{"id":"c3073453-a8c8-4784-a60d-bcc4f153f39d","arxiv_id":"2605.11704","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"ScaleMoGen applies next-scale autoregressive prediction to human motion generation with multi-scale skeletal-temporal bitwise token maps, reporting SOTA FID on HumanML3D.","lead":"ScaleMoGen generates 3D human motion by predicting progressively finer token maps in a coarse-to-fine order, instead of predicting tokens one after another along time. The authors report state-of-the-art FID on HumanML3D and the best CLIP score on SnapMoGen, plus zero-shot text-guided motion editing.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Main configuration (Strategy 5) may not implement the skeletal coarse-to-fine hierarchy claimed in Section 3.1; if it uses flat atomic segments at all scales, the central skeletal-temporal contribution is unverified.","rationale":"The reader's weakest assumption is that the fixed 7-segment skeletal partition is semantically adequate. I identify a more specific and manuscript-internal problem: the selected main configuration may not implement the skeletal hierarchy described in Section 3.1 at all, because Strategy 5 appears to keep the same atomic joint groups at every scale. This is not an external-consensus disagreement; it is a possible mismatch between the proposed method and what was actually run, making it load-bearing for the central contribution. The concrete test—checking S(v) in the released config and re-running with a true skeletal-tree strategy—would settle whether the claimed coarse-to-fine skeletal mechanism contributes to the SOTA numbers. This concern reinforces the need for code release and a fair hierarchy ablation at the main bit-width, so the verdict should remain CONDITIONAL. I set agreement_with_reader to 'partial' because my concern is more severe and more specific than the reader's, though it is related to the skeletal-partition theme.","tokens_in":15791,"tokens_out":5812,"duration_ms":65654,"concrete_test":"Access the released code/supplementary config for Strategy 5 and enumerate S(v), m_v, h_v for v=0..6. Verify whether S(0)={J}, S(V)=7 atomic segments, and recursive refinement holds for intermediate v, or whether m_v=7 for all v. If the partition is flat, re-run the main SnapMoGen and HumanML3D experiments using Strategy 1 or Strategy 2 at 24-bit code with the same evaluation protocol. If FID/CLIP/R-Precision are statistically indistinguishable or worse than the reported Strategy 5 numbers, the skeletal coarse-to-fine mechanism is not the source of the SOTA gains, and the central claim must be repositioned.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central novelty is a multi-scale skeletal-temporal hierarchy: Section 3.1 defines S(0)={J} (whole body), S(V)=atomic segments, with recursive refinement, and Eq. (2) relies on progressively finer skeletal partitions. However, the main configuration selected in supplementary B.1 is Strategy 5, described as a \"7-scale full hierarchy, representing independent anatomical joint groups across all scales.\" That phrasing indicates m_v=7 at every scale (the skeleton is always partitioned into the seven atomic groups), i.e., no spatial coarsening along the skeletal axis—only temporal resolution changes. If so, Tables 1-2 do not actually exercise the skeletal hierarchy that the method is named after. The only relevant ablation, Table 3(A), removes skeletal topology entirely rather than varying the spatial partition; Table 7 compares five hierarchy strategies but was run on HumanML3D with 32-bit code (stated in B.1) and the chosen Strategy 5 is not one of the true skeletal-tree configurations (1-4). Thus the claim that \"skeletal-temporal multi-scale representation\" drives SOTA performance is not directly supported: the gains could come from bitwise temporal multi-scale residual VQ plus fixed atomic joint grouping, not from recursive coarse-to-fine skeletal refinement. This is an internal-consistency risk for the paper's main contribution, and it also weakens the response to the reader's concern about the fixed 7-segment partition—if the spatial partition never changes across scales, the hierarchical refinement argument in Section 3.1 is not what is being evaluated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"ScaleMoGen proposes an autoregressive next-scale prediction framework for text-driven 3D human motion generation. The motion is encoded into a skeletal-temporal latent grid and progressively quantized into multi-scale bitwise token maps; a transformer then predicts the token map at each scale conditioned on accumulated coarser scales, starting from a 1x1 map. The authors claim state-of-the-art results on HumanML3D and SnapMoGen, and additionally propose a training-free text-guided editing mechanism based on preserving/masking tokens at chosen scales, skeletal parts, or semantic-confidence thresholds. The paper includes ablations on skeletal topology, code size, model size, VQ design, CFG scheduling, and sampling efficiency, plus a small user study for editing.","tokens_in":16236,"tokens_out":3506,"duration_ms":37963,"significance":"If the central claims hold, ScaleMoGen would be a meaningful step in discrete