{"id":"95650d70-1776-425a-8480-2254f997a3f6","arxiv_id":"2601.14788","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"RAM couples motion reconstruction with text-to-motion diffusion and adds reconstruction-anchored error guidance, reporting FID 0.032 on HumanML3D with 20 inference steps.","lead":"A text-to-motion diffusion model called RAM adds a motion-reconstruction branch that also serves as a negative reference during inference. On HumanML3D it reports FID 0.032 and R-Precision@1 56.1% with 20 denoising steps, a level previously reached by VQ-VAE-based systems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"REG's causal role in the headline FID is not established: with w1=5, w2=1.5 the update is a large extrapolation, the ablation sweep is on the test set, and no CFG-only or 'bad version' control at matched guidance strength is reported.","rationale":"The reader identifies REG's off-manifold latent encoding as the weakest assumption. I agree that REG is the least secure component, but I would sharpen the concern: the empirical evidence in Table 5 does not isolate REG's proposed mechanism from a generic increase in guidance strength. The combined update in Eq. (11) uses a text coefficient of 1+w1+w2 = 7.5, and the ablation only compares REG/CFG at moderate individual weights, never a CFG-only run at the equivalent effective guidance scale. Because the two guidance terms interact strongly (each alone gives FID ≈0.13, together 0.032), the large headline improvement could be an artifact of a high-guidance extrapolation rather than error-propagation correction. Additionally, the hyperparameters are selected from sweeps on the test split, so the exact w1=5, w2=1.5 combination may be overfit to HumanML3D. This is not a fatal flaw: the paper provides ablations, and the method is plausible. But the central claim that REG 'mitigates error propagation' requires either a derivation of why the Em-conditioned residual is a valid score direction or a control experiment that matches guidance strength. I therefore keep the reader's CONDITIONAL verdict; the concern reinforces it but does not move it.","tokens_in":18779,"tokens_out":10131,"duration_ms":109819,"concrete_test":"Hold out a random 10% of the HumanML3D training split as validation. Tune w1 and w2 only on this validation split. Then compare three inference settings on the standard test set: (i) RAM with validation-tuned (w1,w2); (ii) CFG-only with w1=0 and w2 swept over the effective guidance range implied by Eq. (11) (e.g., 5, 6.5, 7.5, 9) to bracket the text/unconditional coefficient pair; (iii) a 'bad version' control that replaces Em(x_hat_{t+1,s}) with a dropout- or attention-perturbed text-conditioned prediction, matched for total compute. If (ii) or (iii) reaches FID≈0.032, REG's specific mechanism is not the cause of the headline result.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3 defines REG (Eq. 10) as an amplified residual between the text-conditioned prediction D(xt,t,zt) and the reconstruction D(xt,t,Em(x_hat_{t+1,s})). The claim that this residual corrects error propagation requires that Em generalize to off-manifold inputs x_hat_{t+1,s} and that the residual direction point toward clean data. Neither is demonstrated. Moreover, the headline result is not a clean test of REG: in Eq. (11) with w1=5, w2=1.5, the text prediction receives total coefficient 1+w1+w2=7.5, and the unconditional term is -1.5, so the combined update is a large guidance extrapolation. Table 5 shows w1=5 alone (FID 0.128) and w2=1.5 alone (FID 0.132) are both mediocre; only their combination yields FID 0.032. No CFG-only control at equivalent effective guidance strength (e.g., w2 in the 5-9 range) is reported, and hyperparameters are chosen from test-set sweeps without a validation split. Thus the FID gain attributed to REG could be an artifact of guidance-strength/overfitting rather than the proposed error-propagation mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Reconstruction-Anchored Diffusion Model (RAM) for text-to-motion generation. The method has two main components: (1) a motion reconstruction branch that co-trains a motion encoder and a diffusion decoder with self-regularization and motion-centric latent alignment, aiming to map text embeddings into a motion-sensitive latent space; and (2) Reconstructive Error Guidance (REG), an inference-time mechanism that reconstructs the previous denoising estimate with the motion encoder and uses the amplified residual between the current text-driven prediction and this reconstruction to correct error propagation. On HumanML3D, the authors report an FID of 0.032 and R-Precision@1 of 56.1% with 20 inference steps, claiming state-of-the-art performance among diffusion-based methods and competitiveness with VQ-VAE-based