{"id":"83b48aad-c602-4b89-9519-b778fd7446b4","arxiv_id":"2607.06401","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"A perspective article defining world models as finite-resource compression of physical state transitions and outlining a roadmap toward physical AGI via unified representations and interactive simulators.","lead":"This paper proposes a formal definition of 'world models' as compression mechanisms for physical state transitions, plus a staged roadmap from multimodal models to physical AGI. A generalist might read it to understand how the AI field is organizing itself around building machines that understand physical reality, not just language.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The compression-first definition is conceptually sound, but the 'one state, many decoders' principle lacks a guarantee that lossy compression preserves contact-rich physical details needed for control.","rationale":"The reader correctly identified the load-bearing assumption: that a single unified physical representation learned from passive video can support all downstream tasks without critical information loss, particularly for contact-rich manipulation. I agree with this assessment. The paper is a perspective article, and its definition of world models as compression under finite computational resources is internally consistent and well-situated in the literature. The compression framing is genuinely useful for organizing the architectural landscape. However, the roadmap's central architectural proposal—the 'one state, many decoders' principle in Section 7.2—relies on an unproven assumption that lossy compression of internet video can preserve decision-relevant physical details (contact, force, deformation) that are not directly observed in the training data. The paper itself acknowledges this risk in Section 3.2 but does not resolve it. Since the paper is explicitly a roadmap and does not claim empirical validation of this principle, the reader's CONDITIONAL verdict is appropriate. The concern does not invalidate the paper's contribution as a synthesis and roadmap, but it correctly flags the most critical assumption on which the proposed path forward depends. No verdict adjustment is needed.","tokens_in":47688,"tokens_out":759,"duration_ms":323149,"concrete_test":"Construct a controlled experiment using a contact-rich manipulation task (e.g., peg-in-hole insertion or cloth folding). Train a unified representation on internet video + task-specific data as proposed in Section 7.2. Then measure: (1) whether the learned latent state preserves force/contact-relevant information by probing it with a linear decoder for contact events, and (2) whether a planner operating on the compressed state achieves task success comparable to a planner with access to raw proprioceptive/force data. If the probe accuracy or task success drops significantly compared to the raw-state baseline, the 'one state, many decoders' assumption is violated for contact-rich tasks.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central thesis (Definition 2.1, Section 2.1) frames world modeling as 'strictly a compression problem.' This is a reasonable conceptual lens, but the roadmap's core architectural claim—the 'one state, many decoders' principle (Section 7.2)—assumes that a single compact internal representation can be learned from internet-scale video data and simultaneously support rendering, simulation, and planning without discarding decision-relevant physical details. The paper itself acknowledges this risk in Section 3.2: 'excessive abstraction may discard information that is visually small but decision-critical, such as gripper-object contact, thin obstacles, subtle object pose changes, or deformation.' However, it offers no analytical or empirical argument for why the information-theoretic tradeoff of lossy compression would not destroy exactly this class of information. Internet video lacks proprioceptive, force, and tactile channels; a representation learned primarily from this data has no signal from which to compress contact-rich dynamics. The paper's proposed Inverted Pyramid Workflow (Figure 1) attempts to bridge this gap by funneling web data into task-specific datasets, but this does not address whether the *unified representation* itself retains the physical structure needed for contact-rich manipulation. The concern is not that the compression framing is wrong—it is internally consistent and well-motivated—but that the leap from 'compression is the goal' to 'a single compressed state can serve all decoders including control' is presented as a roadmap target without addressing the rate-distortion tradeoff at the level of physical observables.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"This perspective article proposes a scientific definition of world models as compression models of physical state-transition processes under finite computational constraints, and lays out a staged roadmap progressing from unified multimodal models through unified physical representations to foundation-scale interactive simulators. The paper synthesizes a broad literature spanning model-based RL, video generation, 3D representations, causal inference, and embodied AI, organizing it through a two-dimensional taxonomy (functional roles × architectural substrates) and introducing conceptual constructs including the Inverted Pyramid Workflow, the 'one state, many decoders' principle, and a Trinity Architecture for physical AGI. The mathematical formulations (POMDPs, Bayesian filtering, structural causal models, MBRL objectives) are standard and correctly applied.","tokens_in":48377,"tokens_out":1324,"duration_ms":241824,"significance":"The paper's primary contribution is conceptual: it provides a unifying definition and organizational framework for a field that currently lacks