{"id":"867de8fe-a70b-4755-8550-42dd7075720e","arxiv_id":"2608.07598","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"NewtonGS rolls out a 22-dimensional object state (pose, velocity, scale, mass, contact) with hybrid analytic and learned dynamics and maps it to all Gaussians of an object, beating analytic baselines on synthetic trajectories.","lead":"This paper adds a physics-structured object state to 3D Gaussian scenes so objects can be animated from a compact 22-number description instead of per-primitive motion. It reports lower position and velocity errors than five analytic baselines on synthetic motion datasets, though not yet on real reconstructed scenes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"In-distribution State-32 result is selected on the test set: Appendix B.2 says val_id is exactly the ID test set and model selection uses the complete val_id objective, so the ID half of the headline claim is not an out-of-sample comparison.","rationale":"The reader's verdict is correctly CONDITIONAL, but the weakest assumption they name (synthetic-to-real transfer) is an acknowledged scope limitation rather than a defect in the reported numbers. The stronger issue is that the ID half of the headline is not an honest out-of-sample evaluation: the same val_id set is used both for checkpoint selection and for the ID test table. The appendix is unusually transparent about this, but transparency does not remove the circularity. Since the OOD split is not used for selection and still shows the same ordering, the central empirical conclusion is likely to survive a corrected protocol; hence I do not recommend rejecting the paper. The fix is cheap: hold out a true validation split and report both ID and OOD with a checkpoint selected only on that split. Additional strengths that support the core claim include the zero-output initialization of the residuals, the post-hoc branch-removal table, and the explicitly labeled limitations in Appendix G. The state-to-Gaussian affine formula in the main text also differs from the appendix implementation for nonidentity initial orientations (Appendix A.4), which is a secondary inconsistency worth correcting, but it does not affect the State-32 state-rollout claim and is not the load-bearing issue. Overall, the verdict stays CONDITIONAL.","tokens_in":34208,"tokens_out":9766,"duration_ms":95637,"concrete_test":"Create a new validation split from a held-out portion of State-32 (e.g., a disjoint 10% of the training generator or a separate val set with the same profile), select the checkpoint on that split using the same objective, and recompute the ID rows of Table 1. If NewtonGS still beats Physics-Prior-SE(3) on Traj/FDE/Vel by similar margins, the ID concern is minor; if the margins shrink below 3.61%/2.57%/2.34% or reverse, the 'both splits' claim must be qualified to OOD only.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing problem is not the acknowledged real-scene gap but a circularity in the reported ID evaluation. Appendix B.2 states that 'val_id and val_ood are exactly the ID and OOD test sets reported in the main paper' and that 'Model selection uses the complete val_id objective'; Appendix A.6 confirms 'best checkpoint selected by the complete val_id objective.' With checkpoints saved at epoch 1 and every 5 epochs through 100, the checkpoint reported in the ID rows of Table 1 is chosen by minimizing the weighted state loss on the same sequences used to compute those ID metrics. Therefore the ID leg of the abstract's 'on both the ID and velocity-range-shift splits' claim is a selected, not out-of-sample, number. The OOD split is not used in selection, so the OOD leg remains valid. Because the claim is explicitly conjunctive, one supporting leg is methodologically compromised; this should be fixed by selecting on a genuinely held-out split and re-reporting Table 1.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces NewtonGS, a framework for object-level animation of static 3D Gaussian scenes. Each object is assigned a 22-dimensional dynamic state (position, quaternion orientation, linear/angular velocity, anisotropic scale and its rate, mass, restitution, and tangential attenuation). A structured hybrid model, Gaussian-NND, combines analytic translation, quaternion kinematics, gravity, damping, and scale-restoration dynamics with learned continuous and contact residuals; floor contact is treated as a discrete event. Predicted states are mapped back to Gaussian primitives through a shared affine transformation. The method is evaluated on two procedurally generated datasets: State-32 for state rollout and Gaussian-32 for state-to-Gaussian conversion. The main empirical claim is that NewtonGS achieves lower trajectory RMSE, final displacement error, and velocity RMSE than five analytic baselines on both the in-distribution and velocity-range-shift splits of State-32.","tokens_in":34499,"tokens_out":6469,"duration_ms":61706,"significance":"If the headline results survive scrutiny, the paper makes a useful contribution: it offers a compact, controllable object-level state for Gaussian scene animation and shows that a hybrid analytic-plus-residual dynamics model can outperform hand-specified analytic baselines on a large procedural benchmark. The manuscript is unusually transparent: it reports three-seed