{"id":"944efe96-929e-4550-ae83-59a13ed8ded1","arxiv_id":"2512.09423","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"FunPhase encodes motion clips as sinusoidal phase functions and decodes them continuously in space and time, enabling reconstruction, generation, super-resolution, and body completion across skeletons.","lead":"This paper describes a machine learning model that compresses human and animal motion into a small set of sine-wave parameters, then uses those parameters to rebuild, generate, or interpolate movement as a continuous function. The authors claim at least 45% uniform error reduction over older phase-based motion models, but their own tables contradict that number on several metrics.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The abstract's 'uniform improvements of at least 45% across all metrics' is contradicted by Table 1: relative gains on DOG orientation (~37%), DOG ACL (~40%), and 100STYLE ACL (~9%) all fall below 45%.","rationale":"The reader's conditional verdict is well-founded. My stress-test focuses on the one issue that can be settled immediately from the manuscript itself: the abstract's uniform ≥45% improvement claim is arithmetically inconsistent with Table 1 for ACL on 100STYLE (~9% relative reduction) and for orientation/ACL on DOG (~37%/~40%). No external code or new experiment is needed to see this; it directly affects the headline contribution. I do not find an internal inconsistency in the architecture: the single-sinusoid-per-channel constraint in Eq. 1 is a real expressivity limit, but the decoder's query-time positional embeddings and nonlinear cross-attention can in principle represent time-varying structure, so a claim that transitions cannot be represented would need its own experiment rather than being assumed. The reader's other concerns — no error bars, qualitative controller, missing code, non-monotonic 256-channel orientation — are secondary and correctable. Therefore the verdict remains conditional: the method may be sound, but the reported claims must be revised and the key numbers verified before acceptance.","tokens_in":15576,"tokens_out":11183,"duration_ms":109981,"concrete_test":"Recompute the arithmetic in Table 1: for each metric on both DOG and 100STYLE, calculate (DeepPhase − FunPhase)/DeepPhase. If any relative reduction is below 45%, the abstract's uniform-improvement claim fails as stated. To make this a substantive empirical check, recompute Table 1 from the provided checkpoints with at least 3 seeds and report per-seed values and confidence intervals; if the intervals straddle 45% on orientation or ACL, the claim must be restated as dataset- and metric-specific rather than uniform.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim of the paper — 'uniform improvements of at least 45% across all metrics' over prior periodic autoencoders — is not supported by the paper's own Table 1. Computing relative improvement as (DeepPhase − FunPhase)/DeepPhase for each metric: on DOG, DeepPhase-16C vs FunPhase-16C gives orientation 0.54→0.34 (37% reduction) and ACL 1.768→1.062 (40%); on 100STYLE, DeepPhase-32C vs FunPhase-32C gives ACL 1.504→1.371 (8.8%), and even the 256-channel variant gives ACL 1.504→1.378 (8.4%). Thus the uniform-45% statement is false for at least two metrics regardless of channel count. This matters because the abstract frames the contribution around these uniform large gains; if the intended comparison is selective or metric-specific, the text must say so and the numbers must be recomputed. The single-sinusoid stationarity assumption (Eq. 1) is a related representational limitation, but the immediate load-bearing issue is the claim/report mismatch in the headline result.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces FunPhase, a functional periodic autoencoder that represents motion as a continuous spatio-temporal function. Each latent channel is modeled as a single sinusoid with learned amplitude, frequency, phase shift, and offset, following the DeepPhase periodic autoencoder but replacing the frame-based convolutional decoder with a Perceiver-based function decoder that can decode at arbitrary temporal and joint resolutions. The latent parameters are transformed (Cartesian phase, probit frequency, raw offset) and used to train a DiT-based latent diffusion model for class-conditional motion generation. The method is evaluated on reconstruction against DeepPhase and other autoencoders, on generation against MLD, ACMDM, CAMDM, AnyTop, and SinMDM, and on qualitative motion-controller and super-resolution/completion tasks. The paper claims substantial and uniform reconstruction improvements