{"id":"f2f9402a-3692-4964-b469-254cc960dbcb","arxiv_id":"2502.02063","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"CASIM replaces fixed-length text embeddings with token-level cross-attention in text-to-motion models, improving alignment and quality for both diffusion and autoregressive generators.","lead":"The paper puts word-level text tokens directly into motion generation models, allowing each body movement to focus on the words that describe it. This improves the match between text prompts and generated human motions compared to using a single summary embedding.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim is not yet isolated from model capacity: CASIM's aligner adds parameters that baselines are not given, so the reported gains may reflect extra capacity rather than composite-aware token injection.","rationale":"I read the paper as attempting to show that replacing fixed-length CLIP embeddings with token-level embeddings plus a learned text-motion aligner improves text-to-motion generation across architectures. For that claim to hold, the improvement must come from the composite-aware injection itself rather than from the extra parameters and sequence capacity introduced by the aligner. The paper does not provide a parameter-matched control: the baselines are run with their original conditioning, while CASIM versions add projection layers and attention modules (Eqs. 2, 4, 5). The statement in Section 4.1 that other settings follow the baselines only guarantees equal training protocol, not equal model capacity. This is the weakest load-bearing assumption because it directly affects every quantitative claim in Tables 3 and 4. The reader's verdict identified exactly this concern. I also note secondary issues, such as inconsistent baseline numbers between Tables 1 and 3 (e.g., MDM FID 0.489 vs 0.325) and absent error bars, but these strengthen the need for careful controls rather than replacing the capacity confound. The proposed concrete test is feasible: a matched-capacity [CLS]-based control isolates semantic granularity from parameter count. If the control fails to reproduce CASIM's gains, the mechanism is validated; if it matches them, the central claim is overstated. Since the reader already requested parameter-matched ablations and the concern does not force a rejection at this stage, I recommend keeping the CONDITIONAL verdict.","tokens_in":14834,"tokens_out":3542,"duration_ms":38117,"concrete_test":"Run an equal-parameter control on HumanML3D for at least MDM-Decoder and T2MGPT: augment the baseline with the same linear projections and the same number/width of text-motion attention layers as CASIM, but feed only the fixed-length CLIP [CLS] embedding (broadcast or repeated over the sequence length) instead of token-level embeddings. Keep all other settings, training budget, and random seeds identical; report R-Precision Top1/Top3, FID, and MM-Dist. If the parameter-matched [CLS] control reproduces CASIM's gains, the token-level composite mechanism is not the cause; if CASIM beats the control by more than run-to-run noise (the paper reports no error bars), the central claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that token-level embeddings plus the text-motion aligner, rather than fixed-length CLIP [CLS] injection, drive the consistent improvements (Abstract, Tables 3-4). The load-bearing assumption is that the injected mechanism itself, not added model capacity, causes the gains. That assumption is unverified. For the decoder-based diffusion variant, CASIM inserts an MHCA module between motion and text tokens (Eq. 5); for the encoder variant it adds a linear projection and grows the text context from one token to N tokens (Eq. 4). For autoregressive models, the aligner changes the conditioning sequence and adds attention interactions (Eq. 2). Section 4.1 states 'All the other settings follow the baseline methods and hyperparameters remain unchanged,' but this controls training protocol, not parameter count or sequence capacity. No baseline is augmented with an equal-parameter conditioning module that still uses only the fixed-length [CLS] embedding. Table 5 reports useful ablations (encoder/decoder, teacher forcing, CoMo without keywords), but none of these holds capacity fixed while varying semantic granularity. The MLD limitation (Appendix C) actually supports the concern: adding CASIM's machinery can hurt text-motion metrics, so the sign of the effect is not determined by the mechanism alone. Until a parameter-matched [CLS]-based control is run, the stated improvements in R-Precision and MM-Dist are confounded with capacity, and the causal attribution to 'composite aware semantic injection' is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CASIM, a semantic injection module that replaces fixed-length CLIP [CLS] text embeddings with token-level CLIP embeddings and a learned text-motion attention aligner. It instantiates CASIM for autoregressive generators (T2MGPT, CoMo) and diffusion generators (MDM, MotionDiffuse, MoMask), reports quantitative results on HumanML3D and KIT-ML plus a long-horizon DoubleTake study, and includes attention visualizations as an interpretability check.","tokens_in":15118,"tokens_out":3677,"duration_ms":38425,"significance":"If the central claim holds, CASIM is a useful plug-in contribution: it is architecture-agnostic, the authors release code, and the method produces large improvements on several strong baselines, notably MDM and T2MGPT on R-Precision and MM-Dist. The attention visualization in Appendix D is a valuable mechanistic sanity check. However, the central attribution to composite-aware token injection is not yet isolated from model capacity, and the paper's own Appendix C shows a case where the module hurts text-motion alignment, so the contribution is promising but needs additional controlled experiments and more careful claims.","major_comments":[{"comment":"The central claim that token-level composite-aware injection, rather than added model capacity, drives the reported gains is not yet supported. CASIM adds a linear projection, extends the conditioning sequence from one [CLS] token to N token embeddings, and inserts MHSA/MHCA blocks into each baseline, while Section 4.1 only states that other settings follow the baseline. No baseline is augmented with an equal-parameter conditioning module that still uses only the fixed-length [CLS] embedding. Please add a parameter-matched control, for example a baseline whose text side is expanded to the same number of tokens and given the same attention/MLP machinery but with those tokens derived from [CLS] or from input-independent learned vectors. Without such a control, the improvements in Tables 3 and 4 are confounded with capacity.","section":"Section 4.1; Section 3.2 Eq. (2); Section 3.3 Eqs. (4)-(5)"},{"comment":"Baseline numbers for the same method and dataset differ across tables without a clear explanation. MDM is reported with Top1 R-Precision 0.455 in Table 1 but 0.471 in Table 3, and T2MGPT is reported with 0.491 in Table 1 but 0.484 in Table 3. Table 5 suggests these correspond to different MDM diffusion-step settings and T2MGPT teacher-forcing settings, but the paper never states which configuration is used in Table 1 versus Table 3, so the reader cannot verify that the CASIM comparison is against the same baseline configuration. Please clarify the configuration used in each table, and report the CASIM-integrated model under the identical baseline setting for every comparison.","section":"Table 1 vs. Table 3; Table 5"},{"comment":"The paper claims 'consistent improvements' but reports only point estimates from 20 repeats, with no standard deviations, confidence intervals, or significance tests. Several gains are small enough to be within run-to-run noise, such as MoMask FID 0.064 to 0.057 in Table 3. Please provide variance measures or statistical tests for the main comparisons, at least for the headline metrics (R-Precision Top1, FID, MM-Dist).","section":"Section 4.2; Tables 3 and 4"},{"comment":"The Abstract and Section 5 state that CASIM 'consistently improves' motion quality and text-motion alignment across state-of-the-art methods, but Appendix C shows CASIM-MLD degrades R-Precision (0.469 to 0.452) and MM-Dist (3.282 to 3.389) on HumanML3D. The explanation that fixed-length motion latents limit fine-grained text-motion correspondence is reasonable, but the claim of consistency must be qualified to the tested model families. This limitation also reinforces the need to characterize when the mechanism helps rather than to present it as universally effective.","section":"Appendix C, Table 9; Abstract; Section 5"}],"minor_comments":[{"comment":"There is a typo in 'directionh adds another layer of difficulty' that should read 'direction adds another layer of difficulty.'","section":"Section 2.1"},{"comment":"The word 'nuaunced' should be 'nuanced.'","section":"Section 4.3"},{"comment":"The text says 'it sill relies on motion blending' and should read 'it still