terminological consensus. The compression-first framing (Definition 2.1) is a reasonable and potentially useful lens that connects information-theoretic principles to architectural decisions. The two-dimensional taxonomy (Section 2.5, Figure 5) and the functional taxonomy discussion (Section 2.4) are genuinely clarifying for a fragmented literature. The paper does not present machine-checked proofs, reproducible code, or parameter-free derivations, but it does offer falsifiable predictions—particularly the claim that a single shared physical representation can simultaneously support rendering, simulation, and planning (Section 7.2)—which could in principle be tested. The breadth of literature coverage is substantial and the roadmap is concrete enough to guide near-term research. The discussion of counterfactual reasoning (Section 4.6) and physics-informed learning (Section 4.5) are well-grounded treatments that connect world models to established formal frameworks.","major_comments":[{"comment":"Section 7.2 articulates the 'one state, many decoders' principle as the central architectural thesis of the roadmap: a single compact internal representation should support rendering, simulation, and planning simultaneously. The paper itself acknowledges the core risk in Section 3.2: 'excessive abstraction may discard information that is visually small but decision-critical, such as gripper-object contact, thin obstacles, subtle object pose changes, or deformation.' However, no analytical or empirical argument is provided for why lossy compression would preserve exactly this class of information. The concern is sharpened by the Inverted Pyramid Workflow (Figure 1), which proposes learning primarily from internet video—a data source that lacks proprioceptive, force, and tactile channels. The paper should either (a) provide an information-theoretic or empirical argument for why the unified","section":null},{"comment":"representation can retain contact-rich physical details when trained on data that does not contain them, or (b) explicitly acknowledge this as a fundamental limitation and scope the 'one state, many decoders' claim accordingly. As stated, the gap between the compression-first definition (which is sound) and the unified-representation roadmap claim (which is unproven) is the weakest load-bearing link in the argument.","section":null},{"comment":"Section 8 introduces the 'Trinity Architecture' (Agent–Evaluator–World Model) as a cognitive loop for autonomous evolution toward physical AGI. This construct is presented at a conceptual level without connection to the formal POMDP/MBRL framework developed in Sections 2–4. Specifically, the relationship between the Evaluator component and the reward function R in the POMDP formulation (Section 2.2) is unclear, and the claim that the World Model component 'knows the edge of feasible tasks of the current Actor' is not grounded in any formal mechanism. The Trinity Architecture would benefit from either formalization within the existing framework or explicit acknowledgment that it is a speculative conceptual proposal.","section":null}],"minor_comments":[{"comment":"Section 2.5 references 'Figure 1' for the many-to-many mapping of systems to taxonomy categories, but Figure 1 depicts the Inverted Pyramid Workflow. The intended reference appears to be Figure 5. Please verify cross-references.","section":null},{"comment":"Several typos throughout: 'internaml model' (Section 1), 'interchangably' (Section 2.1 footnote), 'adanced' and 'genralist' (Section 4.2), 'reies' and 'dexterous manipulation tasks' (Section 4.2), 'natually' (Section 8). Proofreading pass recommended.","section":null},{"comment":"Section 4.2, paragraph on DreamZero: the sentence beginning 'However, DreamZero does not really resolve the covariate shift...' is grammatically fragmented and should be revised for clarity.","section":null},{"comment":"The paper cites numerous 2026-dated works (e.g., NVIDIA Cosmos Team 2026, ByteDance Seed Team 2026, several arXiv preprints). Given the July 2026 submission date, these may be very recent preprints. Where possible, the authors should verify that cited preprints have stable identifiers and that claims about these systems are accurate as of the cited versions.","section":null},{"comment":"Table 1 is informative but would benefit from a column indicating which specific limitations are most critical for each paradigm, rather than listing them uniformly. This would improve usability as a reference.","section":null},{"comment":"Section 6.6 introduces 'Federated World Models' and 'Trusted Execution Environments' as proposed solutions for privacy-preserving governance. These concepts are introduced without prior context or literature grounding. A brief reference to existing federated learning or secure computation literature would help readers calibrate their expectations.","section":null},{"comment":"Equation (4) presents scaling laws with the notation C≈κND, but the constant κ is not defined in the surrounding text. Please add a definition.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The paper is a perspective article from the Shanghai AI Laboratory Physical Intelligence Team and cites a substantial number of systems from Chinese AI companies (ByteDance, Alibaba, Tencent, Kling). The citation pattern appears balanced overall, with appropriate coverage of Western and international systems (Sora, JEPA, Dreamer, NVIDIA Cosmos, etc.). The self-citation rate is low. The paper's scope is appropriate for a perspective article in a serious AI journal, though the breadth means that individual topics are sometimes treated at a survey level rather than with deep technical analysis. The central conceptual contribution (the compression-first definition) is sound; the main risk is that the roadmap's architectural claims outpace the evidence provided for them."