{"id":"e3dc84a5-16c2-49bf-9f11-33b5fcff6695","arxiv_id":"2508.07863","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Being-M0.5 combines part-aware residual quantization with a 5M-sequence web-video dataset to reach real-time, part-controllable 3D motion generation, though its state-of-the-art claim does not hold on every standard benchmark.","lead":"This paper presents Being-M0.5, a 7-billion-parameter model that converts text instructions, video, and starting poses into 3D human motion in real time, together with a new dataset of over five million motion clips. It is worth reading because controllability and latency are the two bottlenecks that keep motion-generation models out of real products, and this work attacks both directly.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The controllability claims rest on benchmarks drawn from the same auto-annotation pipeline used for training; without an outside-distribution human-validated part-level test, I2PM/I2U gains may reflect template matching rather than kinematic control.","rationale":"The paper's strongest claims are (1) real-time controllable generation and (2) state-of-the-art performance on multiple tasks, enabled by PRQ and HuMo100M. The reader's weakest-assumption analysis correctly identifies the auto-annotation/evaluation pipeline as the most load-bearing element: the new benchmarks are generated by the same machinery that produced the training data, so the model's apparent controllability gains could be due to in-distribution template mapping rather than genuine kinematic understanding. Section 4.2.1 describes part descriptions from Gemini-1.5-Pro and PoseScript, with no reported human validation; Section 5.1.1 samples all new testbeds from HuMo100M; Table 6 demonstrates the I2PM gains only on that in-distribution benchmark. This is not an internal inconsistency in the PRQ mechanism, but it does mean the central controllability claim has not been shown to transfer outside the exact annotation distribution. I also note the identical Table 1 rows for LMM and MotionGPT-v2 as a provenance concern, but it is secondary: even if the baselines are corrected, the controllability benchmarks themselves remain circular. The paper earns credit for Section 5.4.4 acknowledging PRQ's FID trade-off, and Table 4 provides some independent support for the tokenizer, but those do not address the external-validity gap for part-level control. The reader's CONDITIONAL verdict is appropriate; the condition should be external, human-validated evaluation of part-level and unseen-motion control, plus artifact release. My read does not move the verdict.","tokens_in":20414,"tokens_out":4007,"duration_ms":48276,"concrete_test":"Independently re-annotate a held-out set of 200 web-video motion clips (not from HuMo100M) with part-level text using only human annotators, using the same WHAM/PHC SMPL extraction as the paper. Run Being-M0.5-PRQ4, Being-M0.5-PRQ4 without part labels, and a strong baseline such as ParCo or MotionGPT on these clips in a blind protocol; measure left/right part accuracy and human-rated kinematic plausibility. If PRQ4's advantage over the no-part-label variant and over baselines disappears or drops below significance on human-annotated out-of-distribution prompts, the controllability claim is not supported outside the template distribution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of controllable, state-of-the-art motion generation depends on the assumption that the new HuMo-I2PM, I2M, I2LM, and I2U benchmarks measure genuine kinematic control rather than template-matching within the annotation distribution. Section 4.2.1 constructs all part-level labels with Gemini-1.5-Pro prompts plus PoseScript rule templates; Section 5.1.1 then samples the testbeds from the same HuMo100M source and the same hierarchical description styles used for training. Table 6's central I2PM result (PRQ4 with part labels: FID 1.831 vs. 4.281 without part labels) therefore conflates the architecture's benefit with the model's ability to exploit the statistical regularities of the auto-generated prompts. The only external part-level probe, the left-right exchange test in Section 5.3.3, uses 50 sentences and is too small and too close to the training template style to rule out interpolation. The I2U 'unseen' testbed is also drawn from HuMo100M (Section 5.1.1), so novelty is defined relative to the same pipeline. If this concern lands, the headline 'controllable' claim and the I2PM/I2U improvements are not established; PRQ's reconstruction gains in Table 4 would remain, but the controllability and generalization claims would need outside-distribution support. The paper reports no human validation rate for the generated part descriptions, and no artifacts are released to permit independent checks.