relies on motion blending.'","section":"Section 5, Limitations"},{"comment":"The symbol τ denotes the diffusion step in Eq. (3) but denotes the teacher-forcing coefficient in Section 4.2; the footnote clarifies this, but the notation clash is confusing within the same paper.","section":"Section 3.3 vs. Section 4.2"},{"comment":"CoMo's baseline Top1 R-Precision is 0.422 in Table 2 but 0.399 in Table 4 on KIT-ML; if these correspond to different checkpoint or evaluation settings, this should be stated.","section":"Table 2 vs. Table 4"},{"comment":"The word-cloud analysis in Figure 4 is qualitative; including a small quantitative analysis, such as the fraction of test prompts where the most-attended token is an action verb or modifier, would strengthen the claim that attention focuses on motion-critical words.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is simple and appealing: instead of feeding the fixed CLIP sentence embedding into a text-to-motion model, keep the token-level text embeddings and let each motion token attend over them. That buys you 'left hand' vs 'right hand' control and, on the face of it, large improvements in R-Precision and FID across five quite different base models (MDM, MotionDiffuse, T2MGPT, CoMo, MoMask). The paper is honest enough to include a limitation appendix where CASIM hurts MLD's R-Precision, and it does real ablations: encoder vs decoder injection, teacher-forcing rates, CoMo without GPT keywords, CLIP layer choice. The attention visualizations are a nice touch. If the numbers hold, this is a useful plug-in module for the field.\n\nBut the numbers do not clearly hold as reported. The biggest problem is Table 3: the MDM baseline row uses the encoder architecture (FID 0.325, R-Precision 0.471) while the CASIM row uses the decoder architecture (FID 0.165, R-Precision 0.502). The text even says 'FID: 0.325→0.165' as if that were an apples-to-apples comparison. It isn't. The encoder-with-CASIM row in Table 5 actually shows R-Precision dropping from 0.471 to 0.463, with FID improving to 0.265. That's a much more modest and mixed story. Something similar is going on with T2MGPT baselines, which differ across Tables 1, 3, and 5 without any explanation of which configuration is being used. This is not a nitpick; the central claim depends on comparing the same base model with and without CASIM.\n\nThe other soft spot is the one the stress-test note flags: CASIM adds a text-motion aligner plus linear projections, and no baseline gets equivalent added capacity. So part of the gain could just be more parameters. The MLD result actually argues against a pure-capacity story—extra parameters didn't help there—but it doesn't isolate the mechanism in the successful cases. A parameter-matched control that still uses the fixed [CLS] embedding would settle this, and the paper doesn't have one. Also, despite saying results are averaged over 20 runs, there are no error bars or significance tests, which makes it hard to judge whether a 0.01 R-Precision change is real.\n\nBottom line: this is a promising method with a genuine mechanic to test, and the authors are thinking clearly about where it works and where it doesn't. But the current write-up overstates the evidence, and the main tables need cleanup before the claims can be trusted. I'd send it to serious peer review—the idea deserves referee time—but I'd ask for a revised version that fixes the baseline inconsistencies, adds a capacity-controlled ablation, and reports variance. Who's it for? Anyone working on text-to-motion conditioning or on semantic injection for generative models more broadly. They should read it, but with the tables in hand.","headline":"CASIM's token-level injection idea is worth taking seriously, but the main comparison tables mix architectures and baselines in ways that inflate the reported gains.","tokens_in":15643,"tokens_out":3674,"would_cite":true,"duration_ms":36772,"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":"Replacing fixed-length CLIP text embeddings with token-level embeddings plus a learned text-motion aligner improves motion quality, alignment, and retrieval across five text-to-motion models.","keywords":["text-to-motion generation","semantic injection","token-level text embeddings","text-motion alignment","multi-head attention","diffusion models","autoregressive models","human motion generation"],"falsifier":"Train a parameter-matched control in which the same added aligner layers and projections are fed a fixed-length [CLS] condition, for