{"id":"27773b80-d7ff-4633-933c-17b46ba5561a","arxiv_id":"2607.05966","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":4,"one_line_summary":"DreamerV3's imagined rollouts are insensitive to friction changes that cause real gait collapse, revealing that world models extrapolate kinematically rather than dynamically.","lead":"The paper shows that world models like DreamerV3 extrapolate trajectories using simple kinematic rules rather than true physics: when friction changes enough to make a walker fall, the model's imagined rollouts don't notice. This matters because it identifies a specific structural blind spot in AI systems that learn by simulating their own future.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The 5-step conditioning prefix is the load-bearing confound: the WM's friction-invariance could reflect insufficient regime evidence rather than kinematic imagination, and the DR control does not fully close this gap because it also conditions on only 5 steps at evaluation time.","rationale":"The reader correctly identified the most load-bearing concern: the 5-step conditioning prefix is a genuine confound that the paper's controls do not fully resolve. The DR control closes the training-distribution loophole but not the evaluation-time information bottleneck. The joint-noise positive control establishes that the WM extracts kinematic features from the prefix but does not establish that the prefix carries friction information. These are real gaps. However, I do not adjust the verdict from CONDITIONAL because the reader's assessment is already calibrated: the paper has multiple independent controls (3 seeds, DR, actor-horizon ablation, per-step structure, state-choice robustness, joint-noise positive control), ships code and data, and is transparent about this exact limitation. The proposed explicit-conditioning and prefix-length-sweep experiments are the right next steps. The verdict of CONDITIONAL with MODERATE confidence accurately reflects that the finding is suggestive and well-executed but not definitive pending the prefix-length experiment. No adjustment needed.","tokens_in":13343,"tokens_out":4481,"duration_ms":300804,"concrete_test":"Take the DR-trained WM (which has seen friction variation during training and therefore has the capacity to represent friction-conditional dynamics) and sweep the conditioning prefix length: condition on 5, 15, 25, 50, and 64 observed steps from perturbed environments before rolling out free imagination for the remaining T−prefix steps. Compute the friction-sweep regression slope β_WM at each prefix length. If β_WM becomes nonzero (CI excludes zero) at prefix lengths ≥25 (covering one gait period), the 5-step bottleneck is confirmed and the central claim weakens to 'WMs cannot infer friction from short prefixes.' If β_WM remains statistically zero even at 64-step conditioning (~1.6s, well over a gait period), the kinematic-imagination claim is substantially strengthened, as the WM had sufficient regime evidence and still produced friction-invariant rollouts.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the WM's friction-invariant iKCE reflects kinematic imagination, not an information bottleneck. The WM conditions on only 5 observations (125ms, under one gait period) before free imagination. The paper itself argues (§III-B) that friction effects 'accumulate over multiple footfalls' and become detectable only at horizons longer than the gait period — but this argument is applied to the *measurement* horizon (T=64), not the *conditioning* prefix. If 5 steps do not carry enough friction signal for even a dynamically capable WM to infer the regime, then flatness across friction is uninformative: it would hold for any WM regardless of whether it imagines kinematically or dynamically. The domain-randomization control (Appendix B2) addresses the *training-distribution* loophole (the DR WM has seen friction variation and could in principle learn friction-conditional latent dynamics), but it does not address the *evaluation-time* bottleneck: the DR WM still conditions on only 5 steps at test time. The joint-noise positive control (Fig. 2, right) shows the WM extracts *kinematic* information from the 5-step prefix, but this does not establish that the prefix carries *dynamic* (friction) information — kinematic state is directly observable in a single step, while friction must be inferred from its effects on contact dynamics over multiple footfalls. The paper acknowledges this (§V, limitation ii) and proposes the explicit-conditioning experiment as future work, but until that experiment is run, the central claim is supported but not definitively established. This is not a fatal flaw — the controls are thorough and the finding is suggestive — but it is the single point on which the entire argument turns.