motion generation: shifting from temporal next-token prediction to scale-wise coarse-to-fine prediction is a natural and underexplored idea, and the bitwise multi-scale VQ design is a clean way to avoid categorical codebook collapse. The reported HumanML3D FID (0.030) and SnapMoGen CLIP Score (0.693) are competitive with recent strong baselines, and the zero-shot editing concept is attractive. The paper also provides useful ablations (hierarchy strategy, code size, model size, CFG schedule) and reports sampling efficiency. However, the manuscript's central claim that the method exploits a recursive skeletal coarse-to-fine hierarchy is not actually supported by the configuration used for the main results, and the 'state-of-the-art' claim is metric-dependent, so the significance of the contribution as stated is uncertain.","major_comments":[{"comment":"The main configuration (Strategy 5) is described as a '7-scale full hierarchy, representing independent anatomical joint groups across all scales.' This directly contradicts Section 3.1, which defines S(0)={J}, S(V)=atomic segments, and recursive refinement with varying m_v. If m_v=7 at every scale, the spatial partition is fixed atomic and the skeletal axis never coarsens; only temporal resolution changes. Then Eq. (2) and the central 'skeletal-temporal multi-scale' claim are not what Tables 1-2 exercise. Table 7 also uses 32-bit codes for the HumanML3D column, so those comparisons are not aligned with the 24-bit main results. Please state the explicit (h_v, m_v) sequence for Strategy 5, and either run the main experiments with a true skeletal hierarchy or revise the paper's title/abstract/claims to reflect a fixed skeletal grouping with multi-scale temporal refinement.","section":"B.1 + Table 7"},{"comment":"The abstract's 'state-of-the-art' claim is metric-dependent. On SnapMoGen, ScaleMoGen's FID (16.35) is worse than MoMask++ (15.06), while only CLIP Score is better (0.693 vs 0.685). On HumanML3D, SALAD has higher Top-1 R-Precision (0.581 vs 0.577). The abstract cherry-picks FID for HumanML3D and CLIP Score for SnapMoGen. Please qualify the claim per metric and per dataset, and discuss the FID/CLIP tradeoff on SnapMoGen explicitly.","section":"Abstract + Tables 1-2"},{"comment":"The editing user study is based on only 10 examples, and Table 5 reports means and standard deviations without any significance test, inter-rater reliability, or breakdown by participant/example. The statement that ScaleMoGen 'achieves the highest scores across all criteria' is not statistically supported. Since zero-shot editing is a headline contribution, either enlarge the study and add statistical analysis, or substantially soften the claim.","section":"Section 4.3 + Table 5"}],"minor_comments":[{"comment":"Typo: 'skeletal-emporal' should be 'skeletal-temporal'.","section":"Abstract"},{"comment":"'The world-level text embeddings c' should be 'word-level text embeddings'.","section":"Section 3.2"},{"comment":"The descriptions of Strategies 1-5 are verbal and hard to parse; please include an explicit table listing h_v, m_v, and the skeletal partition for each scale for each strategy. Figure 5 shows pooling to the 7-joint atomic topology but not the scale schedule.","section":"B.1"},{"comment":"The row ordering for the 32-bit code is confusing: the text says 'increasing to 32 bits yields the opposite trend' but the FID improves (15.98) and CLIP degrades (0.683); this is opposite to 16-bit but the row order is 16, 32, 24 in the text while the table lists Base, A, B, C. Consider reordering for readability.","section":"Table 3"},{"comment":"The ablation in Table 3(A) removes skeletal topology entirely (binary condition). It does not vary the skeletal partition, so it cannot validate the specific choice of the 7 atomic segments or the recursive refinement hierarchy. Please discuss this limitation.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper compares against MoMask++ (Guo, Hwang, Wang, Zhou), which shares overlapping authorship. The numbers appear taken from the original paper, but given the overlap, the authors should state clearly whether all baselines are reproduced under identical evaluation protocols. The main technical concern is internal consistency: the main configuration as described in B.1 does not exercise the recursive skeletal hierarchy claimed in Section 3.1. This must be resolved before the paper can be considered for acceptance. The 'state-of-the-art' phrasing in the abstract should also be toned down to avoid cherry-picking."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know about ScaleMoGen. First, the core recipe—applying bitwise next-scale autoregressive prediction from VAR/Infinity to human motion, with a 2D skeletal-temporal token map—is a real and reasonable contribution, and the paper shows competitive numbers on HumanML3D and SnapMoGen. Second, the headline claim that a recursive skeletal coarse-to-fine hierarchy drives those results is not actually supported by the configuration used in the main experiments. The supplementary describes the chosen Strategy 5 as a '7-scale full hierarchy, representing independent anatomical joint groups across all scales,' meaning the skeleton is partitioned into the same 7 atomic segments at every scale; only the temporal resolution changes. Section 3.1, by contrast, defines S(0)={J} with progressive refinement. So the main results do not exercise the skeletal hierarchy the paper is named after.