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents Being-M0.5, a 7B vision-language-motion model trained on a newly collected HuMo100M dataset (5M motion sequences, 100M instructional instances). The central contributions are a part-aware residual quantization (PRQ) tokenizer that decomposes motion into five body-part code streams with shared joints, a three-stage curriculum including multi-task instruction tuning, and a HuMo263 motion feature that avoids inverse kinematics. The authors report state-of-the-art or competitive results on HumanML3D/KIT-ML T2M, new in-house benchmarks for instruct-to-motion, part-motion, long-motion, and unseen-motion, reconstruction, prediction, and captioning, and real-time 20–28 FPS inference on the 7B model. Extensive ablations of quantization layers, shared joints, part labels, visual stage, and multi-task composition are also provided.","tokens_in":20746,"tokens_out":6369,"duration_ms":68149,"significance":"If the empirical claims are validated, this is a substantial advance: a million-scale dataset, a compact part-level motion tokenizer, and a real-time 7B model are all useful community assets. The paper's strengths include transparent limitations (e.g., §5.4.4 concedes the FID deficit vs MoMask), extensive ablations (Tables 6–8), and efficiency analysis across GPUs. The main risk is that the evaluation of controllability relies on benchmarks produced by the same automatic annotation pipeline used for training, so the headline 'controllable' and 'unseen generalization' claims require additional outside-distribution human-validated evidence.","major_comments":[{"comment":"The I2PM benchmark is created by sampling 200K instances from HuMo100M, and the part-level labels in HuMo100M are generated with the same Gemini-1.5-Pro prompts and PoseScript rule templates used for training. Thus the central I2PM comparisons in Table 6 (PRQ4 with part labels FID 1.831 vs 4.281 without) do not yet establish kinematic control; they may reflect the model learning template statistics of the annotation pipeline. The paper reports no human validation rate for the generated part descriptions. To support the controllability claim, the authors should provide an outside-distribution part-level test set with human-validated labels and report per-part success metrics. The 50-sentence left-right exchange test in §5.3.3 is too small and too close to the training template style to rule out interpolation.","section":"§5.1.1, §4.2.1, Table 6"},{"comment":"The I2U 'unseen' benchmark is curated by sampling 200K 'novel' sequences from HuMo100M that are excluded from training. Because HuMo100M is also the training distribution, this is a random split rather than an out-of-distribution generalization test. The large FID gap in Table 3 (HuMo100M 8.65 vs MotionX 43.28) may simply reflect distribution overlap between training and test. The authors should define how 'unseen' was measured (e.g., semantic/text-embedding clustering) and evaluate on a genuinely held-out external dataset.","section":"§5.1.1, Table 3"},{"comment":"The paper claims 'state-of-the-art performance across multiple motion generation benchmarks,' but Table 1 shows Being-M0.5-PRQ4 FID 0.056 vs MoMask-RQ6 0.045 on HumanML3D, and §5.4.4 openly concedes that PRQ exhibits higher FID scores on HumanML3D. The SOTA claim is metric-dependent; please qualify it to the metrics where it holds (R@1, R@3, MMDist) or provide an aggregate justification.","section":"Abstract, §5.4.4, Table 1"},{"comment":"HuMo100M integrates existing motion datasets, including Motion-X, 3DPW, PROX, and BEHAVE, and the evaluation uses HumanML3D, KIT-ML, AMASS, and 3DPW. The paper does not disclose whether the test sets of these external benchmarks were excluded from HuMo100M during training. If any of them are included, the benchmark numbers in Tables 1, 4, 5 and Figure 4 could be contaminated. Please state explicitly how each evaluation set was kept out of HuMo100M.","section":"§A.1, §5.1.1, Tables 1 and 4"},{"comment":"The evaluation protocol for the new benchmarks is under-specified. The text says 'all baseline methods utilize identical motion data within each benchmark' but does not state whether MotionGPT/MoMask baselines were fine-tuned on HuMo-I2M/HuMo-I2PM or only evaluated zero-shot. If baselines were