example the same vector repeated at every position, instead of per-token embeddings. If this control matches CASIM's R-Precision and FID gains, the token-level alignment mechanism is not what is doing the work; if it falls short, the composite-aware injection is supported.","tokens_in":14671,"feed_emoji":"🏃","tokens_out":7903,"duration_ms":68330,"temperature":0.7,"pith_summary":"This paper argues that the standard practice of compressing a text prompt into one fixed-length CLIP [CLS] embedding loses the composite, order-sensitive structure of human motion descriptions, and that this loss limits both the quality and the controllability of generated motion. The proposed CASIM mechanism instead keeps the per-token text embeddings and adds a learned text-motion aligner that lets each motion frame attend to the words most relevant to it. Across five state-of-the-art generators—MDM, MotionDiffuse, T2MGPT, CoMo, and MoMask—on the HumanML3D and KIT benchmarks, the paper reports consistent gains in motion quality, text-motion alignment, and retrieval. If correct, the mechanism offers finer control, such as distinguishing 'left hand' from 'right hand' in a prompt, without changing the underlying generator.","feed_headline":"Word-level text beats one-vector prompts for motion generation","feed_subtitle":"Each motion frame attends to individual words, improving text-motion alignment in diffusion and autoregressive models.","key_machinery":"The load-bearing object is the text-motion aligner: a multi-head attention module that uses each motion token as the query and the full set of text token embeddings as keys and values. It is paired with a composite-aware text encoder that outputs per-token embeddings from a pretrained text encoder instead of a single [CLS] vector. The aligner is what converts a static global condition into a dynamic, per-frame condition, and it is the component that the paper claims carries the improvement.","core_discovery":"The central claim is that fixed-length semantic injection is the bottleneck. Where prior text-to-motion models use the CLIP [CLS] token as a global condition, CASIM preserves token-level embeddings and learns a dynamic correspondence between text tokens and motion tokens through multi-head attention—self-attention when the generator is autoregressive, cross-attention when it is a diffusion decoder—so each motion frame can draw on the words that describe it. The paper reports that this injection mechanism, applied on top of existing models with their hyperparameters unchanged, improves R-Precision, FID, and MM-Distance on HumanML3D and KIT-ML, and that attention visualizations show text tokens such as 'wave' attending to early frames and 'sit' to later frames.","pith_inferences":["The paper leaves open whether the gains come from the alignment mechanism or from added parameters: a control giving each baseline the same extra attention layers and projections while keeping fixed-length conditioning would settle this.","The BERT-versus-CLIP ablation hints that the approach is not tied to CLIP; if so, stronger contextualized text encoders could be swapped in without redesigning the aligner.","The MLD result suggests CASIM's benefit depends on motion representations that keep per-frame or per-token structure, so applying it to latent-space generators would require token-wise motion latents rather than a fixed-length vector.","The learned word-frame attention could double as an interpretability and editing tool, letting users see which prompt words drive which motion phases and re-weighting attention to edit motion."],"forward_implications":["Gains transfer across generator families: the paper reports improved R-Precision, FID, and MM-Distance for two diffusion models (MDM, MotionDiffuse) and three autoregressive or tokenized models (T2MGPT, CoMo, MoMask) on HumanML3D and KIT-ML.","Prompt differences of a single word, such as 'left hand' versus 'right hand,' become distinguishable in the generated motion, giving users finer control over body side, direction, and action order.","External semantic augmentation becomes less necessary: CASIM-CoMo without keyword augmentation beats the keyword-augmented CoMo baseline on R-Precision and FID.","In long-term generation via DoubleTake, CASIM improves the quality and text alignment of the generated motion clips, with the largest FID gain at a 20-frame handshake (0.953 to 0.463).","Because CASIM changes only the semantic injection, it can be layered onto existing generators without retraining