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive reading of our manuscript. The two major comments both identify genuine gaps in the argument that we will address in revision. On the first comment, regarding the 'one state, many decoders' principle and the Inverted Pyramid Workflow's reliance on internet video lacking proprioceptive channels, we agree that the manuscript does not adequately address how contact-rich physical details can be preserved when training on data that does not contain them. We will revise Section 7.2 to explicitly scope the claim and acknowledge this as a fundamental limitation requiring complementary data sources. On the second comment, regarding the Trinity Architecture's lack of formal connection to the POMDP/MBRL framework, we agree that Section 8 is presented at a conceptual level without grounding in the formal apparatus of Sections 2-4. We will add an explicit mapping between the Evaluator and the reward function R, clarify the World Model's 'edge of feasible tasks' claim in terms of model uncertainty and reachable-state estimation, and acknowledge the speculative nature of the proposal where formalization is not yet possible.","responses":[{"response":"The referee correctly identifies the weakest load-bearing link in our argument. We do not have an information-theoretic proof that lossy compression trained on internet video will preserve contact-rich physical details, and we agree that the Inverted Pyramid Workflow's reliance on video data lacking proprioceptive, force, and tactile channels creates a genuine gap between the compression-first definition (which is sound as a general principle) and the unified-representation roadmap claim (which is unproven for this data regime). We cannot honestly provide argument (a): there is no existing theoretical result guaranteeing that information absent from the training distribution can be recovered by compression, and empirical evidence from current video-based world models (as we note in Section 6.2) shows that perceptual fidelity does not imply physical precision. We will therefore adopt option (b). Specifically, we will revise Section 7.2 to: (1) explicitly state that the 'one state, many decoders' principle is an architectural thesis whose feasibility for contact-rich domains is not yet established; (2) acknowledge that internet video, while encoding broad physical priors (object permanence, rigidity, kinematic structure), does not contain proprioceptive, force, or tactile channels, and that no compression mechanism can recover information entirely absent from the training data; (3) scope the claim to the modalities present in the training distribution, noting that contact-rich and tactile information require complementary data sources (embodied interaction data, tactile sensors, physics-informed constraints as discussed in Section 4.5); and (4) add a forward reference to Section 6.1 (Data Asymmetry) and Section 6.4 (Sim-to-Real Transfer), which already discuss these bottl","revision_made":"no","referee_comment":"Section 7.2 'one state, many decoders' principle: no argument provided for why lossy compression preserves contact-rich details when trained on internet video lacking proprioceptive/force/tactile channels. Paper should either (a) provide an information-theoretic or empirical argument, or (b) acknowledge as fundamental limitation and scope the claim."},{"response":"We agree that Section 8 is insufficiently connected to the formal framework developed in Sections 2-4, and we will revise accordingly. We can partially formalize the mapping: (1) The Evaluator component corresponds to the reward function R in the POMDP formulation (Section 2.2), assessing trajectory quality against task objectives. We will state this mapping explicitly. (2) The Agent corresponds to the policy pi_phi, and the World Model corresponds to the learned transition model P_hat_theta, both already formalized in Section 4.2. (3) The claim that the World Model 'knows the edge of feasible tasks of the current Actor' can be partially grounded in the model uncertainty and reachable-state estimation literature discussed in Section 4.2: ensemble disagreement (Chua et al., 2018), conservative rollout termination (Yu et al., 2020; Kidambi et al., 2020), and the horizon-limited value gap (Eq. 19) all provide formal mechanisms for estimating where the model's predictions become unreliable, which operationalizes 'the edge of feasible tasks.' We will add these connections explicitly. However, we acknowledge that the Trinity Architecture's claim of autonomous curriculum generation—where the World Model proposes tasks 'just beyond' the Agent's current limits—goes beyond what existing formal mechanisms fully support. The automatic curriculum learning literature (e.g., Plan2Explore's intrinsic motivation, Sekar et al., 2020) provides partial precedents, but the closed-loop self-evolution we describe remains a speculative conceptual proposal. We will state this explicitly in the revised Section 8, marking it as a research direction rather than a formally grounded result.","revision_made":"no","referee_comment":"Section 8 Trinity Architecture presented conceptually without connection to formal POMDP/MBRL framework. Relationship between Evaluator and reward function R unclear. Claim that World Model 'knows the edge of feasible tasks' not grounded in formal mechanism. Should formalize within existing framework or acknowledge as speculative."