{"id":"2bbc6b1f-94ed-4d54-9b8b-4564d6e55551","arxiv_id":"2411.17532","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"FTMoMamba achieves FID 0.181 on HumanML3D by injecting frequency-domain features into the state transition matrix and text features into the output matrix of a Mamba-based diffusion denoiser.","lead":"A new diffusion-based system for generating human motion from text, FTMoMamba, splits motion into low- and high-frequency parts and injects text features directly into the state-space model's matrices. It reports the lowest FID score yet on the HumanML3D benchmark, 0.181, but leaves several implementation details under-specified.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equations (9) and (15) are dimensionally undefined: DWT-derived f_low/f_high and CLIP f_t cannot be added to A and C without projection, so the claimed FreqSSM/TextSSM mechanism is not verifiable as written.","rationale":"The reader's weakest assumption correctly identifies the dimensional ambiguity in Equations (9) and (15) as the load-bearing weak point. The paper's stated contribution relies on modifying the SSM matrices A and C with frequency and text features, yet the shapes of f_low, f_high, and f_t are incompatible with A and C under the standard Mamba parameterization. Without a specified projection or code, the equations cannot be executed as written, meaning the reported FID of 0.181 may come from an implementation that differs from the described mechanism. Additional inconsistencies, such as the claim that FID is reduced by 5.1% relative to MotionGPT (0.232 to 0.181 is 0.051 absolute, not 5.1% relative) and the similar 2.8% wording in the frequency ablation, reinforce the need for exact experimental artifacts. The verdict remains conditional because the core result could survive if the authors provide shape-compatible definitions, release code, and confirm that the FID gain persists when the claimed matrix injections are actually used. The proposed concrete test directly settles whether the equations describe the implemented model.","tokens_in":11774,"tokens_out":3674,"duration_ms":36164,"concrete_test":"Ask the authors to release a minimal forward trace (or code) for FreqSSM/TextSSM, and run a shape assertion: instantiate f_m^mo in R^{B x L x D}, A in R^{D x N}, C in R^{B x L x N}, then execute Eq. (8)-(9) and (15) without any extra projection. If the addition fails or requires broadcasting, the mechanism as written is not implemented. If the released code contains projection layers, compare FID after ablating those projections; if the FID change exceeds the reported +/-0.009 confidence interval, the frequency/text-in-SSM attribution in the paper is not supported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that FTMoMamba improves FID by embedding frequency features into the SSM transition matrix and text features into the output matrix. Equations (8)-(9) and (15) are the exact sites of that claim. In a standard Mamba layer, A has shape (D,N) or (N,), C has shape (B,L,N), and DWT applied along the sequence length of f_m^mo yields f_low,f_high with length L/2 (and channel dimension equal to the hidden width). Equation (9) therefore requires adding a (L/2, D) quantity to A, which has no N dimension; Eq. (15) requires adding f_t in R^{1x256} to C, which is either a parameter or a (B,L,N) tensor. No projection, reshaping, or broadcasting rule is given. This is not a cosmetic issue: if the implementation inserts unseen projection layers or replaces the additions with concatenation/residual branches, then the stated mechanism 'frequency into A, text into C' is not what is being evaluated, and the FID gain cannot be attributed to FreqSSM/TextSSM as described. The paper also provides no code, so the discrepancy cannot be resolved by inspection.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FTMoMamba, a latent diffusion framework for text-to-motion generation that augments a Mamba-based denoiser with two modules: FreqSSM, which injects low- and high-frequency features (obtained via DWT) into the SSM state-transition matrix A, and TextSSM, which injects a CLIP sentence embedding into the SSM output matrix C. The authors report the lowest FID (0.181) on HumanML3D among the compared methods, alongside ablations of the two modules and of frequency-extraction choices. The central claim is that frequency information in A improves static-pose and fine-grained motion generation, and that sentence-level text in C improves text-motion