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"The paper proposes a kinematic-vs-dynamic reframing of long-horizon world-model failure, introducing the imagined Kinematic-Consistency Error (iKCE) as a diagnostic metric. The central claim is that world models imagine kinematically—extrapolating position-velocity-acceleration trajectories consistent with linear kinematic update rules but inconsistent with physical constraints. The diagnostic is instantiated on a DreamerV3 checkpoint trained on DMC walker-walk, with two predicted signatures: (H1) iKCE elevated above matched real-physics rollouts, and (H2) statistical invariance of imagined iKCE across a friction sweep that crosses the gait-collapse boundary. The experimental suite includes three seeds, a domain-randomization control, an actor-horizon ablation, a kinematic-state robustness check, and a joint-noise positive control. The falsifiable regression framework (β with bootstrap CIs) is well-executed.","tokens_in":14075,"tokens_out":1544,"duration_ms":351609,"significance":"The paper makes a genuine conceptual contribution by distinguishing kinematic imagination from dynamic imagination as a structural failure mode, separate from compounding-error and representation-engineering accounts. The iKCE diagnostic is clearly defined, falsifiable, and embodiment-agnostic in principle. The experimental instantiation is thorough: three independent seeds with bootstrap CIs, a domain-randomization control that closes the training-distribution loophole, an actor-horizon ablation, a per-step temporal-structure decomposition, robustness to kinematic-state choice, and a joint-noise positive control that confirms the WM is not insensitive to all perturbations. The release of code, checkpoints, and data is commendable and strengthens reproducibility. The horizon-emergence test (Fig. 3) is a particularly clean result showing the physics-side friction sensitivity grows with measurement horizon while the WM side remains flat at all horizons.","major_comments":[{"comment":"§IV (H2) and §V, limitation (ii): The 5-step conditioning prefix (125ms, under one gait period) is a load-bearing confound for the central claim. The paper argues in §III-B that friction effects 'accumulate over multiple footfalls' and become detectable only at horizons longer than the gait period—this argument is applied to the measurement horizon T=64 but not to the conditioning prefix. If 5 steps do not carry enough friction signal for even a dynamically capable WM to infer the regime, then the WM's flatness across friction is uninformative: it would hold for any WM regardless of whether it imagines kinematically or dynamically. The domain-randomization control (Appendix B2) addresses the training-distribution loophole (the DR WM has seen friction variation during training), but it does not address the evaluation-time bottleneck: the DR WM still conditions on only 5 steps at test time","section":null},{"comment":"§II: The four motivating observations are drawn from the driving VLM/VLA setting, but the diagnostic is instantiated only on DMC walker-walk (a 2D 9-DOF planar system). The paper asserts these observations 'converge on a single structural deficit' but does not test whether the iKCE diagnostic reproduces the kinematic-not-dynamic signature on any driving world model. The connection between §II and §IV is therefore an asserted analogy rather than a demonstrated one. This is acknowledged as a limitation (§V, limitation i) and proposed as future work (Vista, DriveDreamer), but it weakens the bridge between the motivation and the experimental contribution. The paper should either scope the central claim to locomotion world models or provide stronger justification for why the driving-VLM observations necessarily imply the same deficit in DreamerV3.","section":null}],"minor_comments":[{"comment":"Table I: The ratio narrowing from ~180× at T=16 to ~30× at T=64 is attributed to 'per-step dilution from the WM's smooth long-horizon tail.' It would help to clarify whether this dilution is a property of the integrated metric (averaging over more steps where per-step displacement decays) or a genuine change in the WM's per-step behavior at longer horizons.","section":null},{"comment":"Fig. 2 caption: 'iKCE diverges in physics, stays flat in imagination' is slightly misleading given that the physics-side elevation is modest (1.04–1.30×10⁻⁴ in low-µ vs. 0.70–0.92×10⁻⁴ in high-µ). Consider 'iKCE shows friction sensitivity in physics, remains flat in imagination' or similar.","section":null},{"comment":"§III-A, Eq. (1): The kinematic predictor kin(·) is defined as 'any closed-form kinematic predictor (e.g., constant-velocity or constant-acceleration).' The choice of predictor is embodiment-specific and could affect the diagnostic's sensitivity. The paper uses constant-velocity throughout but does not discuss how the choice of null model interacts with the diagnostic signature. A brief remark on this would strengthen the diagnostic.","section":null},{"comment":"Appendix A2: The three-seed regression is reported only for the WM side; the physics-side regression uses seed 0 only ('as physics is not a learned model'). While this is reasonable, a brief note on the physics-side variability across seeds (even if not a learned model, the policy is) would be informative.","section":null},{"comment":"§V, limitation (iii): The per-step displacement decay over the rollout horizon is noted but not quantified. A simple plot of mean per-step |Δx| over t for both channels would let the reader assess how much of the long-horizon iKCE reduction is a motion-magnitude artifact vs. genuinely cleaner imagination. This could be deferred to the appendix.","section":null},{"comment":"The paper uses 'kinematic' and 'dynamic' in a specific sense (§I) that differs from some robotics conventions where 'kinematic' refers to forward/inverse kinematics. A brief note that this is the classical-mechanics usage would prevent reader confusion.","section":null}],"recommendation":"major_revision","confidential_remarks":"The 5-step prefix issue is the key question for this paper. The authors are aware of it (§V, limitation ii) and propose the