not trained on the same instruction templates, the comparison is not apples-to-apples. Specify the training/evaluation protocol for baselines on each new benchmark.","section":"§5.2, Tables 2 and 6"}],"minor_comments":[{"comment":"The title says 'Image-to-Motion (I2M)' but the task is Instruct-to-Motion; please correct the terminology.","section":"§5.2, Table 2 title"},{"comment":"The caption and text use 'HuMo-MP' while §5.1.1 defines 'HuMo-MPI'; standardize the naming.","section":"§5.2, Figure 4 caption"},{"comment":"The body-part bullet list uses 'Left Hand'/'Right Hand' while the main text uses 'left arm'/'right arm'; align the terminology.","section":"Appendix B"},{"comment":"Typos: 'sequential stepts' and 'ecent efforts' should read 'sequential steps' and 'recent efforts'.","section":"§4.1, §1"},{"comment":"The reconstruction loss uses \\tilde{m}_j and the aggregation of shared joints is only described in Appendix B. Define these quantities at first use in Section 3.2.2.","section":"Eq. (4), Appendix B"},{"comment":"The FPS figure reports LLM token-generation throughput via llama.cpp but not end-to-end latency including the motion encoder/decoder and tokenizer post-processing. A full pipeline latency breakdown would strengthen the real-time claim.","section":"§5.3.2"}],"recommendation":"major_revision","confidential_remarks":"The core concern is evaluational circularity: the part-level and unseen benchmarks are drawn from the same auto-annotation pipeline as the training data. The authors should be asked to release the evaluation sets, report annotation validation rates, and add at least one external human-validated part-control benchmark. The paper also does not disclose whether external test sets overlap with HuMo100M; this needs clarification before the benchmark claims can be fully trusted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. The genuinely new thing is PRQ: decomposing motion into five body parts with shared joints and a shared codebook, with part-level supervision from auto-generated descriptions. That goes beyond the upper/lower or body/hand splits in prior work, and the tokenizer results in Table 4 are solid. The design choices are well ablated (Tables 6-8) and the paper is unusually honest in Section 5.4.4: PRQ's FID on HumanML3D is worse than MoMask's, and they say so rather than burying it. HuMo263 also sidesteps the IK bottleneck, which is plausible and useful for real-time.\n\nThe real-time claim (20+ FPS) is credible given the frame-by-frame decoding and measured latency. If HuMo100M ships, it would be a major training resource.\n\nNow the soft spots. The biggest one is circularity. The I2M, I2PM, I2U, and I2LM benchmarks are sampled from HuMo100M, which is also the training distribution, and the part-level labels come from the same Gemini/PoseScript templates used during training. So the I2PM gains in Table 6 may reflect template matching rather than kinematic control. The left-right exchange test is only 50 sentences and stylistically close to the training templates. That doesn't kill the tokenizer contribution, but it means the headline 'controllability' claim is not yet established. I don't see any human validation of the generated part descriptions in the paper.\n\nAlso, no artifacts are released, so the 5M-sequence claim can't be checked. There's at least one provenance question: LMM and MotionGPT-v2 have identical numbers in Table 1, which needs a footnote. And 'state-of-the-art' in the abstract overreaches; Table 1 shows MoMask wins on FID, though the paper does later acknowledge this.\n\nNone of this is a load-bearing flaw in PRQ itself. The core empirical story is defensible. The paper is honest about its trade-offs and the ablations are real.\n\nWho should read it: anyone working on motion tokenization or VLMMs, especially if the dataset is released. It's a serious systems-and-data paper. I'd send it to peer review, conditional on the authors releasing artifacts and either moving the controllability benchmarks out-of-distribution or providing human-validated part-level evaluation. Without that, the controllability claim remains unverified, but the tokenizer and dataset contributions stand on their own enough to warrant referee time.","headline":"Genuinely new part-aware tokenizer with solid ablations and honest reporting; the controllability claims rest on in-distribution benchmarks that need outside validation before they land.","tokens_in":21371,"tokens_out":1745,"would_cite":true,"duration_ms":19566,"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":"Being-M0.5 