alignment.","tokens_in":12069,"tokens_out":5529,"duration_ms":94842,"significance":"If the reported FID result is reproducible and the FreqSSM/TextSSM mechanisms operate as stated, the work would be of interest to the motion-generation community as a lightweight Mamba-based alternative to transformer denoisers. The paper includes standard benchmark comparisons, ablations, and qualitative visualizations, and it explicitly reports confidence intervals for all metrics. However, the load-bearing equations for the two core modules are dimensionally ambiguous, the ablation narrative contradicts the reported table values, and several numerical claims in the text do not match the tables. The manuscript's contribution, as written, is therefore not fully verifiable and will require substantial revision before its claims can be assessed.","major_comments":[{"comment":"The additions in Eq. (9) and Eq. (15) are dimensionally undefined. In a standard Mamba layer A is a parameter of shape (D,N) or (N,), while DWT applied along the sequence length of a latent feature f_m^mo yields f_low and f_high of length L/2 with channel dimension equal to the hidden width; these cannot be added to A without a projection or reshaping rule. Similarly, in Eq. (15) the text embedding f_t ∈ R^{1×256} is added to C, which is either a parameter or a (B,L,N) tensor; no broadcasting or projection is specified. The paper provides no code or implementation details to resolve the discrepancy. Since Eqs. (8)-(9) and (15) are the exact sites of the claimed 'frequency into A, text into C' mechanism, the mechanism as stated is not implementable, and the FID gain cannot be attributed to FreqSSM/TextSSM as described.","section":"§3.2, Eqs. (8)-(9) and (15)"},{"comment":"The notation for the output features is inconsistent. f_m^no is defined in Eq. (7) as the gated output of the FreqMamba branch, but Eq. (11) also defines f_m^no as the FreqSSM output before the linear/gating stage; likewise f_m^vo is defined in both Eq. (14) and Eq. (17), and Eq. (18) then uses both symbols. This makes the forward pass ambiguous: it is unclear whether the final feature z_t' includes the linear/gated outputs, the raw SSM outputs, or both. The equations need to be aligned or renamed before the architecture can be reproduced.","section":"§3.2, Eqs. (7), (11), (14), (17)-(18)"},{"comment":"The ablation discussion is inconsistent with Table 2. The text states that 'TextMamba outperforms FreqMamba in terms of lower FID and greater diversity,' but the row with FID 0.209 (which presumably corresponds to FreqMamba-only, given the sentence order in the paragraph) has both lower FID and higher diversity than the row with FID 0.227 (presumably TextMamba-only). The table also does not show which module is enabled in each row because the checkmark symbols are missing from the FreqMamba/TextMamba columns. The comparison and its interpretation need to be corrected or explicitly labeled.","section":"§4.5, Table 2"},{"comment":"Several percentage claims do not match the tables. In §4.4, the FID is said to be 'reduced by 5.1%' compared to MotionGPT, but Table 1 gives (0.232−0.181)/0.232 ≈ 22%. In §4.5, the combined low/high-frequency configuration is said to decrease FID 'by 2.8% relative to the baseline,' but Table 4 (and Table 2, if row 1 is the baseline) gives a reduction from 0.209/0.263 to 0.181, i.e., roughly 13%–31%. The authors should report correct percentages or clarify what baseline is being used.","section":"§4.4 and §4.5"},{"comment":"The central assumption that 'low frequencies correlate with static poses, and high frequencies align with fine-grained motions' is asserted without direct evidence, and the related claim that the SSM matrix A 'is related to the frequency domain poles' is only supported by informal citations. No spectral analysis, frequency-band ablation against motion categories, or theoretical argument is provided. This weakens the design rationale for Eq. (9); a concrete test, such as visualizing the frequency content of static versus transitional motion segments or ablating with inverted band assignments, would help substantiate the claim.","section":"§1, §3.2"}],"minor_comments":[{"comment":"The section heading 'Expreiments' should be 'Experiments'.","section":"§4.1"},{"comment":"In the implementation details, '50 during interfering' should presumably read '50 during inference.'","section":"§4.3"},{"comment":"The sentence 'When FreqMamba is , use BiMamba instead; When TextMamba is , apply Linear Attention instead' has missing symbols (likely