explicit-conditioning experiment as future work. I think the paper would be substantially strengthened if at least a prefix-length sweep (5 vs. 10 vs. 20 vs. 40 steps) could be run on the existing checkpoints without retraining—this would directly test whether the flatness is an information bottleneck or a structural property. If the authors can show that the WM remains flat even with longer prefixes, the central claim is strongly supported; if they cannot run this experiment, the paper should more prominently caveat the claim and acknowledge that the current evidence is consistent with both kinematic imagination and an evaluation-time information bottleneck. The DR control is a good partial mitigation but does not fully close the gap, as the reader's analysis correctly identifies. Two of the four motivating observations (§II items i and iv) come from the authors' own prior work (EgoDyn-Bench, StyleVLA), which should be transparently disclosed."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive report. Both major comments identify genuine limitations of the current manuscript. We agree that the 5-step conditioning prefix is a load-bearing confound that weakens the H2 claim, and that the bridge from driving-VLM motivation to the DMC instantiation is presently an asserted analogy rather than a demonstrated one. We propose concrete revisions for both.","responses":[{"response":"The referee is correct that the 5-step conditioning prefix is a load-bearing confound, and we agree this weakens the H2 claim in its current form. We acknowledge this honestly in §V, limitation (ii), but the referee is right that the paper does not adequately confront the implication: if 5 steps (125ms, under one gait period) cannot carry enough friction signal for any WM to infer the regime, then the WM's flatness is uninformative because it would hold for a dynamically capable WM as well. We cannot fully resolve this confound within the revision cycle because a conditioning-prefix-length sweep (5–64 observed steps) requires new rollouts and is listed in our own open directions. However, we can and will do the following: (1) Promote the prefix-length limitation from a secondary caveat to a primary bound on the H2 claim, explicitly stating that H2 is established only up to the regime evidence a 5-step prefix can carry. (2) Add the horizon-emergence argument as partial mitigation: the physics-side friction sensitivity grows with measurement horizon T (Fig. 3, β_phys crossing zero between T=32 and T=64), which is consistent with friction signal accumulating over the rollout itself, not just the prefix. This means the WM has 64 steps of its own imagined trajectory in which friction-conditional deviations could appear if the model had learned any friction-conditional latent dynamics—and it does not, even under the DR control where the WM was trained on the full friction range. (3) Reframe the H2 claim as a conditional result: given a 5-step prefix, the DR-trained WM (which has seen friction variation and could in principle encode it) produces friction-invariant rollouts. This is informative about the WM's rollout dynamics even if it does not rule out the possibility that a ","revision_made":"partial","referee_comment":"§IV (H2) and §V, limitation (ii): The 5-step conditioning prefix (125ms, under one gait period) is a load-bearing confound for the central claim. The paper argues in §III-B that friction effects 'accumulate over multiple footfalls' and become detectable only at horizons longer than the gait period—this argument is applied to the measurement horizon T=64 but not to the conditioning prefix. If 5 steps do not carry enough friction signal for even a dynamically capable WM to infer the regime, then the WM's flatness across friction is uninformative: it would hold for any WM regardless of whether it imagines kinematically or dynamically. The domain-randomization control (Appendix B2) addresses the training-distribution loophole (the DR WM has seen friction variation during training), but it does not address the evaluation-time bottleneck: the DR WM still conditions on only 5 steps at test time"},{"response":"The referee is correct that the connection between §II and §IV is an asserted analogy, not a demonstrated one, and we agree the paper should be more honest about this. We cannot run the iKCE diagnostic on a driving world model (Vista, DriveDreamer) within the revision cycle because these models do not expose the ego-pose/state heads needed to compute iKCE, as we note in §V. We will therefore take the referee's first suggestion and scope the central claim. Specifically: (1) The central claim in §I will be revised to state that the kinematic-not-dynamic hypothesis is motivated by observations in the driving-VLM/VLA setting and instantiated on a locomotion world model, and that the claim of kinematic imagination is demonstrated only for the locomotion case. (2) §II will be reframed as motivating evidence—four independent observations that are consistent with the kinematic-fallback hypothesis—rather than as evidence that converges on a single structural deficit that has been demonstrated across embodiments. (3) The abstract will be adjusted to scope the experimental claim to DreamerV3 on DMC walker-walk, with the driving-VLM observations presented as motivation for the diagnostic framework, not as validated instances of it. (4) We will add an explicit statement that the iKCE diagnostic is embodiment-agnostic in principle (the protocol requires only a