{"id":"fb621956-e068-483d-b0ec-407f5eb60d3d","arxiv_id":"2506.17912","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"PlanMoGPT combines progressive coarse-to-fine token planning with a flow-enhanced motion tokenizer to achieve state-of-the-art text-to-motion generation, especially on long sequences.","lead":"This paper presents PlanMoGPT, a language-model-based system that generates 3D human motion from text by first sketching a coarse movement plan and then refining it into fine detail with a learned flow-based decoder. It reports large gains in quality and diversity over prior methods on long motion sequences, though the new long-sequence benchmarks are built by the authors themselves.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Long-sequence SOTA claim rests on self-built benchmarks with limited validation and only two retrained baselines; independent evaluation is needed.","rationale":"The reader's weakest assumption identifies the self-built HumanML3D++/KIT-ML++ benchmarks and limited baseline coverage as the key vulnerability. I agree. The abstract's most concrete quantitative claims—63.8% FID reduction and 49.9% diversity increase—are reported only on these new datasets, so if they are unreliable or unrepresentative, the headline contribution fails. The paper's own text limits the validation to 100 samples with 86% reliability, and only MoMask and T2M-GPT are retrained; other strong baselines are absent. There is also a potential metric issue: standard FID extractors are designed for short motions, and the paper does not specify how long-sequence FID is computed. These concerns are not accusations of wrongdoing; they are standard empirical robustness requirements for a new benchmark. A conditional verdict is appropriate. I would not reject the paper because the method is well-motivated and shows credible gains on standard HumanML3D; but the long-sequence SOTA claim needs independent validation. Hence verdict unchanged from the reader's CONDITIONAL.","tokens_in":16012,"tokens_out":7417,"duration_ms":75251,"concrete_test":"Evaluate PlanMoGPT and the retrained MoMask/T2M-GPT on the Infinite Motion long-sequence benchmark (or an independently constructed long-sequence split) using the original baselines' training recipes adapted for long sequences (extended positional embeddings and longer training). If PlanMoGPT's FID improvement over MoMask/T2M-GPT is below about 20% relative, or is not reproduced, the long-sequence SOTA claim is not robust; if it is reproduced, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline result—63.8% FID improvement (0.380 to 0.141) and 49.9% MModality gain on long-sequence generation—is demonstrated only on HumanML3D++ and KIT-ML++, datasets constructed by the authors. Section 4.1 reports that only 100 motion-text pairs were manually inspected, with 86% deemed reliable; no description is given of how clip boundaries are blended, how GPT-4 merged texts are aligned to transitions, or how FID is computed for sequences up to 50s. The standard T2M feature extractor used for FID is trained on short motions, so long-sequence FID values may not be comparable to established benchmarks. Furthermore, the only baselines retrained on these new splits are MoMask and T2M-GPT; no BAMM, MotionDiffuse, MotionGPT, or long-motion-specific method is included. The original training recipes for these baselines target at most 10s motions, and no adaptation (e.g., extended positional encoding, longer training, adjusted batch size) is described. If the dataset is noisy or the baselines are under-tuned, the 63.8% FID improvement is not evidence of SOTA against the broader field. This is load-bearing because the abstract's primary quantitative claim is precisely this long-sequence improvement.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PlanMoGPT, an LLM-based text-to-motion framework that combines progressive multi-interval planning (generating tokens at intervals T=4, T=2, then T=1) with a flow-enhanced motion tokenizer (downsampling rate 2, codebook size 4096, and a flow-matching decoder). The authors claim state-of-the-art results on HumanML3D and introduce two new long-sequence benchmarks, HumanML3D++ and KIT-ML++, built by concatenating 2-5 motion clips and merging texts with GPT-4. The headline claim is a 63.8% FID improvement on long-sequence generation (0.380 to 0.141) and a 49.9% MModality improvement, which they interpret as resolving a diversity-quality trade-off in non-LLM methods.","tokens_in":16282,"tokens_out":7376,"duration_ms":78452,"significance":"If