}],"tokens_in":47679,"tokens_out":1082,"duration_ms":172382,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"This is a perspective article from the Shanghai AI Lab team that tries to define what a world model is and lay out a roadmap toward physical AGI. The punchline: the compression-centric definition is useful and the taxonomy is genuinely clarifying, but the central architectural claim—that one compact representation can serve rendering, simulation, and planning simultaneously—gets asserted as a target without anyone grappling with the rate-distortion math that would tell us whether it's achievable. What's new and good: Definition 2.1 (world model as compression of state-transition processes under finite compute) is a clean framing that subsumes a lot of scattered work. The two-dimensional taxonomy (function × architecture) is the best organizational contribution here—it cleanly separates 'what does the model output' from 'how is world knowledge represented' and maps existing systems (Sora, JEPA, Dreamer, Cosmos, Marble) onto both axes without forcing them into ill-fitting buckets. The survey sections on MBRL, counterfactual reasoning, and physics-informed learning are competent and well-referenced. The Chain-of-Imagination section (reasoning as imagined action-consequence pairs in latent space rather than verbal chains) is a genuinely interesting idea, even if it's still conceptual. The soft spot is real but should be sized correctly. The paper's roadmap (Section 7.2) claims that a single unified physical representation can be learned from internet-scale video and simultaneously support rendering, simulation, and planning via 'one state, many decoders.' The stress-test concern lands here: internet video lacks proprioceptive, force, and tactile channels, so a representation learned primarily from that data has no signal from which to compress contact-rich dynamics. The paper itself acknowledges this risk in Section 3.2 ('excessive abstraction may discard information that is visually small but decision-critical'), but then doesn't address the rate-distortion tradeoff analytically or empirically. The Inverted Pyramid Workflow (Figure 1) is a data pipeline, not a solution to the representation problem. That said, this is a perspective article, not a methods paper. It's proposing a research direction, not claiming to have solved it. The gap between 'compression is the goal' and 'a single compressed state serves all decoders' is presented as a roadmap target, which is fair for what this is. The Trinity Architecture (agent-evaluator-world model loop) is reasonable as a conceptual blueprint but adds little beyond what the self-improvement literature already describes. Who gets value: researchers entering the world model space who want a map of the territory, and people working on one subproblem (video generation, MBRL, representation learning) who want to see how their piece connects to others. It deserves a serious referee. The synthesis is real, the taxonomy is useful, and the compression framing is worth engaging with—even if the unified-representation claim needs either analytical grounding or a concrete demonstration to move from aspiration to evidence.","headline":"Solid synthesis with a real soft spot in the central architectural claim","tokens_in":48367,"tokens_out":678,"would_cite":true,"duration_ms":61791,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"A world model is a compression problem, not a generation problem","keywords":["world models","lossy compression","physical representation","embodied AI","model-based reinforcement learning","POMDP","world action models","sim-to-real transfer"],"falsifier":"A contact-rich manipulation task where a visually plausible but physically incorrect compressed representation causes a robot to fail in ways that a less compressed but more physically faithful representation would not—demonstrating that the compression objective and the control objective are fundamentally in tension rather than aligned.","tokens_in":47805,"feed_emoji":"🗜️","tokens_out":1387,"duration_ms":149578,"temperature":0.7,"pith_summary":"The paper argues that the term 'world model' has become overloaded across AI subfields—applied to video generators, reinforcement-learning dynamics models, and embodied planners without a shared definition. The authors propose a single formal definition: a world model is a compression of the physical world's state-transition processes, built under finite computational resources. Under this definition, the core task is not to generate realistic pixels or simulate futures, but to distill high-dimensional sensory data into a compact internal representation that preserves the physical, causal, and dynamical structure needed for reasoning, planning, and control. Generation and simulation are downstream decodings of this representation, not the objective itself. The paper frames understanding as primary and prediction as secondary—a model must first identify what is happening and why before it can predict what will happen. It situates this definition within a POMDP agent-environment loop, proposes a two-dimensional taxonomy (functional role × architectural substrate) that supersedes prior one-dimensional classifications, and lays out a staged roadmap: first unify modalities, then distill them into a single shared physical representation from which rendering, simulation, and planning are all decoded, and finally scale this into foundation-level interactive simulators. The paper also introduces an 'Inverted Pyramid' data pipeline that funnels internet-scale video down to compact task-aligned robot training data, arguing that data diversity—not architecture or compute—sets the ceiling on physical generalization.","feed_headline":"A world model is a compression problem, not a generation problem","feed_subtitle":"Redefining the field's most debated concept: the real task is distilling physical structure from pixels, and everything else is just a","key_machinery":"The load-bearing mechanism is the unified physical representation—a single, compact, learned internal state encoding geometry, motion, material properties, appearance, semantics, uncertainty, and interaction