their motion tokenizers or denoisers from scratch."],"supporting_citations":[{"why":"Source of the pretrained text encoder whose [CLS] embedding is the fixed-length baseline and whose token embeddings CASIM reuses.","marker":"(Radford et al., 2021)"},{"why":"MDM baseline; the paper reports large FID and R-Precision gains from CASIM on both encoder and decoder variants.","marker":"(Tevet et al., 2023)"},{"why":"T2MGPT autoregressive baseline; CASIM improves its R-Precision from 0.491 to 0.539 on HumanML3D.","marker":"(Zhang et al., 2023a)"},{"why":"CoMo baseline; CASIM-CoMo without external keyword augmentation outperforms keyword-augmented CoMo on R-Precision and FID.","marker":"(Huang et al., 2024)"},{"why":"MoMask baseline; CASIM improves its already strong FID, FID, and retrieval scores.","marker":"(Guo et al., 2023)"},{"why":"MotionDiffuse diffusion baseline; CASIM improves its R-Precision, FID, and MM-Distance.","marker":"(Zhang et al., 2022)"},{"why":"HumanML3D dataset and the text-motion matching network used to compute all metrics.","marker":"(Guo et al., 2022a)"},{"why":"KIT-ML dataset; provides the small-scale benchmark where CASIM-MDM shows large R-Precision gains.","marker":"(Plappert et al., 2016)"},{"why":"DoubleTake long-term generation framework; CASIM's motion clips improve FID and text alignment within it.","marker":"(Shafir et al., 2024)"},{"why":"MLD baseline; its failure to benefit from CASIM marks the paper's stated limitation tied to fixed-length motion latents.","marker":"(Chen et al., 2023)"}],"fun_headline_variants":["Word-level attention improves text-to-motion generation","Each motion frame picks its own text words","Token-level injection boosts motion quality and alignment","Dynamic text-motion attention beats fixed CLIP embedding","Attend to words, not one vector, for better motion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper attributes the gains to token-level semantic injection, but because each CASIM-augmented model receives extra attention layers and projections that the baseline lacks, the improvements could in principle come from the added model capacity rather than from the alignment mechanism itself.","fun_headline_variants_meta":{"raw":{"variants":["Word-level attention improves text-to-motion generation","Each motion frame picks its own text words","Token-level injection boosts motion quality and alignment","Dynamic text-motion attention beats fixed CLIP embedding","Attend to words, not one vector, for better motion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000278,"raw_usage":{"total_tokens":1623,"prompt_tokens":883,"completion_tokens":740,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":667}},"tokens_in":499,"tokens_out":740,"duration_ms":7609,"temperature":1.0,"reasoning_tokens":667,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T13:28:37.411544+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a parameter-matched control in which the same added aligner layers and projections are fed a fixed-length [CLS] condition, for example the same vector repeated at every position, instead of per-token embeddings. If this control matches CASIM's R-Precision and FID gains, the token-level alignment mechanism is not what is doing the work; if it falls short, the composite-aware injection is supported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MDM baseline; the paper reports large FID and R-Precision gains from CASIM on both encoder and decoder variants."},{"cited_title":"Como: Controllable motion generation through language guided pose code editing, 2024","cited_arxiv_id":null,"evidence_quote":"CoMo baseline; CASIM-CoMo without external keyword augmentation outperforms keyword-augmented CoMo on R-Precision and FID."},{"cited_title":"G., Wang, S., and Cheng, L","cited_arxiv_id":null,"evidence_quote":"MoMask baseline; CASIM improves its already strong FID, FID, and retrieval scores."},{"cited_title":"The KIT motion-language dataset","cited_arxiv_id":null,"evidence_quote":"KIT-ML dataset; provides the small-scale benchmark where CASIM-MDM shows large R-Precision gains."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DoubleTake long-term generation framework; CASIM's motion clips improve FID and text alignment within it."},{"cited_title":"Executing your commands via motion diffusion in latent space","cited_arxiv_id":null,"evidence_quote":"MLD baseline; its failure to benefit from CASIM marks the paper's stated limitation tied to fixed-length motion latents."}],"review_version":1}