from the submitted artifact.","major_comments":[{"comment":"The full text of the submission is the VimoRAG paper, not the MAPF-World paper promised by the title and abstract. No section or equation describes the MAPF-World model, its autoregressive future-state/action prediction, the automatic map generator, the training protocol, or the MAPF benchmarks. Consequently, the abstract's central claims of superiority over state-of-the-art learnable solvers, zero-shot out-of-distribution generalization, and 96.5%/92% efficiency gains are completely unverifiable from the submitted text. This is not a local gap but the absence of the entire artifact needed to evaluate the contribution.","section":"Abstract, first/last paragraphs"},{"comment":"The abstract asserts \"Extensive experiments demonstrate that MAPF-World outperforms state-of-the-art learnable solvers\" and reports specific efficiency numbers, but the manuscript contains no MAPF experiments, no baseline comparisons, no dataset descriptions, and no evaluation metrics. The provided Tables 1–8 and Figures 1–13 all belong to VimoRAG and concern text-to-motion generation, so no numerical evidence bears on the MAPF claims. The paper therefore fails to satisfy the basic requirement that its central empirical assertions be accompanied by methods and results.","section":"Abstract, first/last paragraphs"},{"comment":"The proposal hinges on \"explicitly modeling environmental dynamics... through future state and actions prediction,\" but the manuscript provides no definition of the prediction target, no loss function, and no evidence that the predicted futures are grounded in actual environment transitions as opposed to generated by the decision policy itself. Without such material, the claimed mechanism of \"informed, coordinated, and far-sighted decision-making\" cannot be assessed even in principle.","section":"Abstract, paragraph 2"}],"minor_comments":[{"comment":"The title, author list, and abstract of the manuscript do not match the title, authors, or subject matter of the body; this suggests a submission error that must be corrected before any review can proceed.","section":"Title and metadata"},{"comment":"The abstract cites MAPF-related concepts and benchmarks, but the reference list contains only citations for motion generation and video retrieval; no MAPF-related literature is cited anywhere in the manuscript.","section":"References"},{"comment":"The Limitations subsection addresses latency of video-augmented motion generation and is unrelated to the MAPF-World claims; it does not acknowledge the absence of experimental validation for the abstract's assertions.","section":"Section 5 (Discussion)"}],"recommendation":"reject","confidential_remarks":"I strongly suspect an upload or compilation error, because the body is a complete, unrelated paper. As submitted, the manuscript cannot be reviewed on the merits. Rather than treating this as a technical rejection of the underlying idea, the editor may wish to invite the authors to resubmit the correct MAPF-World manuscript; however, under the current submission's text, the central claims are entirely unsupported and the appropriate editorial disposition is reject."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Plainly: the claim you'd be reviewing is not in the submission. The abstract describes MAPF-World, a path-finding world model trained with 96.5% fewer parameters and 92% less data, but the full text is VimoRAG, a paper about retrieval-augmented 3D motion generation with different authors, title, and arXiv number. That is not a missing proof or a hidden assumption; it is the absence of the artifact being claimed.\n\nCredit where it is earned: the abstract's idea is a plausible and worthwhile direction. Autoregressive world models are established in RL and imitation learning, and applying them to MAPF with future state and action prediction is a sensible extension. The efficiency numbers, if supported, would make learned MAPF noticeably more practical for logistics and multi-robot fleets. None of that can be checked in the supplied text. The VimoRAG material that is actually present has the normal furniture of a reviewable paper — experiments, ablations, latency analysis, a limitation statement, and public resources — but it is not the MAPF paper, and no equation or table in it touches the MAPF claims.\n\nThe soft spot is disqualifying: the abstract and the body are irreconcilable. There is no architecture, no baseline comparison, no map generator, no training protocol for MAPF-World. The reader's worry that the predicted \"future states and actions\" might be generated by the same policy that makes the decisions, rather than by a model grounded in true environment transitions, may be exactly right — but that is a downstream question. Right now there is nothing to evaluate, and this is not a gap a referee can fix with a revision.\n\nWho this is for: nobody, in this state. The MAPF-World abstract is a promise; the VimoRAG text is a different project. If the authors submit the actual MAPF paper, the idea deserves a serious referee. As delivered, the responsible move is to return it without full peer review.","headline":"The submission is not internally coherent: the abstract announces MAPF-World, an autoregressive action world model for multi-agent path finding, but the supplied full text is an unrelated paper on video-based 3D motion generation (VimoRAG).","tokens_in":21240,"tokens_out":2360,"would_cite":false,"duration_ms":24883,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MAPF-World claims that a decentralized pathfinding policy which predicts future states and joint actions before each decision plans more coordinately and generalizes better than much larger reactive solvers, using far less data.","keywords":["multi-agent path finding","world model","autoregressive action model","future state prediction","decentralized planning","zero-shot generalization","real-world map generation"],"falsifier":"Run MAPF-World on long-horizon, out-of-distribution maps with the future-prediction branch disabled or supplied with random futures instead of its own predictions. If success rates and collision counts stay essentially the same, then look-ahead is not the load-bearing mechanism and the world-model