over prior periodic autoencoders and competitive generation quality.","tokens_in":15888,"tokens_out":5966,"duration_ms":54620,"significance":"If the results hold, FunPhase would be a valuable contribution: it unifies reconstruction, generation, super-resolution, and partial-body completion in a single interpretable and skeleton-agnostic phase manifold. The functional formulation, Perceiver-based architecture, and diffusion over phase-transformed parameters are well motivated and clearly described. The empirical evaluation includes multiple baselines, ablations, and downstream tasks. The paper also provides honest limitations, though the core claims need tightening to match the reported numbers.","major_comments":[{"comment":"The abstract states 'uniform improvements of at least 45% across all metrics' over prior periodic autoencoders. Table 1 does not support this. Relative improvements (DeepPhase to FunPhase) are: DOG Orientation 0.54→0.34 (37%), DOG ACL 1.768→1.062 (40%), 100STYLE ACL with 32C 1.504→1.371 (8.8%), and with 256C 1.504→1.378 (8.4%); 100STYLE Orientation with 256C is 0.35→0.20 (43%). Thus at least two metrics fail the 45% threshold for every compared variant. The claim must be corrected to reflect metric-specific or 'most metrics' improvements, or the comparison must be recomputed.","section":"Abstract and Table 1"},{"comment":"The text says 'Notably, despite operating in a more compact latent space imposed by the periodic parameterization, FUNPHASE still surpasses the Function Autoencoder in reconstruction accuracy.' This is not uniformly true. Comparing FunPhase-256C with Function AE-256C: Position improves (0.36 vs 0.74), but Orientation is worse (0.20 vs 0.05) and NPSS is worse (0.75 vs 0.59). The claim should be qualified to specific metrics (e.g., position and physical plausibility) or the numbers must be revisited.","section":"Section 5.1, Table 1"},{"comment":"The model assumes each latent channel is a single sinusoid with constant amplitude, frequency, phase shift, and offset over the entire 1-second window. This is a stationarity assumption that cannot represent non-stationary segments such as transitions, speed changes, or contact events within a window. The limitation section only acknowledges loss of detail at very low sampling rates, not this representational restriction. Given the paper's claims of generality across skeletons and datasets, the authors should either supply an analysis of reconstruction error on non-periodic or transitional motions, or explicitly scope the method to approximately periodic locomotion.","section":"Section 4.1, Eq. (1); Section 5.4"}],"minor_comments":[{"comment":"The dataset name is typeset as '100STYLES' (missing space and extra S). Also 'MLD-V AE' and 'ACMDM-AE' have inconsistent spacing.","section":"Table 1"},{"comment":"The column 'ACCL' is likely meant to be 'ACL' (Average Curve Length) as used in Table 1. Please make consistent.","section":"Table 2"},{"comment":"The conditioning latent is written 'DF unP hase' in the equation; likely a typo for 'FunPhase'.","section":"Section 7.2 (Supplementary)"},{"comment":"The text refers to 'Table 1 of the main paper' when comparing Function Diffusion and FunPhase; the relevant table is Table 2.","section":"Section 8.3 (Supplementary)"},{"comment":"The sentence 'this results in clear gains across all metrics' is too strong. In Table 2, FunPhase has slightly worse Coherence and ACCL than Function Diff, and worse Diversity than MLD and CAMDM. Please qualify.","section":"Section 5.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's main technical contribution seems sound, but the abstract's 'uniform improvements of at least 45%' claim is contradicted by the paper's own Table 1. This is a load-bearing quantitative claim that will be noticed by any careful reader. I recommend major revision to correct the claim and to qualify the comparisons that are not uniformly positive. The stationarity assumption also deserves more explicit treatment, as it bears on the generality claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"FunPhase is a legitimate step forward: it takes DeepPhase's periodic latent idea, makes the decoder functional via Perceiver cross-attention, and adds diffusion on the phase parameters. That combination is new as far as the references show, and it buys real practical breadth — skeleton-agnostic training, arbitrary temporal sampling, super-resolution, partial-body completion — with quantitative gains over DeepPhase on most metrics.