kinematic state vector and a perturbation axis) but has been demonstrated on only one embodiment. We believe this scoping is honest and does not diminish the paper's contribution, which is the diagnostic framework and its falsifiable instantiation, not a claim of universality.","revision_made":"yes","referee_comment":"§II: The four motivating observations are drawn from the driving VLM/VLA setting, but the diagnostic is instantiated only on DMC walker-walk (a 2D 9-DOF planar system). The paper asserts these observations 'converge on a single structural deficit' but does not test whether the iKCE diagnostic reproduces the kinematic-not-dynamic signature on any driving world model. The connection between §II and §IV is therefore an asserted analogy rather than a demonstrated one. This is acknowledged as a limitation (§V, limitation i) and proposed as future work (Vista, DriveDreamer), but it weakens the bridge between the motivation and the experimental contribution. The paper should either scope the central claim to locomotion world models or provide stronger justification for why the driving-VLM observations necessarily imply the same deficit in DreamerV3."}],"tokens_in":13175,"tokens_out":1364,"duration_ms":335171,"standing_objections":["The 5-step conditioning prefix confound cannot be fully resolved in this revision. A prefix-length sweep (5–64 observed steps) requires new rollouts and is listed in our open directions. We can bound the claim and provide partial mitigation via the horizon-emergence test and DR control, but we cannot eliminate the confound that a 5-step prefix may not carry enough friction signal for any WM to infer the regime.","The driving-WM cross-anchor (Vista, DriveDreamer) cannot be completed in this revision because these models do not expose the ego-pose/state heads needed to compute iKCE. The bridge from §II motivation to §IV experiments remains an asserted analogy rather than a demonstrated one."]},"desk_editor":{"model":"glm-5.2","letter":"The main thing to know: this paper introduces a genuinely new diagnostic — iKCE, a test-time metric measuring how far imagined rollouts deviate from a constant-velocity kinematic null, paired with a friction-sweep protocol that tests whether the world model responds to physical regime changes. The finding is clear: DreamerV3's imagined iKCE is statistically flat across a friction sweep that crosses the gait-collapse boundary, while real-physics iKCE responds. This is a real contribution. The kinematic-vs-dynamic reframing is not in the DreamerV3 or MBPO literature, and the falsifiable regression framework with bootstrap CIs across 3 seeds is well-executed. The domain-randomization control is the right experiment — the DR-trained WM saw the full friction range at training time and still produces flat iKCE at evaluation. The actor-horizon ablation and joint-noise positive control are thorough. Code and checkpoints are released. Credit is earned here. The soft spot is real and load-bearing: the WM conditions on only 5 observations (125ms, under one gait period) before free imagination. The paper itself argues that friction effects accumulate over multiple footfalls, but applies that argument to the measurement horizon, not the conditioning prefix. If 5 steps don't carry enough friction signal for even a dynamically capable WM to infer the regime, then flatness across friction is uninformative — it would hold for any WM regardless of whether it imagines kinematically or dynamically. The DR control addresses the training-distribution loophole but not this evaluation-time bottleneck: the DR WM still conditions on 5 steps at test time. The joint-noise positive control shows the WM extracts kinematic information from the prefix, but kinematic state is directly observable in a single step while friction must be inferred from contact dynamics over multiple footfalls — so the positive control doesn't establish that the prefix carries dynamic information. The paper acknowledges this honestly (§V, limitation ii) and proposes an explicit-conditioning experiment as future work. That experiment — appending friction indicators to the observation and sweeping prefix length — is exactly what would close the gap. Until it runs, the central claim is supported but not definitively established. The scope mismatch between the broad driving-VLA framing in §II and the single-embodiment DMC walker experiment is a secondary concern; the motivating observations are suggestive but the actual evidence base is one system. This paper is for researchers working on world-model evaluation and model-based RL who want a cheap, falsifiable structural diagnostic. It deserves a serious referee. The 5-step prefix issue is the thing to press on in review — if the authors can run even a partial prefix-length sweep or an explicit-conditioning pilot, that would move the result from suggestive to convincing.","headline":"New diagnostic shows DreamerV3 imagines kinematically, not dynamically — but a 5-step conditioning prefix leaves the central claim plausible rather than definitive","tokens_in":14430,"tokens_out":647,"would_cite":true,"duration_ms":190333,"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":"World models imagine kinematics, not physics","keywords":[],"falsifier":"Extend the conditioning prefix from 5 steps to 32-64 steps of observed data, or explicitly append a friction indicator to the world model's observation. If the world model's iKCE then begins responding to the friction sweep, the