runs with small standard deviations, provides full procedural-generator specifications and exact evaluation formulas in the appendix, and candidly documents failure modes and the observation-to-state gap in Appendix G. These strengths are substantial. The main caveats are that the reported improvements over baselines are modest (roughly 2 to 5 percent relative error reduction), the comparison set contains no learned non-physics baseline, and the in-distribution evaluation is compromised by the checkpoint-selection protocol described in Appendix B.2. The real-scene gap is acknowledged by the authors and is not, by itself, a fatal flaw.","major_comments":[{"comment":"Appendix B.2 states that \"val_id and val_ood are exactly the ID and OOD test sets reported in the main paper\" and that \"Model selection uses the complete val_id objective,\" and Appendix A.6 confirms that the best checkpoint is selected by that objective. Because the ID rows of Table 1 are computed on val_id, the in-distribution half of the abstract's claim is a test-set-selected number rather than an out-of-sample result: the checkpoint is chosen to minimize the loss on the very sequences used to compute the reported ID metrics. This is a load-bearing methodological problem. Please introduce a genuine validation split disjoint from both test splits, select checkpoints on it, and re-report Table 1. The OOD comparison, which is not used for selection, remains informative.","section":"Appendix B.2 / §4.2"},{"comment":"The main-text affine map in Eq. (8), A_t = R_t diag(rho_t) R_0^T, is not the transform implemented in Appendix A.4, Eq. (19), Aimpl_t = R(q_t ⊗ q*_0) diag(rho_t). These coincide only for q0 = I or isotropic scale; for a general initial orientation they differ because the scale is applied in different frames. Since the state-to-Gaussian conversion is a stated contribution and is used in the Gaussian-32 experiments, the paper should either report results using the Eq. (8) transform or explicitly adopt the implemented version in the main text and explain the discrepancy.","section":"Eq. (8) vs. Appendix A.4, Eq. (19)"},{"comment":"The experiments compare NewtonGS only with hand-specified analytic baselines; no learned, non-physics baseline is trained on the same data and loss, and the paper itself lists \"learned baselines\" as future work. Because a central motivation is that the physics structure improves learned dynamics, a feedforward or recurrent trajectory model of comparable capacity is needed to establish that the structured prior, rather than the neural residual alone, is responsible for the reported gains. The post-hoc branch removal in Table 14 does not answer this question, since the remaining model is not independently trained without the branch.","section":"§4.4 and Conclusion"}],"minor_comments":[{"comment":"Appendix B.4 states that the Common-12 manifest contains 84 videos across seven methods, but Table 3 reports results for only five methods. Please either include the CogVideoX and NewtonGen rows or explicitly state why they are omitted from the table.","section":"Appendix B.4 / Table 3"},{"comment":"The sentence \"The smoothness term penalizes Lpen second temporal differences of position and, with relative weight 0.1, scale\" is garbled; it should read something like \"The smoothness term penalizes second temporal differences of position and, with relative weight 0.1, scale, and is included in Lsmooth.\"","section":"§3.5"},{"comment":"The caption describes the comparison as \"exploratory\" and notes that L4GM receives different inputs, which is appropriate; consider adding a sentence in the main text of §4.6 clarifying that the velocity comparison is not a controlled same-input comparison.","section":"Table 2 caption"}],"recommendation":"major_revision","confidential_remarks":"The val_id/test-set selection issue is the key methodological concern and must be fixed before the paper can be accepted. If the authors re-run checkpoint selection on a truly held-out split and the ID advantage persists, combined with the OOD result, I would support publication. The paper is transparent and technically detailed, but the contribution would be strengthened by a learned non-physics baseline and by resolving the Eq. (8)/Eq. (19) inconsistency."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about NewtonGS. The architecture is a clean, honest integration: a 22-dim object state, hybrid analytic-plus-residual dynamics with zero-init residuals, discrete floor contact, and a shared affine transform for Gaussian primitives. The experimental reporting is unusually candid, with Appendix G explicitly conceding that headline results start from labeled Z0, that observation-to-state inference is not evaluated, and that real multi-view scenes are future work. That candor is real and earns credit.\n\nThe bigger problem is one the paper doesn't concede. The in-distribution leg of the headline claim is selected on the test set. Appendix B.2 says val_id is exactly the ID test set reported in the main paper, and Appendix A.6 says the best checkpoint is chosen by the complete val_id objective. So the ID rows in Table 1 come from the same sequences used to select the checkpoint, not an out-of-sample number. The OOD split is not used in selection, so the OOD leg stands, and the architecture itself is not invalidated. But the abstract's \"both the in-distribution and velocity-range-shift splits\" is a conjunctive claim with one compromised leg. This needs a true held-out validation split and a re-reported Table 1.