\n\nThe main problem is the abstract's headline claim. The version I read says 'uniform improvements of at least 45% across all metrics.' Table 1 doesn't support that: on DOG, orientation improves 37% and ACL 40%; on 100STYLE, ACL improves 8.8% for the 32-channel variant and 8.4% for 256 channels. So the uniform-45% statement is simply wrong. If that phrasing has been removed in the final version, fine, but the claim was live in the abstract used for the version under discussion and must be retracted or recomputed.\n\nOther soft spots are minor to moderate. The 256-channel variant having higher orientation error than the 32-channel one (0.20 vs 0.10) is odd and unexplained. No error bars appear on the reconstruction numbers. The phase-transformation ablation in the supplement shows essentially no FID gain (1.28 vs 1.27), even though the paper credits much of its generation quality to that transformation — the bigger gains come from the phase formulation itself, so the text should be more careful about which component is doing what. The single-sinusoid-per-channel-per-window assumption is a real representational limit; the limitation section concedes loss of fine detail at low sampling rates but doesn't test the stronger version head-on. Code is promised only 'upon publication,' which is normal but not ideal.\n\nNone of this kills the architecture. The external comparisons against CAMDM, ACMDM, MLD, AnyTop, and SinMDM are credible, the method is clearly specified, and the central claim — that a functional decoder plus phase manifold works well across tasks — holds up. This paper deserves a serious referee. If I were the editor I'd send it out and ask the authors to fix the abstract, compute the numbers honestly, add error bars where feasible, and address the non-monotonicity and the phase-ablation wording.","headline":"A real architectural step forward in phase-based motion models, but the abstract's 'uniform at least 45%' claim is contradicted by the paper's own Table 1.","tokens_in":16379,"tokens_out":2855,"would_cite":true,"duration_ms":27248,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"FunPhase models each motion clip as a handful of sinusoids and reconstructs motion as a continuous function, cutting reconstruction error by at least 45% versus prior periodic autoencoders.","keywords":["motion generation","phase manifold","periodic autoencoder","functional representation","latent diffusion","skeleton-agnostic","motion super-resolution","motion completion"],"falsifier":"Take a real 1-second motion clip that contains a clear gait transition, such as a walk-to-run change within the window, encode it with FunPhase, and measure the reconstruction error at the transition frames versus steady-state frames. If the single-frequency-per-channel model is truly capturing motion structure, the error should spike at the transition, revealing the representation's blind spot for non-stationary content.","tokens_in":15427,"feed_emoji":"🏃","tokens_out":2237,"duration_ms":25098,"temperature":0.7,"pith_summary":"The paper introduces FunPhase, a periodic functional autoencoder that turns motion into a continuous spatio-temporal function rather than a sequence of frames. Its central claim is that decomposing each latent channel into a single sinusoid—phase, amplitude, frequency, and offset—yields a compact, interpretable phase manifold that supports reconstruction, generation, super-resolution, body completion, and phase-aligned control in one skeleton-agnostic framework. The authors argue that this phase-structured function space acts as a strong inductive bias, reducing reconstruction error substantially compared to prior periodic autoencoders while matching state-of-the-art motion generation quality. A sympathetic reader would care because the method promises to unify previously separate tasks in motion learning and to make phase-based representations compatible with modern generative diffusion models.","feed_headline":"Motion as a function: sinusoidal latents cut error 45%","feed_subtitle":"A periodic autoencoder replaces frame-by-frame decoding with continuous phase functions, unifying generation, control, and super-resolution","key_machinery":"The load-bearing mechanism is the channel-wise sinusoidal parameterization of the latent space: every latent channel is modeled as a_c * sin(2π(f_c * T − s_c)) + b_c, with parameters extracted by a differentiable FFT and a learned phase regressor. This single-frequency-per-channel representation enforces periodicity and creates a smooth phase manifold, which the decoder queries as a continuous function. The accompanying domain transformations (phase-to-Cartesian, probit on frequency) make these parameters suitable for Gaussian diffusion, while the Perceiver encoder/decoder pairs