kinematic-imagination diagnosis would be weakened in favor of an information-bottleneck explanation.","tokens_in":13530,"feed_emoji":"🦿","tokens_out":846,"duration_ms":134592,"temperature":0.7,"pith_summary":"The paper argues that when today's learned world models roll out imagined trajectories into the future, they extrapolate surface-level motion patterns—position, velocity, acceleration sequences that satisfy simple kinematic update rules—without capturing the underlying physical constraints (friction, contact, mass) that govern real motion. The authors introduce a diagnostic called the imagined Kinematic-istency Error (iKCE), which measures how far each predicted next state in an imagined rollout deviates from a closed-form kinematic extrapolation of its predecessor. The key insight is counterintuitive: low iKCE does not certify dynamic fidelity; it certifies kinematic imagination. The true signature of dynamic imagination is the opposite—iKCE should be positive, grow with horizon, and respond when physical conditions (like friction) cross a regime boundary. The authors test this on a publicly released DreamerV3 checkpoint trained on a walking robot task. They sweep friction across a range that crosses the boundary where the trained policy's gait collapses. Real-physics rollouts show elevated kinematic residuals in the low-friction regime where slipping accumulates over multiple footfalls. The world model's imagined rollouts remain statistically flat across the entire friction sweep—blind to the regime change that catastrophically affects the policy's behavior. A joint-noise control confirms the world model is not insensitive to all perturbations: it responds to kinematic perturbations (joint-position noise) but not to dynamic ones (friction). This pattern holds across three independently trained checkpoints and under a domain-randomization control where the world model was trained on the full friction range. The paper positions kinematic fallback as a third, structurally distinct layer of long-horizon world-model failure, alongside representation-stability engineering and error-compounding bounds.","feed_headline":"World models imagine kinematics, not physics","feed_subtitle":"Friction changes break a walking policy but leave its world model's imagined rollouts unchanged—revealing a structural blind spot.","key_machinery":"iKCE: per-step deviation of imagined rollout states from a constant-velocity kinematic extrapolation. Friction sweep: perturbation of MuJoCo friction across 13 values crossing the empirical gait-collapse boundary. Joint-noise sweep: kinematic positive control adding Gaussian noise to joint-position observations. Log-log regression slope: falsifiable statistic testing whether iKCE responds to friction. Horizon-emergence test: re-integration of per-step traces at T=8,16,32,64 showing the physics-side dynamic signature emerges with horizon while the WM side remains flat at all horizons.","core_discovery":"The central object is the imagined Kinematic-Consistency Error (iKCE), a per-step diagnostic measuring deviation from a closed-form kinematic null along an imagined rollout, paired with a physical-regime perturbation protocol. The core discovery is the kinematic-not-dynamic signature: across a friction sweep crossing the gait-collapse boundary, DreamerV3's imagined iKCE is statistically flat (slope CI contains zero across three seeds and a domain-randomization control), while matched real-physics iKCE responds (slope excludes zero). The world model responds to kinematic perturbations (joint noise) but not to dynamic ones (friction). This flatness persists even when the world model is trained","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["World models imagine kinematics, not dynamics","DreamerV3 ignores friction shifts in imagined rollouts","Imagined rollouts ignore physical regime changes","World models imagine motion but ignore physics","Kinematic-not-dynamic signature in imagined rollouts"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The diagnostic conditions the world model on only five observed steps (125 milliseconds, under one gait period) before free imagination begins. If the world model cannot extract enough regime information from this short prefix to know what friction level it is operating in, its flatness across the friction sweep could reflect an information bottleneck rather than a structural inability to imagine dynamically.","fun_headline_variants_meta":{"raw":{"variants":["World models imagine kinematics, not dynamics","DreamerV3 ignores friction shifts in imagined rollouts","Imagined rollouts ignore physical regime changes","World models imagine motion but ignore physics","Kinematic-not-dynamic signature in imagined rollouts","World model rollouts ignore dynamic shifts"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":913,"prompt_tokens":521,"completion_tokens":392,"prompt_tokens_details":null},"tokens_in":521,"tokens_out":392,"duration_ms":29087,"temperature":1.0,"reasoning_tokens":336,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T19:46:22.314635+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"Extend the conditioning prefix from 5 steps to 32-64 steps of observed data, or explicitly append a friction indicator to the world model's observation. If the world model's iKCE then begins responding to the friction sweep, the kinematic-imagination diagnosis would be weakened in favor of an information-bottleneck explanation.","supporting_citations":[],"review_version":1}