{"id":"cea6cf55-8705-4d92-ba1c-854f39b403a7","arxiv_id":"2508.07775","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"A neural controlled differential equation model with smooth path filtering is proposed for extrapolating 3D rotation trajectories under non-conservative, noisy conditions.","lead":"This paper proposes a deep-learning method to predict the future rotation of moving objects by learning from noisy, sparse orientation measurements on the 3D rotation manifold. The approach is designed to work even when forces are not conserved, which could help robotics, augmented reality, and other systems that track spinning or tumbling objects.","discovery_kind":"unclear","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The submitted full text is a different paper (CT denoising), so the SO(3) forecasting claims have no supporting methods or experiments and cannot be verified.","rationale":"The reader correctly identified that the submission is internally inconsistent: the abstract describes SO(3) forecasting while the full text is about CT denoising. This is the most load-bearing concern because without the actual methods and experiments, the abstract's claims cannot be evaluated. The reader's stated weakest_assumption about smoothing preserving true dynamics is a plausible secondary concern—if the full text were present, that premise would indeed be critical—but it is not the primary reason the paper is unverdictable. I partially agree with the reader: the same root cause is identified, but the load-bearing issue is the missing content, not the specific smoothing assumption. My verdict remains UNCHANGED because the appropriate disposition is already UNVERDICTED; no new information moves it. The honest outcome is that the central claim is unsupported by the submitted manuscript, and the concrete test is to obtain and inspect the real arXiv paper to see whether the claims survive. This is not an ad hominem criticism; it is a straightforward observation about the evidence provided.","tokens_in":3161,"tokens_out":2712,"duration_ms":29335,"concrete_test":"Retrieve the actual arXiv:2508.07775 full text (e.g., from arXiv) and check whether it contains: (1) a precise definition of the SO(3) Savitzky-Golay path and a proof or analysis of its noise-reduction and dynamics-preservation properties; (2) a formulation of the Neural CDE on SO(3) with well-defined vector fields and a method for training from noisy pose states; (3) experiments that compare extrapolation against constant-velocity/energy-conservation baselines under non-conservative forces and show generalization to unseen physical parameters. If any of these components is missing or does not support the abstract's claims, the central claim remains unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in the abstract asserts robust SO(3) extrapolation via Neural CDEs guided by SO(3) Savitzky-Golay paths, agnostic to energy/momentum conservation and generalizing to unknown physical parameters. However, the provided full text is arXiv:2508.07788, an entirely different paper on low-dose CT denoising. There is no description of the SO(3) method, no derivation of the Neural CDE, no definition or analysis of the SO(3) Savitzky-Golay path, and no experiments on rotational dynamics. Consequently, every element of the central claim is unsupported by the submitted manuscript. This is not a flaw in the proposed approach itself but a fundamental evidentiary gap: the technical content needed to assess correctness, novelty, or robustness is absent. Even the abstract's internal consistency cannot compensate for the missing derivations and empirical validation. The paper as submitted is therefore unverifiable, and any scientific verdict would be speculation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript, as identified by its arXiv number and abstract, claims a method for forecasting continuous non-conservative dynamical systems on SO(3). The proposed approach is said to combine Neural Controlled Differential Equations with SO(3) Savitzky-Golay smoothed paths, to be agnostic to energy and momentum conservation, robust to input noise, and capable of generalizing to unknown physical parameters. However, the supplied full text is a completely different paper on low-dose CT denoising (ALDEN), with no equations, derivations, experiments, or algorithmic details related to SO(3) forecasting. Consequently, the technical content needed to assess the central claims is absent from the submitted manuscript.","tokens_in":3355,"tokens_out":1335,"duration_ms":17477,"significance":"If the claims in the abstract were substantiated, the work could be significant for computer vision and robotics: robust SO(3) extrapolation without conservation-law assumptions would address a real gap in current methods, and the proposed module-level integration would be practically attractive. The stated code availability is a positive sign, though the code is not part of the submitted text. However, significance cannot be evaluated without the actual method and experiments. As submitted, the manuscript provides no derivations, no architectural details, no training procedure, no evaluation protocol, and no results. Therefore the potential significance is real but entirely unverified.","major_comments":[{"comment":"The submitted full text is a different paper: 'Anatomy-Aware Low-Dose CT Denoising via Pretrained Vision Models and Semantic-Guided Contrastive Learning' (arXiv:2508.07788), concerning CT denoising. None of the claimed SO(3) forecasting content appears. There is no description of Neural CDEs, no definition or analysis of SO(3) Savitzky-Golay paths, no derivation of the dynamics model, no training procedure, and no experiments on rotational motion. Every load-bearing element of the abstract's central claim is therefore unsupported by the manuscript as submitted. This is not a presentation issue; it is a fundamental evidentiary gap that prevents any verification of correctness, novelty, or robustness.","section":"Full Text (entire