the long-sequence results hold, the paper would make a useful contribution by addressing the local-dependency problem in token-based motion generation and by demonstrating a concrete mechanism for coarse-to-fine LLM planning. The paper's strengths include a well-motivated and clearly ablated progressive planning design, competitive short-sequence results on HumanML3D against strong baselines, and detailed ablation studies covering codebook size, downsampling rate, plan intervals, and flow-matching steps. However, the long-sequence state-of-the-art claim currently rests on self-built benchmarks with only two retrained baselines, and the headline percentage improvement is arithmetically incorrect. These issues are load-bearing for the paper's central claim, so the manuscript needs substantial revision.","major_comments":[{"comment":"The long-sequence SOTA claim is supported by only two retrained baselines, T2M-GPT and MoMask, on the self-built HumanML3D++ and KIT-ML++ datasets. No comparisons are provided against BAMM, MotionDiffuse, MotionGPT, or long-motion-specific methods such as Infinite Motion [25], which is cited for the dataset construction. Consequently, the abstract's claim of state-of-the-art performance on long-sequence generation is not established against the broader field; the authors should either add these baselines or temper the claim to a comparison against the two retrained methods.","section":"§4.1, Table 2"},{"comment":"The paper does not specify how FID and MM-Dist are computed for sequences up to 50 seconds. The standard T2M feature extractor used in these metrics was trained on short HumanML3D motions, and it is not stated whether it is applied to full long sequences or to chunks, nor whether its validity for long inputs was verified. Additionally, the flow-matching decoder is described as splitting motions into 64-frame clips and re-stitching them after inference; the same detail is not given for evaluation. This is load-bearing because all long-sequence quality numbers depend on an evaluation protocol that is currently underspecified.","section":"§4.1, evaluation protocol"},{"comment":"The new benchmarks are validated with only 100 manually checked motion-text pairs, of which 86% were deemed reliable. No description is provided of how clip boundaries are blended, how the GPT-4 merged texts are aligned to the transitions, or whether the 14% unreliable pairs are excluded from the test set. At this validation scale, the benchmark noise could materially affect the reported FID values; the authors should provide a more systematic validation protocol or release the datasets so other groups can independently assess them.","section":"§4.1, dataset quality"},{"comment":"The claimed improvement \"from 0.380 to 0.141\" is a relative FID reduction of (0.380 - 0.141)/0.380 = 0.629, i.e., 62.9%, not 63.8%. The abstract and contributions section should be corrected to state 62.9%.","section":"Abstract, Table 2"},{"comment":"T2M-GPT and MoMask are retrained using their source code on the new long-sequence datasets, but the paper does not describe any adaptations for sequences up to 50 seconds, such as extended positional encodings, longer training schedules, or adjusted batch sizes. Since these methods were originally designed for motions of at most about 10 seconds, the comparison may systematically disadvantage them. The authors should state what adaptations were made and provide evidence that the retrained baselines are not under-tuned.","section":"§4.1, baseline retraining"}],"minor_comments":[{"comment":"The interpolation y_t is not explicitly defined between t=0 and t=1; please specify that y_t = (1-t) y_0 + t y_1 (or otherwise) so that the conditional flow-matching objective is unambiguous.","section":"§3.1, Eq. (2)"},{"comment":"The offset b is sampled from {1,...,Tmax}, but Tmax is not defined; if it is the largest plan interval, please state this explicitly.","section":"§3.2"},{"comment":"The checkmark layout in Table 4 makes it difficult to determine which interval combination each row corresponds to; please align the checkmarks explicitly with the T=8, T=4, and T=2 columns.","section":"§4.5, Table 4"},{"comment":"There is a typo in the text: \"an MMdoality of 2.524\" should read \"an MModality of 2.524\".","section":"§4.2"},{"comment":"Table 3 states that all compared models are base versions, but the surrounding text refers to \"MoMask and BAMM\" without this qualification; please make the distinction between base and full