state jointly. From this state, rendering (producing visual observations), simulation (propagating physical dynamics under action), and planning (evaluating counterfactual futures to select actions) are recovered as distinct decoding operations. The POMDP loop provides the formal scaffolding: the agent maintains a Bayesian belief over latent states, the world model approximates both the transition kernel and observation model, and the planner optimizes over the resulting belief-continged","core_discovery":"The central conceptual move is the redefinition of a world model as a compression of physical state transitions under finite resources, coupled with the principle that understanding should be primary while prediction serves it. This reframing dissolves the boundary between renderers, simulators, and planners: they become different decoding operations on a single compressed internal state. The paper calls this the 'one state, many decoders' principle and argues that the field's central open problem is discovering what compact internal structure can preserve sufficient physical and semantic information to support all downstream projections of an embodied intelligence.","pith_inferences":["The compression framing implicitly assumes a favorable information-theoretic structure: that the physical regularities relevant for control occupy a low-dimensional manifold within high-dimensional sensory data. If the decision-relevant physical state for contact-rich manipulation is actually high-dimensional (e.g., requiring fine-grained deformation fields or friction distributions), then lossy c","The claim that data diversity sets the ceiling while architecture only affects efficiency echoes scaling-law arguments from language modeling, but physical data may not obey the same power-law regime: the long tail of rare physical events (edge cases in contact, fracture, anomalous dynamics) may be so structurally diverse that no finite corpus captures sufficient coverage, making the 'ceiling' a m","The Trinity Architecture's self-curriculum loop assumes the world model can reliably identify the edge of the agent's capability. If the model's uncertainty estimates are poorly calibrated—a known problem in deep generative models—the curriculum will either propose trivial tasks (stalling progress) or impossible ones (wasting interaction), turning the self-improvement loop into a random walk."],"forward_implications":["If the compression-first definition is correct, then evaluating world models by visual fidelity (FID, FVD) is measuring the wrong thing; the right metrics should test whether the compressed state preserves decision-relevant physical structure—contact dynamics, causality, object permanence—sufficient for downstream control.","The 'one state, many decoders' principle implies that current systems maintaining separate representations for rendering (radiance fields), simulation (meshes/particles), and planning (occupancy grids) are architecturally suboptimal; a single shared substrate should eventually replace them.","The Inverted Pyramid pipeline implies that the path to generalist robot policies runs through internet-scale video, not through more robot data collection—data diversity from passive video, not interaction volume, sets the generalization ceiling.","The Trinity Architecture (Agent–Evaluator–World Model) implies a self-improving loop where the world model generates curricula at the edge of the agent's current capability, suggesting that world models are not just predictors but active drivers of embodied skill acquisition.","If understanding is primary and prediction is derivative, then architectures that skip explicit representation learning (pure pixel-to-pixel video generators) will plateau as physical reasoners regardless of scale, because they lack the compressed state that carries causal and dynamical structure."],"fun_headline_variants":["World models should compress physical state, not generate pixels","One compressed state, many decoders: a unified theory of world models","Understanding over prediction: reframing the world model debate","Renderers, simulators, and planners are all decoders on one state","The open problem: what compact structure preserves physical semantics"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The paper assumes that a single, unified, compact internal representation can be learned from internet-scale video data that preserves sufficient physical structure—geometry, contact, dynamics, causality—to support rendering, simulation, and planning simultaneously. This 'one state, many decoders' principle presupposes that the information loss inherent in lossy compression will not destroy decision-relevant physical details, which remains unproven for complex, contact-rich, ","fun_headline_variants_meta":{"raw":{"variants":["World models should compress physical state, not generate pixels","One compressed state, many decoders: a unified theory of world models","Understanding over prediction: reframing the world model debate","Renderers, simulators, and planners are all decoders on one state","The open problem: what compact structure preserves physical semantics"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":503,"prompt_tokens":419,"completion_tokens":84,"prompt_tokens_details":null},"tokens_in":419,"tokens_out":84,"duration_ms":56617,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T06:53:03.160464+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"A contact-rich manipulation task where a visually plausible but physically incorrect compressed representation causes a robot to fail in ways that a less compressed but more physically faithful representation would not—demonstrating that the compression objective and the control objective are fundamentally in tension rather than aligned.","supporting_citations":[],"review_version":1}