models; consistent but more modest gains are reported on KIT-ML.","tokens_in":19194,"tokens_out":5914,"duration_ms":63072,"significance":"If the reported results are robust, the paper makes a meaningful empirical contribution: it demonstrates that a diffusion-based text-to-motion model can close the FID gap with VQ-VAE-based models, and it proposes a novel guidance mechanism that is conceptually interesting. The paper is thorough in its ablation coverage—component ablations, latent-space training variants, hyperparameter sensitivity, inference-efficiency experiments, and a user study—and the benchmark numbers are internally consistent. However, the central causal claim for REG is not convincingly isolated: the reported improvement may be attributable to guidance-strength effects or test-set overfitting rather than the proposed error-propagation-correction mechanism. Because this is the paper's headline contribution, the manuscript currently overclaims its evidence.","major_comments":[{"comment":"The ablation does not isolate REG from guidance-strength effects. Eq. (11) can be rewritten as (1+w1+w2)D(xt,t,zt) - w1 D(xt,t,zm,t+1) - w2 D(xt,t,∅). With w1=5 and w2=1.5, the conditional term is scaled by 7.5, far beyond any CFG-only value reported (the largest is w2=4.5). Table 5 shows w1=5 alone gives FID 0.128 and w2=1.5 alone gives 0.132, yet their combination gives 0.032. A CFG-only control at matched effective guidance strength (e.g., w2=6 or 7.5) is missing. Without this control, the FID gain attributed to REG could be an artifact of the large combined extrapolation rather than the proposed error-propagation correction.","section":"Section 3.3, Eq. (11), Table 5"},{"comment":"REG assumes that Em(sm, xhat_{t+1,s}) is a meaningful encoding of an off-manifold estimate and that the residual D(xt,t,zt) - D(xt,t,zm,t+1) points toward clean data. However, Em is trained only on clean motions (Eq. (1)); no evidence is provided that Em generalizes to the noisy/off-manifold inputs encountered during sampling. The mechanism is supported only by the ablation in Table 5. Please add diagnostics: compare the distribution of zm,t+1 with training-time latents, measure the correlation between the REG residual and the clean-data direction, and include a control where zm,t+1 is replaced by a random or perturbed latent. Without such analysis, the claim that REG corrects error propagation is not established.","section":"Section 3.3, Eq. (10), Section 3.1, Eq. (1)"},{"comment":"The hyperparameters w1, w2, β, τ, w_sr, w_latent, and d_E are selected based on test-set metrics, and no validation split is described. Since the headline FID of 0.032 is the result of these sweeps, the reported numbers may reflect test-set overfitting rather than generalizable performance. Please specify the validation protocol used for model selection, or re-report the results with a clear train/validation/test split. At minimum, show that the selected hyperparameters are stable on a held-out portion of the data.","section":"Sections 4.2, 4.5, 4.7; Tables 5, 7, 8, 9"}],"minor_comments":[{"comment":"The text says 'treating only identical indices as positive pairs,' but the formula is actually a uniformity loss: the numerator exp(sim(zi,zi)/τ) is constant (sim=1 after normalization), and the loss pushes all latents apart. Please revise the wording to avoid implying a supervised positive-pair contrastive loss.","section":"Section 3.2, Eq. (7)"},{"comment":"There are typos such as 'inital' (should be 'initial'), and several table captions have 'T able' instead of 'Table'. Please proofread.","section":"Table 3 caption and table formatting"},{"comment":"The code is only promised ('will be released'). Given that REG is an inference-time heuristic whose exact implementation matters, please include a pseudocode algorithm for the inference loop and, ideally, release the code and checkpoints to support reproducibility.","section":"Section 6 and Abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper will likely attract attention for its strong FID on HumanML3D, but the central REG attribution gap and the test-set-based hyperparameter selection are significant. I recommend major revision rather than rejection because both issues are addressable with additional experiments and a clearer validation protocol. If the authors can provide a CFG-only control at matched guidance strength and diagnostics for the REG mechanism, the paper could be a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper likely has a real result in the two-stream training, but the headline FID number is not cleanly attributable to the REG mechanism because the paper never tests CFG at the same effective guidance strength. That doesn't sink the paper, but it should be fixed before I'd trust the central claim.