\n\nWhat is genuinely new: the combination of bitwise residual quantization (from binary spherical quantization) with a multi-scale temporal token map and a transformer that predicts next-scale binary maps. The VQ reconstruction comparison in Table 4 is strong (FID 0.12, MPJPE 2.73), and the ablations on code size, model size, and CFG schedule are informative. The zero-shot editing idea—masking tokens at specific skeletal/temporal locations and resampling with a target text—is clever and worth building on.\n\nThe soft spots, in order. The main one is the hierarchy mismatch above. The paper's own hierarchy ablation (Table 7) shows Strategy 2, a genuine skeletal tree, outperforms Strategy 5 on HumanML3D; Strategy 5 only wins on SnapMoGen, and those HumanML3D numbers were run with 32-bit codes rather than the 24-bit used elsewhere. That does not demonstrate that recursive skeletal coarsening is the source of the gains. It could be the temporal multi-scale plus fixed atomic grouping that matters. There is also a minor internal tension in Section C, which claims paired limbs are grouped at coarser scales, even though the main config keeps all 7 groups separate at every scale. Other issues are smaller: the SnapMoGen 'state-of-the-art' claim is metric-dependent because FID is worse than MoMask++ (16.35 vs 15.06), the editing user study has only 10 examples and no statistics, and code is not released. The overlap with MoMask++ authors is worth noting but not disqualifying.\n\nWho should read this: anyone working on discrete motion representation or efficient text-to-motion generation. The recipe is useful even if the narrative oversells the skeletal hierarchy. It deserves a serious referee—the benchmark numbers are strong enough that the field will care—but the referee should push on the hierarchy question. As it stands, I would treat the central claim as conditional: either re-run with a true skeletal hierarchy or re-scope the contribution to temporal next-scale modeling with fixed skeletal segments.","headline":"Strong results, but the main configuration doesn't actually use the recursive skeletal hierarchy the paper is built around—temporal multi-scale plus fixed atomic segments may be the real story.","tokens_in":16676,"tokens_out":5385,"would_cite":true,"duration_ms":49530,"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":"ScaleMoGen claims that text-driven human motion generation is best framed as coarse-to-fine next-scale prediction over a hierarchy of discrete skeletal-temporal token maps, rather than as next-token prediction along time, and that this fram","keywords":["multi-scale discrete tokenization","next-scale prediction","autoregressive motion generation","text-to-motion","bitwise quantization","skeletal hierarchy","zero-shot motion editing","coarse-to-fine generation"],"falsifier":"Prompt a released ScaleMoGen to generate a motion requiring independent finger counting (e.g., 'counts to three with fingers raised one by one'). If the output cannot show distinct finger configurations—because fingers are not atomic segments in the 7-part hierarchy—then the skeletal hierarchy does not cover all human motions, falsifying the central claim that this multi-scale representation is sufficient for text-to-motion.","tokens_in":15754,"feed_emoji":"🏃","tokens_out":10675,"duration_ms":104449,"temperature":0.7,"pith_summary":"ScaleMoGen sets out to show that text-driven human motion generation is best approached not as next-token prediction in time, but as coarse-to-fine next-scale prediction over a hierarchy of discrete token maps. The paper builds a multi-scale skeletal-temporal representation: motion is quantized into residual token maps of increasing temporal and skeletal granularity, using bitwise quantization that preserves the kinematic hierarchy, and a transformer predicts each finer map conditioned on all coarser ones and the text. On two standard benchmarks, this yields the best reported fidelity and text-alignment scores. The same representation makes training-free text-guided motion editing possible by selectively preserving or resampling tokens at chosen scales and body parts. If correct, this reframes a generation problem that was temporal into one that is structural.","feed_headline":"Coarse-to-fine motion generation beats token-by-token prediction","feed_subtitle":"Multi-scale skeletal token maps lift text-to-motion fidelity and enable training-free editing.","key_machinery":"The load-bearing object is the multi-scale skeletal-temporal token map: a tuple of discrete grids (time and skeletal segments) at scales v=0..V, produced by residual bitwise quantization where each latent dimension becomes a binary label, so the effective vocabulary is 2^d. The operators that make it work are the topology-aware downsampling and upsampling functions, which move between the full-resolution latent grid (n x j x d, with j=7 atomic segments) and each scale's coarser grid while respecting