models consistent.","section":"§4.3"}],"recommendation":"major_revision","confidential_remarks":"The core methodological idea is interesting, and the short-sequence experiments are solid enough to support a competitive claim. However, the long-sequence SOTA claim is the paper's headline contribution and it currently rests on self-built benchmarks with limited validation and only two retrained baselines. I would encourage the editor to ask the authors to release the new datasets and evaluation code during revision; without independent verification, the 62.9% FID improvement figure should not be presented as a field-wide state-of-the-art result. The missing comparison with Infinite Motion [25], which is cited for dataset construction, is a notable omission that should be addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know about arXiv:2506.17912. First, the core engineering is real: the authors combine progressive multi-interval token planning (predict every 4th token, then every 2nd, then all) with a VQ-VAE that uses a 4096-codebook at stride 2 and a flow-matching decoder, and the package beats the published diffusion and token baselines on the standard HumanML3D benchmark (R@1 52.6, FID 0.048, MModality 2.971). That is a genuine result for LLM-based text-to-motion. Second, the headline long-sequence claim—63.8% FID improvement, 49.9% diversity gain—is built on HumanML3D++ and KIT-ML++, self-constructed datasets with only two baselines retrained. That part is not yet established.\n\nWhat is actually new: the progressive planning mechanism is a clean idea and the ablation in Table 4 shows the 4+2 interval combination gives the best R@1/FID trade-off. The flow-enhanced tokenizer beats residual VQ-VAE in generation quality even when the residual stack reconstructs better (Table 7), which is a useful data point for the field. The paper is also honest: it openly reports suboptimal KIT-ML results and attributes them to dataset scale/frame rate, and the diversity-quality curve in Figure 3 is informative.\n\nSoft spots, in order of severity.\n\n1. The abstract's '63.8%' is wrong: 0.380 to 0.141 is 62.9%. Minor but sloppy.\n\n2. The long-sequence benchmark is under-documented. Only 100 pairs were manually checked (86% reliable). No details on how clip boundaries are blended, how GPT-4 merged texts align to transitions, or how FID is computed for 50s sequences with a feature extractor trained on 10s motions. The two retrained baselines use their original recipes with no adaptation for longer sequences, so they are at a disadvantage. This makes the long-sequence SOTA claim provisional.\n\n3. The 'resolves the diversity-quality dilemma' phrasing is too strong given the above; the standard HumanML3D evidence supports a softer version.\n\nThe math in the method is straightforward and the evaluation on held-out splits is a measurement, not a fit, so the circularity concern does not land. The citation pattern looks normal for the subfield.\n\nWho it's for: anyone working on LLM-based motion generation or tokenized temporal generation. It deserves a serious referee: the core method is solid and the long-sequence benchmark, if properly documented, would be a useful resource. I'd send it to review with the expectation of a major revision that addresses the benchmark construction and the arithmetic.","headline":"A well-engineered LLM text-to-motion system with real gains on HumanML3D, but the headline long-sequence SOTA rests on self-built benchmarks that need stronger validation.","tokens_in":16819,"tokens_out":2232,"would_cite":true,"duration_ms":23416,"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":"PlanMoGPT claims that LLM-based text-to-motion generation becomes state-of-the-art when motion tokens are planned coarse-to-fine and decoded with flow matching.","keywords":["text-to-motion synthesis","large language models","motion tokenization","VQ-VAE","flow matching","progressive planning","long-sequence motion generation","motion diversity"],"falsifier":"Retrain MoMask, BAMM, MotionDiffuse, and T2M-GPT on HumanML3D++ and KIT-ML++ with matched compute and hyperparameter tuning, and evaluate all methods on an independently human-validated long-sequence test set built without language-model text merging. If the best retrained baseline comes within noise of FID 0.141 on HumanML3D++, or if removing progressive planning from PlanMoGPT leaves FID unchanged, then the paper's central attribution fails.","tokens_in":15835,"feed_emoji":"🤸","tokens_out":8997,"duration_ms":86361,"temperature":0.7,"pith_summary":"This