is a 7-billion-parameter vision-language-motion model that generates controllable 3D human motion in real time from text, video, or pose inputs, trained on a new 100-million-instruction motion dataset.","keywords":["human motion generation","vision-language-motion model","real-time inference","part-level control","residual vector quantization","large-scale dataset","text-to-motion","instruction following"],"falsifier":"Take a random sample of the auto-generated part-level captions from the HuMo100M pipeline and have independent raters judge whether each caption matches the limb motion in the paired video; a high mismatch rate would undercut the I2PM gains and the 76% left-right exchange score, since those numbers depend on the same annotation machinery.","tokens_in":20199,"feed_emoji":"🕺","tokens_out":7393,"duration_ms":82174,"temperature":0.7,"pith_summary":"Being-M0.5 sets out to show that the bottleneck in human motion generation is controllability, not just quality: a practical model must follow arbitrary instructions, start from any pose, generate long sequences, handle unseen actions, and move individual limbs on command. The paper claims that all five can be solved at once with a 7-billion-parameter model trained on HuMo100M, a dataset of over five million web-extracted motion sequences and 100 million instruction instances with part-level annotations, long concatenated motions, and text-aligned video clips. On standard benchmarks it reports a HumanML3D FID of 0.056 and top-1 R-precision of 0.535, while running at 20–28 frames per second with 4-bit quantization. The enabling mechanism is part-aware residual quantization, which turns motion into per-limb discrete codes that can be decoded frame by frame. If the claims hold, motion generation moves from offline batch synthesis to interactive use in games, film, and robotics.","feed_headline":"7B model turns text and video into 3D motion at 28 fps","feed_subtitle":"Part-aware tokenization plus a 100-million-instruction dataset give limb-level control in real time.","key_machinery":"Part-aware residual quantization (PRQ) is the load-bearing mechanism. Standard residual quantization iteratively refines a whole-body latent vector through stacked codebooks; PRQ instead splits every pose into five anatomically meaningful parts before quantization, applies residual refinement independently to each part, and shares a single codebook and shared joints across parts. This does two jobs: it makes limb identity legible to the language model (a part token can be conditioned on a phrase like 'raise your left arm'), and it enables frame-by-frame decoding, because each frame's part codes are produced progressively rather than waiting for all quantization layers. A second mechanism is","core_discovery":"The paper's central claim is that fine-grained part-level control and real-time speed are compatible in one vision-language-motion model, and that both come from the tokenizer rather than from a bigger backbone. Being-M0.5 decomposes each body pose into five parts — left arm, right arm, left leg, right leg, torso — with shared joints across parts, and quantizes each part with four layers of residual vector quantization against a shared codebook of 1,024 entries. This part-aware residual quantization (PRQ) expands the number of representable configurations multiplicatively without extra storage, and its frame-by-frame decoding lets the language model stream motion codes instead of waiting for","pith_inferences":["The auto-annotation pipeline is the part most likely to be stress-tested: until a human validation rate is reported for the part-level captions, the magnitude of the I2PM gains should be treated with caution.","The five-way part decomposition is a design choice, not a natural kind; a finer or activity-dependent decomposition (e.g., splitting hands, or grouping by functional synergies) is a testable extension that could change the control granularity.","If HuMo100M is released as claimed, the dataset itself may end up mattering more than the model: it would enable third parties to test whether the controllability gains come from scale, from task diversity, or from the PRQ architecture, by controlling for each factor.","The left-right exchange benchmark (76%) is a promising cheap probe; extending it to a larger, more varied prompt set with intermediate part annotations would give a more reliable measure of part-level control."],"forward_implications":["If the reported results hold, a single 7B model can replace separate