\n\nThe genuinely new piece is the combination of a reconstruction branch trained with a uniformity-style self-regularization and a stop-gradient motion-centric alignment, plus an inference-time guidance that uses the reconstruction encoder on the previous step's prediction. The evaluation is careful: the incremental ablation is clean, and Table 6 compares competing latent-space strategies under identical architecture and decoder, which is more than most papers in this area do. The efficiency analysis showing most of the FID gain comes from early denoising steps is a nice supporting observation. Benchmarks include confidence intervals and standard metrics, and the KIT-ML numbers are coherent.\n\nWhere it gets softer. The REG term is essentially a heuristic. Eq. 10-11 have no derivation, and the paper itself says the self-correction capability is a hypothesis. The stress-test concern lands: in Eq. 11 the text prediction carries coefficient 1+w1+w2=7.5, while the unconditional term is -1.5, so the combined update is a large guidance extrapolation. Table 5 reports CFG-only sweeps up to w2=4.5 (FID 0.095), but never at the effective strength of the full REG configuration. So the dramatic drop from 0.132 to 0.032 could in part be guidance strength rather than the proposed error-correction mechanism. The paper needs a matched-strength CFG-only control, and ideally a 'bad version of itself' baseline in the style of Karras et al., to pin down REG's contribution.\n\nAlso, the hyperparameter sweeps in Tables 5, 7, 8, and 9 appear to be on the test set with no validation split described. That's a real overfitting risk. The 'state-of-the-art' phrasing overstates the numbers: RAM ties LaMP on FID and trails Salad on R-Precision on HumanML3D. No code or checkpoints are released yet either.\n\nWho is this for? Anyone working on text-to-motion or latent-space diffusion. The two-stream alignment result is solid enough that the paper deserves referee time, and the REG question is a good test of healthy skepticism. I'd suggest the editor send it to a reviewer who is comfortable with both diffusion guidance and representation learning.\n\nBottom line: worth engaging with, but the REG attribution and the validation split need to be addressed before this is a clean result.","headline":"The two-stream training and ablations are solid, but the headline FID gain from REG is not cleanly attributed — the paper never tests CFG at matched guidance strength — so the empirical claim is conditional until that's fixed.","tokens_in":19627,"tokens_out":5566,"would_cite":true,"duration_ms":47746,"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":"RAM improves diffusion-based text-to-motion generation by anchoring the denoising process to a motion-centric latent space and using reconstruction errors as guidance, achieving state-of-the-art FID and semantic alignment on HumanML3D.","keywords":["text-to-motion generation","diffusion models","motion latent space","reconstruction guidance","error propagation","self-regularization","latent alignment","classifier-free guidance"],"falsifier":"On a held-out set of real motions, compute the cosine similarity between the REG residual (D(xt,t,zt) − D(xt,t,Em(x̂t+1,s))) and the true error direction (x0 − x̂t,s); if the average similarity is not positive across early steps, REG is not correcting toward the data. Alternatively, run the same 20-step sampler with REG replaced by a residual computed from a randomly reconstructed latent: if FID remains at 0.032, the specific motion encoder is not the cause.","tokens_in":18656,"feed_emoji":"🏃","tokens_out":3925,"duration_ms":43085,"temperature":0.7,"pith_summary":"The paper argues that two key weaknesses of diffusion-based text-to-motion models—the gap between pre-trained text embeddings and actual motion dynamics, and error accumulation during iterative denoising—can both be addressed by leveraging a motion reconstruction branch. RAM co-trains this branch with the diffusion decoder, then at inference uses it to reconstruct the previous denoising step's output and amplifies the difference between the current text-driven prediction and that reconstruction. The claimed result is a large improvement in motion realism (FID 0.032 on HumanML3D) and semantic alignment (R-Precision@1 56.1%) with only 20 sampling steps, placing diffusion methods on par with or ahead of VQ-VAE-based generators. A sympathetic reader would care because it suggests diffusion models do not have to sacrifice fidelity for flexibility.","feed_headline":"Diffusion text-to-motion hits FID 0.032, rivaling VQ methods","feed_subtitle":"A reconstruction branch and error guidance close the fidelity gap between diffusion and VQ-VAE motion generators.","key_machinery":"The central object is the motion-centric latent space