paper tries to show that LLMs can beat non-LLM methods in text-to-motion generation if motion is represented as fine-grained discrete tokens that are generated in a coarse-to-fine plan, then refined by a flow-matching decoder. It identifies token granularity as the bottleneck: too-fine tokens make LLMs copy nearby tokens and drift from the text, while too-coarse tokens lose detail. The proposed PlanMoGPT first generates a sparse skeleton plan, then fills in half-resolution tokens, then the full sequence, while a 4096-entry codebook at half downsampling with flow-based decoding preserves motion nuance. On long-sequence benchmarks built by stitching 2-5 motion clips, the paper reports FID improving from 0.380 to 0.141 and diversity rising 49.9% relative to a leading baseline. If right, this would make LLM-based motion generation both higher quality and more diverse than current diffusion approaches.","feed_headline":"LLM motion generator cuts long-sequence FID from 0.380 to 0.141","feed_subtitle":"Progressive coarse-to-fine planning plus a finer codebook also lifts diversity by 49.9%.","key_machinery":"The central machinery is a two-part pipeline. First, a fine-grained VQ-VAE tokenizer with downsampling rate 2 (rather than 4) and codebook size 4096 (rather than 512) converts motion into discrete tokens with less quantization loss; a flow-matching decoder then treats the token-decoded motion as time-zero of an ODE and integrates a learned vector field to add back detail. Second, the LLM generates tokens in three hierarchy levels: every-4-frames plan tokens, every-2-frames plan tokens, and the full token sequence, with special transition tokens separating levels. The sparse upper levels remove the strong local autocorrelation that makes fine-grained token prediction drift from the text, while lower levels supply detail; the ablation shows interval-4 plans improve semantic alignment, interval-2 plans improve FID, and combining them is cumulative.","core_discovery":"The paper's central claim is that the apparent ceiling on LLM-based text-to-motion generation is set by motion token granularity, not by LLM capacity. Fine-grained discrete motion tokens make next-token prediction too locally predictable, so the model leans on adjacent tokens and drifts from the text; coarse tokens erase detail. PlanMoGPT attacks both sides: a tokenizer that runs at half the downsampling rate with an eight-times larger codebook (4096 entries) keeps detail, and a flow-matching decoder refines the coarse reconstruction back toward real motion; on the language side, an LLM first emits a sparse every-4-frames plan, then every-2-frames, then the full token sequence, so each level anchors the next. On the new long-sequence benchmarks, the method reports FID 0.141 versus 0.380 for MoMask on HumanML3D++, and MModality 2.538 versus 1.693, which the paper reads as resolving the diversity-quality dilemma. The authors also report suboptimal results on KIT-ML, attributing the shortfall to that dataset's smaller scale and lower temporal resolution.","pith_inferences":["Editorial inference: the coarse-to-fine plan-then-fill recipe is a general strategy for autoregressive generation over any locally autocorrelated discrete sequence, so it could transfer to video token prediction, gesture generation, or long-form audio.","Editorial inference: the ablation evidence suggests the 4096-entry codebook helps only when paired with a decoder that can repair quantization loss; a fair external test would vary codebook size and decoder type independently across more LLM backbones.","Editorial inference: because HumanML3D++ and KIT-ML++ were made by stitching clips and merging text with an LLM, independent human annotation of a larger sample would clarify whether the long-sequence gains reflect genuine long-motion understanding or distributional artifacts of the new benchmarks.","Editorial inference: the method's flat FID-versus-diversity curve predicts that pushing non-LLM baselines to comparable MModality by raising sampling temperature should degrade their FID; if it does not, the claimed resolution of the dilemma would need revision."],"forward_implications":["If the reported numbers hold, LLM-based text-to-motion generators can match or beat diffusion-based methods on quality while generating markedly more diverse motions, so the previous gap between LLM and non-LLM routes is not a hard limit.","Long-sequence generation becomes a relative strength rather than a failure mode: the biggest FID