checkmark/cross marks), making the ablation design unclear. These should be filled in or described in prose.","section":"§4.5"},{"comment":"The table header interleaves 'R-Precision ↑' with the module indicator columns, and the column labels for FID and R-Precision are not cleanly separated; this makes the table hard to read. The module-enabled indicators should be explicitly shown.","section":"Table 2"},{"comment":"The claim that FTMoMamba 'achieves significant improvements over the baseline method (i.e., MLD) in terms of R-Precision, FID, MM-Dist, Diversity, and MModality' is not fully supported by Table 1: MModality degrades from 2.600 (MLD) to 2.277 (FTMoMamba). The sentence should be revised to acknowledge this metric.","section":"§4.4"}],"recommendation":"major_revision","confidential_remarks":"The paper's central FID claim is plausible, but the dimensional ambiguity in Eqs. (9) and (15) is a serious verifiability issue, and the ablation narrative contradicts Table 2. Since no code is provided, the authors must either give precise shapes/projections for the additions or confirm that the actual implementation differs from the written equations. I would also flag that several related works are cited as arXiv preprints without publication status, and the novelty relative to existing Mamba-based motion generation (e.g., Motion Mamba) should be more clearly delineated after the technical issues are resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper reports an impressive FID of 0.181 on HumanML3D, beating several recent systems including MotionGPT and Motion Mamba, and the core idea—feeding DWT low/high-frequency components into the Mamba A matrix and CLIP text features into the C matrix—is a reasonable new combination. The ablations are clean enough to show both modules help. But the two equations that define the mechanism, (9) and (15), are dimensionally undefined. In standard Mamba, A has shape (D,N) or (N,), while C is a parameter or has shape (B,L,N); the DWT output and f_t are (…, L/2, D) and (1,256). Adding them requires projection or broadcasting rules that are not given. Without code or a clarifying derivation, the stated 'frequency into A, text into C' claim is not verifiable. If the implementation actually uses projection layers or concatenation, the attribution of the FID gain to FreqSSM/TextSSM is unsubstantiated. The paper also has two smaller issues: the percentage improvements in Sections 4.4 and 4.5 don't match the tables (5.1% and 2.8% look like absolute differences, not relative), and R-Precision and MM-Dist are not improved over Motion Mamba, so the claimed text-motion consistency advantage is weaker than the abstract implies. On the plus side, the paper is honest about these limitations in the conclusion, the baseline comparison is standard, and the FID gain is consistently reproduced across ablations. This is a plausible incremental contribution, not a conceptual breakthrough. I'd send it to review, but only with the expectation of a major revision that resolves the dimension issue and releases code or detailed shapes. For myself, I wouldn't cite it until the mechanism is clarified.","headline":"A strong FID number with a plausible but underspecified mechanism; the core frequency-in-A / text-in-C equations don't type-check as written and there's no code, so the result is not yet reproducible.","tokens_in":12544,"tokens_out":3308,"would_cite":false,"duration_ms":27778,"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":"The paper claims that a Mamba-based diffusion denoiser reaches the lowest text-to-motion FID on HumanML3D, 0.181, by putting wavelet frequency features in the A matrix and sentence-level text features in the C matrix of the state space…","keywords":["text-to-motion generation","diffusion model","state space model","Mamba","frequency domain","discrete wavelet transform","text-motion alignment","HumanML3D"],"falsifier":"If a run of the released checkpoints on HumanML3D does not reproduce a FID near 0.181, or if replacing the A-matrix addition in Equation (9) with an ordinary residual connection to the hidden state leaves FID statistically unchanged, then the frequency-in-the-A-matrix mechanism is not what carries the reported gain.","tokens_in":1738,"feed_emoji":"🕺","tokens_out":2134,"duration_ms":69575,"temperature":0.7,"pith_summary":"The paper argues that text-to-motion diffusion models fail in two distinct ways: they miss static poses such as sitting or lying, and they miss fine-grained transitions such as stumbling or