the kinematic tree; the recursive-refinement property ensures coarser segments are unions of finer ones. This representation converts generation into next-scale prediction: a transformer, conditi","core_discovery":"The paper's central claim is that human motion is naturally a coarse-to-fine structure, and that a generative model should mirror this: start with a token that encodes the whole body's global motion, then progressively add finer temporal resolution and finer skeletal decompositions. To carry this, ScaleMoGen defines a hierarchy of token maps where each scale has its own temporal resolution and skeletal partition (satisfying completeness and recursive refinement), with the finest scale at 7 atomic segments—root, spine, head, two arms, two legs. The motion encoder produces a continuous latent grid, and residual bitwise quantization turns it into binary token maps at every scale; topology-aware","pith_inferences":["The coarse-to-fine token hierarchy suggests a generic recipe for other structured temporal signals (audio, gestures, graphs): predict a hierarchical latent map from coarse to fine via residual bitwise quantization. A testable extension would apply the same architecture to speech or music generation and compare next-scale prediction against next-token baselines.","The fixed 7-segment skeletal partition may be too coarse for actions that live in hands or face; a learnable hierarchy—discovering segments from data—could extend the method to fine-grained control. The paper's own ablation shows dataset-specific sensitivity to hierarchy design, so a data-driven partition is a natural next step.","The zero-shot mask could be used beyond editing: e.g., motion interpolation between two text prompts by progressively blending token maps at different scales, or controllable generation where the user specifies which body part follows which text. These are not explored in the paper.","The scale-adaptive guidance schedule (stronger text guidance at finer scales) implies fine details are more text-dependent than global structure; a similar scheduling principle might improve other autoregressive or diffusion systems, though this is an editorial extrapolation."],"forward_implications":["If the next-scale formulation is correct, autoregressive motion generation no longer needs long temporal token sequences: a constant number of scale steps (7 in the paper) produces full motions, cutting inference cost relative to diffusion samplers that require many denoising steps.","Text-guided editing becomes a token-preservation operation: users can keep coarse tokens for global structure, or mask specific body parts and time intervals, and the model resamples only the masked tokens conditioned on the target text—no extra training.","Bitwise quantization gives an exponentially large discrete vocabulary (2^d) that the paper argues preserves fine motion detail better than fixed codebooks, which could make discrete token approaches more competitive with continuous diffusion on fidelity.","Scaling the predictor to 2.2B parameters did not improve alignment and slightly hurt the fidelity metric on current datasets, implying that for text-to-motion, data quantity—not model capacity—is the current limiting factor."],"fun_headline_variants":["Next-scale autoregression: human motion from coarse to fine","Scale-by-scale autoregressive motion beats token prediction","Coarse-to-fine token maps for sharper text-driven motion","Multi-scale skeleton tokens enable training-free motion editing"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the fixed, hand-defined skeletal hierarchy (whole body, then upper/lower body, then seven atomic segments: root, spine, head, arms, legs) together with its temporal upsampling schedule is sufficient to express the semantic structure of every motion that text can describe; if some motion requires a finer or differently grouped joint partition, the token maps cannot represent it and the generation claim fails.","fun_headline_variants_meta":{"raw":{"variants":["Next-scale autoregression: human motion from coarse to fine","Scale-by-scale autoregressive motion beats token prediction","Coarse-to-fine token maps for sharper text-driven motion","Multi-scale skeleton tokens enable training-free motion editing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001294,"raw_usage":{"total_tokens":5103,"prompt_tokens":714,"completion_tokens":4389,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":458,"completion_tokens_details":{"reasoning_tokens":4325}},"tokens_in":458,"tokens_out":4389,"duration_ms":30507,"temperature":1.0,"reasoning_tokens":4325,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T14:13:37.618871+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Prompt a released ScaleMoGen to generate a motion requiring independent finger counting (e.g., 'counts to three with fingers raised one by one'). If the output cannot show distinct finger configurations—because fingers are not atomic segments in the 7-part hierarchy—then the skeletal hierarchy does not cover all human motions, falsifying the central claim that this multi-scale representation is sufficient for text-to-motion.","supporting_citations":[],"review_version":2}