gains appear on HumanML3D++ and KIT-ML++, where sequences reach roughly 50 seconds.","The diversity-quality trade-off observed in non-LLM methods is not intrinsic: PlanMoGPT keeps FID and R-precision nearly flat as multimodality rises, according to the paper's diversity analysis.","The flow-enhanced decoder, not just the larger codebook, drives the FID gain: ablations show the flow version beats residual VQ-VAE at the same codebook and downsampling settings, and the chosen 4096/2 configuration gives the best generation FID.","Training is comparatively practical: the paper reports convergence in 25 epochs versus 500 for MoMask-style training, which would make the LLM route more accessible."],"supporting_citations":[{"why":"T2M-GPT is the token-based baseline whose local-dependency failure motivates the granularity diagnosis and whose codebook and downsampling choices the tokenizer is compared against.","marker":"[56]"},{"why":"MoMask is the non-LLM baseline with the best reported numbers, and its residual VQ-VAE is the approach PlanMoGPT's flow-enhanced tokenizer is designed to outperform; its FID and diversity numbers anchor the central comparisons.","marker":"[13]"},{"why":"MotionGPT supplies the prior LLM-based framework with coarse tokenization, the approach that PlanMoGPT refines with progressive planning.","marker":"[19]"},{"why":"Flow matching provides the ODE and vector-field machinery used by the flow-enhanced motion decoder to reconstruct detail lost in quantization.","marker":"[28]"},{"why":"The procedure for merging 2-5 motion clips into long sequences is adopted to build the HumanML3D++ and KIT-ML++ benchmarks that carry the headline long-sequence results.","marker":"[25]"},{"why":"HumanML3D is the standard benchmark on which the method reports state-of-the-art R-precision and diversity, and it is the source of the HumanML3D++ long-sequence split.","marker":"[14]"},{"why":"KIT-ML is the smaller, lower-frame-rate benchmark used to build KIT-ML++ and to test whether the method transfers to lower temporal resolution.","marker":"[32]"}],"fun_headline_variants":["PlanMoGPT: FID 0.380 to 0.141 on long sequences","Text-to-motion LLM: 63.8% FID win, 49.9% diversity gain","Coarse-to-fine planning fixes LLM text-to-motion gap","Token granularity is key: PlanMoGPT sets new bar","LLM motion: diversity-quality trade-off resolved"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that HumanML3D++ and KIT-ML++, built by stitching 2-5 motion clips and merging their texts with an LLM, fairly represent long text-to-motion generation; only a 100-sample subset was manually checked (86% reliable) and only two baselines were retrained on these splits.","fun_headline_variants_meta":{"raw":{"variants":["PlanMoGPT: FID 0.380 to 0.141 on long sequences","Text-to-motion LLM: 63.8% FID win, 49.9% diversity gain","Coarse-to-fine planning fixes LLM text-to-motion gap","Token granularity is key: PlanMoGPT sets new bar","LLM motion: diversity-quality trade-off resolved"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000918,"raw_usage":{"total_tokens":3990,"prompt_tokens":1045,"completion_tokens":2945,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":661,"completion_tokens_details":{"reasoning_tokens":2843}},"tokens_in":661,"tokens_out":2945,"duration_ms":24163,"temperature":1.0,"reasoning_tokens":2843,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:22:31.271659+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain MoMask, BAMM, MotionDiffuse, and T2M-GPT on HumanML3D++ and KIT-ML++ with matched compute and hyperparameter tuning, and evaluate all methods on an independently human-validated long-sequence test set built without language-model text merging. If the best retrained baseline comes within noise of FID 0.141 on HumanML3D++, or if removing progressive planning from PlanMoGPT leaves FID unchanged, then the paper's central attribution fails.","supporting_citations":[{"cited_title":"Infinite Motion: Extended Motion Generation via Long Text Instructions","cited_arxiv_id":"2407.08443","evidence_quote":"The procedure for merging 2-5 motion clips into long sequences is adopted to build the HumanML3D++ and KIT-ML++ benchmarks that carry the headline long-sequence results."},{"cited_title":"Generating diverse and natural 3d human motions from text","cited_arxiv_id":null,"evidence_quote":"HumanML3D is the standard benchmark on which the method reports state-of-the-art R-precision and diversity, and it is the source of the HumanML3D++ long-sequence split."}],"review_version":1}