stepping over an object. It also argues that these failures come from ignoring frequency structure in the denoising latent space and from treating text as a shallow concatenated condition rather than as part of the state dynamics. To fix this, FTMoMamba builds a latent diffusion denoiser from Mamba state space blocks, with a Frequency State Space Model that injects low- and high-frequency wavelet features into the transition matrix A, and a Text State Space Model that injects sentence-level CLIP features into the output matrix C. On HumanML3D, the paper reports an FID of 0.181, compared with 0.421 for MLD and 0.232 for MotionGPT, while keeping R-Precision and diversity competitive. If correct, this shows that frequency-domain and cross-modal conditioning can be carried inside the existing parameters of an SSM rather than added as extra branches.","feed_headline":"FID 0.181: wavelets and text state spaces drive motion generation","feed_subtitle":"By folding low/high-frequency motion features into the SSM's A matrix and sentence text into C, FTMoMamba beats MLD on HumanML3D.","key_machinery":"The machinery is FreqSSM plus TextSSM inside a Mamba-based diffusion denoiser. Mamba is a linear-time state space sequence model whose discrete update is $h_t = \\bar{A} h_{t-1} + \\bar{B} x_t$ and whose output is $y_t = C h_t$; the paper's central move is to treat the learnable matrices A and C as conditioning channels rather than fixed parameters. FreqSSM decomposes the motion feature with a discrete wavelet transform, enhances the low and high bands with convolution, then adds them to A with learnable weights so the transition dynamics carry frequency-specific control. TextSSM sums the sentence-level CLIP embedding with C, so the observation mapping carries text-motion alignment. These two modified SSMs are embedded in FreqMamba and TextMamba blocks, which form the denoiser of a latent diffusion model trained to predict noise in the VAE latent space.","core_discovery":"The central discovery is that the A and C matrices of a Mamba state space model are the right places to inject two kinds of conditioning. Equation (9) constructs a frequency-aware transition matrix $A_n = A + \\alpha f_{\\text{low}} + \\beta f_{\\text{high}}$, where $f_{\\text{low}}$ and $f_{\\text{high}}$ come from applying a discrete wavelet transform to the motion feature, so the low-frequency band guides static poses and the high-frequency band guides fine-grained motion. Equation (15) constructs a sentence-level output matrix $C_s = f_t + C$, where $f_t$ is the frozen CLIP sentence embedding, so textual semantics are aligned with the motion state at every output step. The paper claims that this dual injection yields the lowest FID among compared methods on HumanML3D at 0.181, with the ablations showing that DWT beats FFT for the frequency split, that both frequency bands contribute, and that the two SSM modules are complementary.","pith_inferences":["A testable extension outside this paper is to apply the same A-matrix injection to other Mamba-based generative models, such as video or audio diffusion, where static background versus transient events mirrors static pose versus fine motion.","One could ablate the design by replacing the sentence-level text vector with token-level CLIP features; the paper's mechanism depends on a single embedding $f_t$, so this would clarify whether sentence-level aggregation is essential.","The dimensional alignment between the wavelet features and the square matrix A, and between the text embedding and the output matrix C, is not specified in the paper; a clean reproduction would need a projection, and checking the released code would settle whether Equations (9) and (15) run as written.","A controlled study that keeps the denoiser capacity fixed and toggles only the matrix injections on and off would isolate whether the mechanism itself, rather than the overall architecture, carries the FID gain."],"forward_implications":["If the reported FID holds, Mamba-based denoisers can outperform Transformer-based motion diffusion baselines without extra cross-attention branches, because conditioning is folded into the SSM matrices themselves.","The low-frequency/high-frequency split gives an interpretable control handle: low-frequency bands steer static poses and high-frequency bands steer transitions, which could make motion editing more targeted.","Injecting sentence-level text into the C matrix aligns text and motion at every state update, which should reduce the mismatch that plain