\n\nOther soft spots are smaller. Gains over the best analytic baseline are 2 to 5 percent relative, real but thin. There is no learned non-physics baseline, so we don't learn whether the physics prior is what buys the margin; the post-hoc branch removal in Appendix D is a reliance test, not a retrained ablation. The Common-12 video comparison is overclaimed in the main text, since Appendix C.2 calls it an unmatched proxy inventory. No code or data are linked.\n\nWhat's genuinely new: the package of an explicit object state, hybrid dynamics, and shared affine transform, plus a procedural benchmark with 32 motion families and exact state supervision. The per-motion results in Table 16 are fairly reported, including the classes where NewtonGS loses to the prior.\n\nThis deserves a serious referee, but the revision must fix the validation protocol and tone down the video-proxy language before it is citable. If you work on Gaussian animation, the state design and benchmark are useful regardless.","headline":"Solid, honest systems integration, but the ID leg of the headline claim is selected on the test set; the OOD leg holds and the architecture is still worth referee time.","tokens_in":34944,"tokens_out":2117,"would_cite":true,"duration_ms":21571,"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":"NewtonGS proposes that animating a static Gaussian scene reduces to rolling out one 22-dimensional physical state per object and applying the predicted pose and scale as a shared affine transform to all object Gaussians.","keywords":["3D Gaussian Splatting","object-level dynamics","neural ODE residuals","physics-structured learning","contact events","state-to-Gaussian transform","scene animation","quaternion kinematics"],"falsifier":"On a held-out set of real reconstructed Gaussian objects with automatically lifted states, compare NewtonGS's trajectory RMSE against the best analytic baseline over contact-rich motions; if the hybrid's advantage is not reproduced or the lifted states are unstable, the paper's central claim would be refuted. A more direct check is to feed the model a real video with unknown object state and see whether the predicted 22-dimensional rollout matches the observed motion.","tokens_in":34005,"feed_emoji":"🎬","tokens_out":5772,"duration_ms":51993,"temperature":0.7,"pith_summary":"NewtonGS argues that animating a static 3D Gaussian scene can be reduced to rolling out one compact physical state per object and then applying that state to the object's Gaussians with a single shared affine transform. The object state has 22 numbers: position, orientation quaternion, linear and angular velocity, anisotropic scale and scale rate, mass, restitution, and tangential attenuation. The dynamics model keeps the analytic Newtonian structure—gravity, damping, quaternion kinematics, scale restoration, and a discrete floor-contact bounce—and adds two small neural residual networks that start at zero and learn only the corrections. On a procedurally generated corpus of 32 motion families, the hybrid predicts positions, endpoints, and velocities with lower error than five hand-specified analytic baselines on both the in-distribution and velocity-range-shift splits. If the claim holds, the practical message is that editable object-level states, rather than per-Gaussian deformation fields, can be the control handle for Gaussian scene animation.","feed_headline":"A 22-number object state beats five analytic physics baselines","feed_subtitle":"Hybrid Newtonian-plus-neural dynamics rolls out pose, scale, and contact to animate Gaussian scenes.","key_machinery":"The load-bearing object is Gaussian Neural Newtonian Dynamics (Gaussian-NND), a 22-dimensional state per object combined with a continuous analytic prior, zero-initialized neural residuals, and a discrete contact map. The state's variables are chosen so that the same numbers that drive the ODE also define a deterministic affine map for all Gaussians in the object, separating translation, rotation, and anisotropic scale while preserving covariance positive semidefiniteness. RK4 integrates the hybrid field; a floor event outside the continuous solver handles instantaneous velocity changes. This machinery is what lets the model predict one trajectory per object rather than one trajectory per Gaussian primitive.","core_discovery":"The paper's central claim is that a physics-structured hybrid is the right way to predict object-level motion for Gaussian animation: fixed analytic dynamics alone underfit the synthetic motion families, while an unstructured network would sacrifice interpretability and control. Gaussian-NND therefore writes state evolution as a structured ODE with $\\dot p=v$, $\\dot q=\\tfrac12 q\\otimes[0,\\omega]$, gravity and damping in $\\dot v$, scale restoration in $\\dot u$, and a discontinuous horizontal-floor event that applies restitution and tangential attenuation, with learned residual terms $r_v,r_\\omega,r_s$ and an impulse correction $\\Delta v,\\Delta\\omega$ added at contacts. The predicted pose and scale are converted to a relative rotation-and-scale matrix $A_t=R(q_t\\otimes