manuscript)"},{"comment":"The abstract asserts that the approach is 'agnostic to energy and momentum conservation' and 'generalizes well to trajectories with unknown physical parameters.' These are strong claims that require a formal model of non-conservative forces, a definition of what 'agnostic' means mathematically, and an experimental protocol that varies inertial parameters and external torques. None of this is present. Without the method section or experiments, these claims are unverifiable and cannot be checked for internal consistency or circularity.","section":"Abstract (lines 1-6)"},{"comment":"The abstract mentions 'simulation and various real-world settings,' yet the supplied full text contains no experimental section, no datasets, no baselines, no metrics (e.g., geodesic error on SO(3)), and no comparisons to existing extrapolation methods. The claimed robustness to input noise and generalization to unknown physical parameters is not demonstrated anywhere. Even if the correct full text were supplied, the abstract alone would be insufficient; the current submission provides no empirical evidence whatsoever.","section":"Full Text (no experiments section)"}],"minor_comments":[{"comment":"The title and abstract identify the paper as 'Forecasting Continuous Non-Conservative Dynamical Systems in SO(3)', but the full text has a different title, different authors, and a different subject area. This is a severe mismatch that should be corrected at the source; the arXiv submission appears to have been overwritten or misassociated.","section":"Title and Metadata"},{"comment":"The abstract uses terms such as 'Neural Controlled Differential Equations' and 'SO(3) Savitzky-Golay paths' without definitions or references. In a standalone abstract this may be acceptable, but with no full text these terms cannot be contextualized. If the paper is resubmitted, please include proper references to prior CDE work and Savitzky-Golay filtering on manifolds.","section":"Abstract"},{"comment":"The reference list is entirely from the CT denoising paper and does not include any citations to relevant SO(3) forecasting, Neural CDE, or manifold smoothing literature. This further confirms that the full text does not correspond to the abstract.","section":"References"}],"recommendation":"reject","confidential_remarks":"The manuscript as submitted is scientifically unassessable because the full text is a different paper. This is not a matter of disputed interpretation or an internal inconsistency that could be fixed by revision; the core technical content is simply missing. The arXiv record appears to have a mismatch between the abstract and the PDF. I recommend rejection of this submission, with the possibility that the authors re-submit the correct paper containing the actual SO(3) method. The code link in the abstract may be useful, but it is not part of the submitted text and cannot substitute for the missing manuscript. Please verify the integrity of the submission process."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: the submitted PDF is a different paper. The abstract describes forecasting non-conservative dynamics in SO(3) with Neural CDEs and Savitzky-Golay paths; the full text is an anatomy-aware low-dose CT denoising paper (arXiv:2508.07788). Nothing in the attached text describes the SO(3) method, its equations, experiments, or comparisons. This is not a fixable weakness in the science; it is a missing manuscript, and that is the dominant fact of this submission.\n\nCredit where it is due: the abstract names a real gap. Extrapolating noisy pose trajectories on SO(3) without assuming constant velocity or energy conservation matters for vision and robotics, and using SO(3) Savitzky-Golay smoothing as a guiding signal for a Neural CDE is a plausible combination. The authors also point to public code. If the actual paper delivers what the abstract promises, it could be a useful module for non-inertial object tracking.\n\nThe soft spots are large. The abstract contains no citations to prior manifold forecasting work, no quantitative results, and no derivations—so even on its own it would be hard to evaluate. But the full-text mismatch makes everything unverifiable: the definition of the SO(3) Savitzky-Golay path, the handling of non-conservative torques, the claimed generalization to unknown physical parameters, and the empirical claims in simulation and real-world settings. The reader's take and the stress-test note are both right: this is an evidentiary gap, not a demonstration of a flaw in the approach.\n\nRecommendation: desk reject this submission. Send it back for a corrected PDF if this was a file mix-up; the idea is worth a look. But no referee can review the SO(3) claims on the basis of an abstract alone with a CT paper attached. If the correct manuscript shows up, I would then send it to someone who works on Riemannian manifold learning or trajectory forecasting.","headline":"The submission is two different papers: the abstract is a plausible SO(3) forecasting method, but the full text is a CT denoising paper, so there is nothing to peer review.","tokens_in":3831,"tokens_out":3227,"would_cite":false,"duration_ms":35656,"reading_group":"no","serious_thinker":"unclear","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that rotational trajectory forecasting on $SO(3)$ can be made resilient to noise and generalizable to non-conservative, non-inertial systems by conditioning a Neural CDE on $SO(3)$-valued Savitzky-Golay smoothed paths.","keywords":["SO(3)","neural controlled differential equations","Savitzky-Golay smoothing","rotational motion forecasting","non-conservative dynamics","noisy pose estimation","extrapolation"],"falsifier":"Generate a rigid body tumbling under a known non-conservative torque, add noise to short observed segments, train the model on those segments only, then extrapolate beyond the training horizon and compare with the ground-truth rotation. If the forecast error matches a constant-velocity baseline once the noise is