explanation of the gains is unsupported.","tokens_in":20338,"feed_emoji":"🤖","tokens_out":10525,"duration_ms":91693,"temperature":0.7,"pith_summary":"Multi-agent path finding (MAPF) asks how to move many agents from starts to goals without collisions, and decentralized learnable solvers scale well but act reactively: each choice uses only the current local view, so coordination over long horizons suffers. This paper proposes MAPF-World, an autoregressive action world model that couples situation understanding with action generation by predicting future states and future actions, then conditioning decisions on those predicted futures. The paper claims this look-ahead outperforms state-of-the-art learnable solvers on MAPF benchmarks, including zero-shot transfer to out-of-distribution maps, while training with 96.5% smaller model size and 92% less data. A secondary contribution is an automatic map generator grounded in real-world layouts for training and evaluating solvers. If correct, the work suggests that foresight, not raw scale, is what decentralized multi-agent planners lack.","feed_headline":"Look-ahead planning lets a tiny model beat big MAPF solvers","feed_subtitle":"Predicting futures before acting beats reactive planning with 96.5% fewer parameters and 92% less data.","key_machinery":"The central object is an 'action world model': an autoregressive model, trained alongside the decentralized MAPF policy, that unifies situation understanding and action generation by predicting future states and future actions step by step, then feeding those predicted futures back into the decision. It carries the argument because it is the mechanism intended to replace reactive, observation-only decision-making with look-ahead: the policy conditions each action on where the model predicts the joint configuration of agents and environment will be, and the paper attributes better coordination and long-horizon performance to this explicit modeling of spatial features and temporal dependencies.","core_discovery":"The central claim is that decentralized MAPF solvers degrade in complex, long-term scenarios because they are reactive policies with weak accounts of environmental temporal dynamics and inter-agent dependencies. MAPF-World answers by unifying situation understanding and action generation in one autoregressive model that explicitly predicts future states and future actions, so each decision is informed by where the joint configuration of agents and environment is heading rather than by immediate local observations alone. The paper attributes the improved situational awareness, coordination, and far-sightedness of its solver to this explicit modeling of spatial features and temporal dependencies. In experiments, MAPF-World is reported to outperform state-of-the-art learnable solvers, to generalize zero-shot to out-of-distribution cases, and to do so with a 96.5% smaller model and 92% less training data, aided by a new real-world-grounded map generator.","pith_inferences":["If the result is correct, it implies that the bottleneck for learned MAPF solvers is not capacity but foresight: deliberately predictive objectives could substitute for the scale that foundation-model-based solvers rely on. This is my extrapolation, not the paper's claim.","The same predict-then-act recipe transfers naturally to other decentralized sequential-decision settings, such as warehouse robot fleets, traffic intersections, and social navigation, where reactive policies also degrade over long horizons. This is an editorial inference.","The abstract does not show that predicted futures are checked against true environment transitions; if the future-prediction head merely re-expresses the decision policy, the reported gains could come from the auxiliary loss's regularization rather than from genuine look-ahead. This is an editorial caveat, not something the paper claims."],"forward_implications":["On the paper's benchmarks, MAPF-World outperforms prior state-of-the-art learnable solvers while using 96.5% fewer parameters and 92% less training data.","The solver exhibits zero-shot generalization to out-of-distribution maps, meaning its performance transfers to layouts unlike those seen in training.","The new automatic map generator, grounded in real-world layouts, provides a more practical training and evaluation protocol than purely synthetic benchmark maps.","Future-state and future-action prediction is presented as the mechanism that fixes the reactive-policy limitation of prior decentralized learnable solvers."],"supporting_citations":[],"fun_headline_variants":["Future-state prediction powers smaller, data-light MAPF solver","Autoregressive world model boosts MAPF with 96.5% fewer parameters","Predicting agent futures beats reactive MAPF with 92% less data","World model predicts futures for far-sighted multi-agent planning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approach assumes that the future states and actions the model predicts faithfully reflect how the multi-agent environment actually evolves; if those predictions come from the same decision process that chooses the actions, the claimed benefit of looking ahead has no support.","fun_headline_variants_meta":{"raw":{"variants":["Future-state prediction powers smaller, data-light MAPF solver","Autoregressive world model boosts MAPF with 96.5% fewer parameters","Predicting agent futures beats reactive MAPF with 92% less data","World model predicts futures for far-sighted multi-agent planning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000682,"raw_usage":{"total_tokens":3110,"prompt_tokens":973,"completion_tokens":2137,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":2062}},"tokens_in":589,"tokens_out":2137,"duration_ms":16234,"temperature":1.0,"reasoning_tokens":2062,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:24:37.952310+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run MAPF-World on long-horizon, out-of-distribution maps with the future-prediction branch disabled or supplied with random futures instead of its own predictions. If success rates and collision counts stay essentially the same, then look-ahead is not the load-bearing mechanism and the world-model explanation of the gains is unsupported.","supporting_citations":[],"review_version":1}