produced by a transformer motion encoder Em, together with the Reconstructive Error Guidance (REG) term. The diffusion decoder is trained to both reconstruct motion from Em's latent and generate from text latents, while two objectives shape the space: self-regularization, a softmax cross-entropy that pushes different motion latents apart, and motion-centric latent alignment, which pulls text latents toward motion latents while almost fully stop-gradients on Em. At inference, the final prediction is x̂t,s = D(xt,t,zt) + w1(D(xt,t,zt) − D(xt,t,Em(x̂t+1,s))) plus the CFG term, where the reconstruction of the previous estimate acts as a delib","core_discovery":"RAM claims that a motion-centric latent space, learned by reconstructing motion sequences and regularized by self-regularization and motion-centric latent alignment, gives text embeddings a target space that carries kinematic structure. At sampling time, passing the previous step's estimate through the same motion encoder and subtracting the resulting prediction from the current text-conditioned prediction yields a residual that, when amplified, steers the sample back toward clean motion. On HumanML3D, the combined system reports FID 0.032 and R-Precision@1 56.1% with 20 inference steps, surpassing previous diffusion methods and outperforming most VQ-VAE-based models.","pith_inferences":["REG is, in principle, a training-free plug-in for any two-stream diffusion model that already has a reconstruction encoder; it could be lifted to other generation domains, provided the encoder is calibrated on off-manifold inputs.","The paper's ablations leave open whether REG's benefit comes from true error correction or from a generic sharpening of the prediction; a direct test would compare it against adding a random or unconditional residual of the same magnitude.","Because the motion encoder is only trained on clean motion, REG's premise is fragile: if it is not robust to off-manifold estimates, the amplified residual could inject noise; a testable extension is to also train the encoder on denoised or intermediate samples.","The self-regularization loss likely increases semantic resolution of the latent space, so a natural extension is to evaluate the same learned latents on motion-text retrieval tasks."],"forward_implications":["Diffusion-based text-to-motion models can reach the FID levels of VQ-VAE models, narrowing the historical fidelity gap.","Applying REG only in early denoising steps already yields most of the FID gain, confirming that error propagation is dominated by early steps.","REG and classifier-free guidance are complementary: CFG improves text-motion alignment while REG improves realism, and both can be combined.","The model is efficient: with 20 inference steps and REG enabled, average inference time per sentence remains below a standard 50-step diffusion baseline.","Motion-centric alignment with a nearly frozen motion encoder outperforms bidirectional alignment and contrastive strategies on realism, while matching them on semantic metrics."],"fun_headline_variants":["Diffusion motion hits FID 0.032 via reconstruction guidance","RAM: anchor diffusion in motion space, cut FID to 0.032","Error-guided sampling lifts text-to-motion to FID 0.032","Motion-latent supervision gets FID 0.032, rivaling VQ"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"REG assumes that the motion encoder, trained only on clean motion sequences, produces a useful 'weakened' prediction when fed an off-manifold previous estimate, such that the amplified residual points toward clean data rather than adding noise.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion motion hits FID 0.032 via reconstruction guidance","RAM: anchor diffusion in motion space, cut FID to 0.032","Error-guided sampling lifts text-to-motion to FID 0.032","Motion-latent supervision gets FID 0.032, rivaling VQ"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00048,"raw_usage":{"total_tokens":2210,"prompt_tokens":742,"completion_tokens":1468,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":1384}},"tokens_in":486,"tokens_out":1468,"duration_ms":12431,"temperature":1.0,"reasoning_tokens":1384,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T09:03:48.860205+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a held-out set of real motions, compute the cosine similarity between the REG residual (D(xt,t,zt) − D(xt,t,Em(x̂t+1,s))) and the true error direction (x0 − x̂t,s); if the average similarity is not positive across early steps, REG is not correcting toward the data. Alternatively, run the same 20-step sampler with REG replaced by a residual computed from a randomly reconstructed latent: if FID remains at 0.032, the specific motion encoder is not the cause.","supporting_citations":[],"review_version":1}