q_0^*)\\operatorname{diag}(s_t\\oslash s_0)$, and every Gaussian of the object updates as $g_i^t=A_t(g_i^0-p^0)+p^t$, $\\Sigma_i^t=A_t\\Sigma_i^0A_t^\\top$. The reported experiments show lower trajectory RMSE, final-displacement error, and velocity RMSE than five analytic baselines on State-32's in-distribution and velocity-range-shift splits, with the explicit caveat that these evaluations start from labeled initial states and do not test observation-to-state inference.","pith_inferences":["Editorial inference: if the same hybrid is applied to real reconstructed scenes, its success will hinge on the state-lifting step; the paper's own appendix flags that PCA lifting can flip axes or leak masks, so a natural test is to measure how much of the State-32 advantage survives automatic lifting.","Editorial inference: the 22-dimensional state could be extended to multi-object interaction by replacing the single floor event with pairwise contact and broad-phase collision, which the current model explicitly excludes.","Editorial inference: since State-32 targets are generated from closed-form formulas, a stronger stress test is to train the same architecture on a simulator with contact-rich, articulated, or non-analytic dynamics and compare against the analytic prior; the paper lists real scenes as future work.","Editorial inference: the shared affine transform means 'non-rigid deformation' in the benchmark covers only aggregate anisotropic scaling, so a genuine deformation field would require a larger state or per-part transforms."],"forward_implications":["Users could animate a reconstructed Gaussian object by editing roughly ten physical quantities—position, velocity, scale, mass, restitution, and friction—instead of specifying per-Gaussian motion.","Because the dynamics cost is independent of the Gaussian count, scenes with millions of primitives remain animatable with the same learned model.","The zero-initialized residual design gives a safe default: before learning, the rollout is exactly the analytic physics prior, so failures can be attributed to the prior rather than to a black-box network.","The rotation-and-scale affine update keeps covariance matrices valid, so rendered primitives stay coherent under large predicted motion.","The velocity-range-shift results suggest the learned corrections are not merely memorizing training speeds, although the benchmark remains synthetic."],"supporting_citations":[{"why":"Supplies the 3D Gaussian scene representation whose means and covariances are later transformed by the predicted affine map.","marker":"Kerbl et al., 2023"},{"why":"Provides the neural-ODE formulation that justifies learned continuous residual dynamics integrated with RK4.","marker":"Chen et al., 2018"},{"why":"Contributes neural event functions for ordinary differential equations, the basis for the discrete floor-contact event.","marker":"Chen et al., 2021"},{"why":"Motivates Neural Newtonian Dynamics as a structured physical state that guides generation, which NewtonGS adapts to 3D Gaussian scene animation.","marker":"Yuan et al., 2026"},{"why":"Fixes the unit-quaternion convention used for singularity-free orientation and the quaternion kinematic update.","marker":"Shoemake, 1985"},{"why":"Supplies the fixed-step RK4 integrator used for continuous rollout.","marker":"Hairer et al., 1993"}],"fun_headline_variants":["22-D state rolls Gaussian scenes, beats five physics baselines","Newtonian-plus-neural dynamics: object states animate Gaussians","Gaussian-NND hybrid outperforms five analytic motion models","State-32: structured state rollout wins over analytic baselines","Physics-structured state for Gaussian animation, lower errors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline numbers come from procedurally generated trajectories with the true initial state handed to the model, so the central premise is that such labeled object states—or a lifting procedure that recovers them—are available for real Gaussian objects.","fun_headline_variants_meta":{"raw":{"variants":["22-D state rolls Gaussian scenes, beats five physics baselines","Newtonian-plus-neural dynamics: object states animate Gaussians","Gaussian-NND hybrid outperforms five analytic motion models","State-32: structured state rollout wins over analytic baselines","Physics-structured state for Gaussian animation, lower errors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000538,"raw_usage":{"total_tokens":2637,"prompt_tokens":1053,"completion_tokens":1584,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":669,"completion_tokens_details":{"reasoning_tokens":1502}},"tokens_in":669,"tokens_out":1584,"duration_ms":12019,"temperature":1.0,"reasoning_tokens":1502,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:30:54.559239+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a held-out set of real reconstructed Gaussian objects with automatically lifted states, compare NewtonGS's trajectory RMSE against the best analytic baseline over contact-rich motions; if the hybrid's advantage is not reproduced or the lifted states are unstable, the paper's central claim would be refuted. A more direct check is to feed the model a real video with unknown object state and see whether the predicted 22-dimensional rollout matches the observed motion.","supporting_citations":[],"review_version":1}