smoothed away, the claimed advantage of the guided CDE for non-conservative dynamics would be falsified.","tokens_in":3055,"feed_emoji":"🔄","tokens_out":6250,"duration_ms":63546,"temperature":0.7,"pith_summary":"The paper presents a method for extrapolating the rotational motion of rigid objects in 3D when observations are noisy and the dynamics need not conserve energy or momentum. Existing $SO(3)$ forecasting methods often assume constant velocity or energy conservation; this work drops those assumptions by pairing a neural controlled differential equation with Savitzky-Golay smoothed paths on the rotation manifold. The authors argue that this combination produces forecasts that tolerate input noise and generalize to unknown physical parameters in both simulation and real-world settings. If correct, it would make rotational trajectory forecasting applicable to tumbling, driven, or otherwise non-inertial objects rather than only free-spinning ones.","feed_headline":"Forecast 3D rotation paths without energy conservation assumptions","feed_subtitle":"Learns non-conservative motion from noisy poses, so objects under external forces and torques can be extrapolated.","key_machinery":"The central object is the Neural Controlled Differential Equation (Neural CDE) on the rotation manifold $SO(3)$, driven by an $SO(3)$-valued Savitzky-Golay path. The Savitzky-Golay path is a locally polynomial smoothing of the noisy rotation observations, adapted to the manifold, and it provides a differentiable control signal encoding local rotational dynamics without assuming a global conservation law. The CDE learns to continue this signal, so the extrapolator is agnostic to energy and momentum while remaining grounded in the geometry of rotations.","core_discovery":"The central claim is that a Neural Controlled Differential Equation driven by an $SO(3)$-adapted Savitzky-Golay path can extrapolate rigid-body rotations under external torques and non-conservative forces, without relying on energy or momentum conservation. The guided path supplies a stable, differentiable control signal in which noise has been smoothed on the manifold, and the CDE learns the vector field that continues that path. The authors report that the model generalizes to trajectories with unknown physical parameters and is robust to input noise, while being easy to insert into existing pose-estimation and tracking pipelines.","pith_inferences":["If the noise-robustness claim holds, a natural extension is 6-DOF forecasting by coupling this $SO(3)$ rotation forecast with a separate translation model; the product-manifold structure would require a joint smoothing scheme rather than treating rotation alone.","The Savitzky-Golay window size and polynomial order become hyperparameters that likely interact with noise level and rotational speed; an adaptive or learned smoothing scale would be a testable improvement.","The same guided-CDE design should apply to other Lie-group trajectories such as $SE(3)$ or unit quaternions whenever a manifold-adapted smoothing exists, not only to pure rotations.","Because the method avoids conservation-law assumptions, it may also extend to human motion or articulated rotation, where internal muscle torques make dynamics strongly non-conservative."],"forward_implications":["Rotational extrapolation pipelines can replace constant-velocity or energy-conserving priors with a learned CDE module, extending forecasts to objects under external forces and torques.","Noisy pose inputs become usable for forecasting because manifold-aware smoothing is built into the driving path before extrapolation begins.","The method transfers to new trajectories with unknown physical parameters, indicating the learned dynamics are not tied to a specific inertia tensor or torque profile.","As a modular component, it can be added to existing pose-estimation and tracking systems without redesigning the rest of the pipeline.","Explicit estimation of physical quantities such as moment of inertia is no longer required for forecasting rotational motion, since the model learns dynamics directly from noisy states."],"supporting_citations":[],"fun_headline_variants":["Predict 3D rotations under external forces without energy assumptions","Extrapolate noisy SO(3) motion with Neural CDEs","Forecast non-conservative rotations from noisy pose data","Learn rotational dynamics without energy or momentum assumptions","Robust SO(3) extrapolation for non-inertial systems"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that the $SO(3)$-adapted Savitzky-Golay smoothing removes observation noise without distorting the true underlying rotation, and that a model trained on such smoothed noisy states extrapolates correctly to unseen external torques and unknown inertia.","fun_headline_variants_meta":{"raw":{"variants":["Predict 3D rotations under external forces without energy assumptions","Extrapolate noisy SO(3) motion with Neural CDEs","Forecast non-conservative rotations from noisy pose data","Learn rotational dynamics without energy or momentum assumptions","Robust SO(3) extrapolation for non-inertial systems"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000521,"raw_usage":{"total_tokens":2354,"prompt_tokens":734,"completion_tokens":1620,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":478,"completion_tokens_details":{"reasoning_tokens":1537}},"tokens_in":478,"tokens_out":1620,"duration_ms":14113,"temperature":1.0,"reasoning_tokens":1537,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T21:50:18.719373+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a rigid body tumbling under a known non-conservative torque, add noise to short observed segments, train the model on those segments only, then extrapolate beyond the training horizon and compare with the ground-truth rotation. If the forecast error matches a constant-velocity baseline once the noise is smoothed away, the claimed advantage of the guided CDE for non-conservative dynamics would be falsified.","supporting_citations":[],"review_version":1}