handle variable skeleton topology via graph positional encodings.","core_discovery":"The paper's core claim is that motion can be encoded as a continuous function of time and skeleton position using a periodic autoencoder whose latent space is parameterized by per-channel sinusoids. Each latent channel carries four scalars—phase shift, amplitude, frequency, and offset—obtained via a differentiable FFT layer, and the decoder evaluates these sinusoids at arbitrary spatio-temporal coordinates through Perceiver-based cross-attention. This functional formulation replaces discrete frame decoding, enabling arbitrary temporal resolution, generalization across skeletons and datasets, and probabilistic generation via latent diffusion performed directly on the phase parameters. The aut","pith_inferences":["If single-sinusoid channels suffice for a one-second window, the strongest implication is that local periodicity is the dominant temporal structure in atomic motion segments; an editorially natural test is whether modeling two or three sinusoids per channel (or allowing frequency drift) would unlock longer, non-stationary windows without losing interpretability.","The functional decoder could be coupled with physics-based priors (contact forces, dynamics) more naturally than frame-based decoders, since the output is a differentiable function of time and joint coordinates; this is an extension the paper names as future work.","The phase manifold's circular structure might support direct interpolation and blending in latent space (e.g., style transfer by adding phase vectors), similar to how phase manifolds helped motion matching in earlier work; the paper only hints at control via a Mixture-of-Experts controller.","A concrete falsifiable extension of the paper's logic: if the sinusoidal bottleneck captures true motion structure, then reconstruction error should degrade sharply when the same clip contains two distinct gaits within one window; testing this would separate the power of the representation from the capacity of the Perceiver decoder."],"forward_implications":["Motion reconstruction, prediction, generation, super-resolution, and partial-body completion can share one learned phase manifold, simplifying pipelines that currently require separate models.","Because decoding is a continuous function evaluated at query coordinates, the same trained model can output motion at any frame rate or skeleton resolution without retraining.","Phase-structured latents stabilize latent diffusion: the paper reports lower FID and better physical plausibility than diffusion without the phase bottleneck, suggesting periodicity is a useful generative prior for motion.","The compact 4-parameter-per-channel representation enables skeleton-agnostic training across human and animal datasets, opening the door to cross-species motion transfer and control."],"fun_headline_variants":["Continuous motion latents cut error 45%","Phase functions fuse motion generation and prediction","Sinusoidal latents sample motion at any speed","Functional phase autoencoder: 45% lower error","Motion as continuous functions: 45% less error"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The claim rests on assuming that every one-second motion window can be faithfully represented as a fixed sum of independent single-frequency sinusoids, one per latent channel, with constant amplitude, frequency, phase, and offset throughout the window.","fun_headline_variants_meta":{"raw":{"variants":["Continuous motion latents cut error 45%","Phase functions fuse motion generation and prediction","Sinusoidal latents sample motion at any speed","Functional phase autoencoder: 45% lower error","Motion as continuous functions: 45% less error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000153,"raw_usage":{"total_tokens":1010,"prompt_tokens":676,"completion_tokens":334,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":420,"completion_tokens_details":{"reasoning_tokens":262}},"tokens_in":420,"tokens_out":334,"duration_ms":4424,"temperature":1.0,"reasoning_tokens":262,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T17:25:46.536290+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a real 1-second motion clip that contains a clear gait transition, such as a walk-to-run change within the window, encode it with FunPhase, and measure the reconstruction error at the transition frames versus steady-state frames. If the single-frequency-per-channel model is truly capturing motion structure, the error should spike at the transition, revealing the representation's blind spot for non-stationary content.","supporting_citations":[],"review_version":1}