text-to-motion, pose-conditioned prediction, long-motion, and part-control systems with one real-time system.","A 100-million-instruction motion dataset of this size would give the community a common training ground for scaling experiments in motion generation, analogous to the role of large image-text datasets.","Part-level discrete codes with shared joints could make interactive motion editing possible, since individual limbs can be re-decoded without regenerating the whole body.","The frame-by-frame decoding result suggests that residual-quantization tokenizers, not just diffusion or autoregressive whole-body models, can meet real-time latency budgets."],"supporting_citations":[{"why":"The standard text-to-motion benchmark whose data and motion representation define the primary evaluation.","marker":"[3]"},{"why":"The residual-quantization masked motion model that serves as the central baseline; its RQ formulation is the starting point for PRQ.","marker":"[1]"},{"why":"Introduces residual quantization, the mechanism PRQ adapts to part-level codes.","marker":"[10]"},{"why":"The earlier million-scale motion dataset whose web-video pipeline and baselines HuMo100M scales up.","marker":"[9]"},{"why":"A large motion model whose multi-task evaluation protocols are reused for motion prediction and in-between benchmarks.","marker":"[21]"},{"why":"An LLM-based motion generator used as the baseline for the new instruction-following testbeds.","marker":"[13]"},{"why":"An autoregressive motion model with finite scalar quantization used as a comparison point.","marker":"[39]"},{"why":"The method that estimates 3D body motion from web video, the first stage of the data pipeline.","marker":"[41]"},{"why":"The refinement policy applied to estimated motions to improve quality.","marker":"[42]"},{"why":"The large multimodal model used to auto-generate part-level text descriptions.","marker":"[44]"}],"fun_headline_variants":["Part-aware tokens give limb-level control at real-time speed","Real-time motion generator with body-part precision","5M clips + 100M instructions: motion at 28 fps","Streaming codes enable real-time limb control","Being-M0.5: real-time, controllable vision-language-motion"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The part-control results rest on the accuracy of automatically generated limb-level text descriptions; if those captions often disagree with which limb actually moves, the reported gains would come from mapping templates to the training distribution rather than from genuine kinematic control.","fun_headline_variants_meta":{"raw":{"variants":["Part-aware tokens give limb-level control at real-time speed","Real-time motion generator with body-part precision","5M clips + 100M instructions: motion at 28 fps","Streaming codes enable real-time limb control","Being-M0.5: real-time, controllable vision-language-motion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000261,"raw_usage":{"total_tokens":1459,"prompt_tokens":803,"completion_tokens":656,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":575}},"tokens_in":547,"tokens_out":656,"duration_ms":7233,"temperature":1.0,"reasoning_tokens":575,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T21:49:36.220701+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random sample of the auto-generated part-level captions from the HuMo100M pipeline and have independent raters judge whether each caption matches the limb motion in the paired video; a high mismatch rate would undercut the I2PM gains and the 76% left-right exchange score, since those numbers depend on the same annotation machinery.","supporting_citations":[{"cited_title":"Scaling large motion models with million-level human motions","cited_arxiv_id":null,"evidence_quote":"The earlier million-scale motion dataset whose web-video pipeline and baselines HuMo100M scales up."},{"cited_title":"Large motion model for unified multi-modal motion generation","cited_arxiv_id":null,"evidence_quote":"A large motion model whose multi-task evaluation protocols are reused for motion prediction and in-between benchmarks."},{"cited_title":"Motiongpt: Finetuned llms are general-purpose motion generators","cited_arxiv_id":null,"evidence_quote":"An LLM-based motion generator used as the baseline for the new instruction-following testbeds."},{"cited_title":"ScaMo: Exploring the Scaling Law in Autoregressive Motion Generation Model","cited_arxiv_id":"2412.14559","evidence_quote":"An autoregressive motion model with finite scalar quantization used as a comparison point."}],"review_version":1}