concatenation of text and motion features leaves behind.","The ablation showing DWT beats FFT supports the claim that separating low and high frequency bands matters more than having a global frequency spectrum.","Using both frequency bands together lowers FID compared with either band alone, indicating that static-pose and fine-motion information are complementary rather than redundant."],"supporting_citations":[{"why":"Defines the Mamba and SSM form whose A and C matrices the paper modifies.","marker":"[10]"},{"why":"Supplies the latent diffusion baseline, the VAE encoder and decoder weights that FTMoMamba reuses, and the MLD FID value it claims to beat.","marker":"[5]"},{"why":"Provides the frozen CLIP text encoder that produces the sentence-level embedding $f_t$ injected into the C matrix.","marker":"[24]"},{"why":"Motivates the idea that high-frequency information captures dynamic actions, which FreqSSM uses for fine-grained motion.","marker":"[3]"},{"why":"Motivates the idea that low-frequency information identifies static or non-significant regions, which FreqSSM uses for static poses.","marker":"[18]"},{"why":"Provides the MotionGPT comparison point, whose FID of 0.232 is the strongest prior result that FTMoMamba's 0.181 FID is claimed to beat.","marker":"[17]"},{"why":"Provides the prior Mamba-based motion generation baseline used as comparison and as evidence that Mamba is suitable for motion sequences.","marker":"[36]"}],"fun_headline_variants":["FTMoMamba: frequency and text SSMs hit FID 0.181","Low and high frequency bands steer motion in SSM","Wavelet-split SSM and sentence embeddings cut FID","Dual SSM injections achieve lowest FID on HumanML3D","Text and frequency state spaces drive motion generation"],"cache_read_input_tokens":14720,"weakest_assumption_plain":"The load-bearing premise is that low- and high-frequency wavelet features can be added directly into the state transition matrix A and the sentence text embedding directly into the output matrix C of the Mamba layer, with dimensions that align so the state equations still hold; the paper does not specify the projection.","fun_headline_variants_meta":{"raw":{"variants":["FTMoMamba: frequency and text SSMs hit FID 0.181","Low and high frequency bands steer motion in SSM","Wavelet-split SSM and sentence embeddings cut FID","Dual SSM injections achieve lowest FID on HumanML3D","Text and frequency state spaces drive motion generation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000168,"raw_usage":{"total_tokens":1279,"prompt_tokens":984,"completion_tokens":295,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":600,"completion_tokens_details":{"reasoning_tokens":208}},"tokens_in":600,"tokens_out":295,"duration_ms":3818,"temperature":1.0,"reasoning_tokens":208,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:59:21.559585+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"If a run of the released checkpoints on HumanML3D does not reproduce a FID near 0.181, or if replacing the A-matrix addition in Equation (9) with an ordinary residual connection to the hidden state leaves FID statistically unchanged, then the frequency-in-the-A-matrix mechanism is not what carries the reported gain.","supporting_citations":[{"cited_title":"Learn- ing transferable visual models from natural language super- vision","cited_arxiv_id":null,"evidence_quote":"Provides the frozen CLIP text encoder that produces the sentence-level embedding $f_t$ injected into the C matrix."},{"cited_title":"Frequency domain analysis of human mo- tions in surveillance video","cited_arxiv_id":null,"evidence_quote":"Motivates the idea that high-frequency information captures dynamic actions, which FreqSSM uses for fine-grained motion."},{"cited_title":"Differentiable Frequency-based Disentanglement for Aerial Video Action Recognition","cited_arxiv_id":"2209.09194","evidence_quote":"Motivates the idea that low-frequency information identifies static or non-significant regions, which FreqSSM uses for static poses."},{"cited_title":"Motiongpt: Human motion as a foreign language","cited_arxiv_id":null,"evidence_quote":"Provides the MotionGPT comparison point, whose FID of 0.232 is the strongest prior result that FTMoMamba's 0.181 FID is claimed to beat."},{"cited_title":"Motion mamba: Efficient and long sequence motion generation","cited_arxiv_id":null,"evidence_quote":"Provides the prior Mamba-based motion generation baseline used as comparison and as evidence that Mamba is suitable for motion sequences."}],"review_version":1}