{"id":"93c364b3-7d47-46ab-b319-334c091207b8","arxiv_id":"2607.08741","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":6,"one_line_summary":"An autoregressive diffusion model with a hybrid explicit-root/latent-body representation generates real-time, controllable 3D human motion from text and spatial constraints.","lead":"ARDY is a real-time system that generates 3D human motion from streaming text prompts and kinematic constraints like waypoints or keyframes. A smart generalist might read it to understand how generative models can power interactive character animation in games and robotics.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"Constraint adherence validated primarily on ground-truth-sampled goals; robustness to arbitrary/adversarial constraints is under-tested, but not enough to overturn the central claim.","rationale":"The paper is well-executed: the hybrid representation, two-stage denoiser, and variable-context autoregressive formulation are clearly motivated, and the ablations in Tables 2–3 provide genuine evidence for each design choice. The public benchmark results (Tables 4–5) are competitive against strong baselines, and the interactive demo demonstrates practical capability. The reader's identified weakest assumption (two-stage vs. one-stage) is a reasonable observation but not truly load-bearing—the ablation in Table 2 does show a meaningful gap on constraint metrics (e.g., waypoint error 0.024 vs 0.164 m), and even if the gap were smaller, the overall system would still work. My concern is more substantive: the 'native control' claim is strongest exactly where the evaluation is weakest. The proprietary-dataset ablations use ground-truth-sampled constraints (always achievable), the perturbation test is mentioned but not quantified, and the public benchmark shows notably worse raw constraint error (4.15 cm vs ~2.5 cm). This doesn't invalidate the paper—ARDY still outperforms baselines without optimization—but it means the practical robustness of 'native' control under arbitrary user inputs remains an open question. The verdict should remain ACCEPT; this is a limitation worth noting, not a disqualifying flaw. The promised code release would allow independent verification.","tokens_in":24960,"tokens_out":3142,"duration_ms":178340,"concrete_test":"On the Bones Rigplay test set, evaluate constraint adherence under systematically degraded constraint plausibility: (1) perturb body-level constraints (joint positions, full-body keyframes) with Gaussian noise at 5/10/20 cm magnitude, and (2) specify physically implausible constraints (e.g., root trajectory passing through walls, end-effector targets requiring self-penetration). Report joint position/rotation errors and motion quality (FID, skating) for each perturbation level. If errors scale roughly linearly with perturbation magnitude and FID remains stable, the native control claim is robust. If errors spike non-linearly or FID degrades substantially at 10 cm perturbation, the model's constraint adherence is overfit to ground-truth-like inputs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that ARDY 'natively learns controllable generation' without test-time optimization or RL policies (Sec 1, Sec 3.3). The strongest quantitative support for constraint adherence comes from Tables 2–3 on the proprietary Bones Rigplay dataset, where spatial constraints are 'sampled directly from the ground-truth test set' (Sec 5.1). This means every constraint is guaranteed physically plausible and consistent with natural human motion. The only robustness test mentioned is adding 'slight random perturbations to the global translation and heading of a subset of sampled constraints' (Sec 5.1), but no quantitative results are reported for perturbed vs. unperturbed settings, and body-level constraints (joint positions/rotations, full-body keyframes) are never perturbed. On the public HumanML3D benchmark (Table 4), the raw (non-optimized) joint position error is 4.15 cm—substantially worse than the ~2.5 cm on Bones Rigplay and far from the 0.30 cm achieved with test-time optimization. This gap raises the question of whether the strong 'native' constraint results on Bones Rigplay reflect the architecture or the favorable constraint distribution. The claim is not circular or unsound—ARDY clearly outperforms MaskControl* (46.18 cm) and DiP (9.20–17.64 cm) without optimization—but the absolute quality of 'native' control is dataset-dependent and may degrade on constraints that deviate from natural motion distributions, which is exactly the use case for interactive applications where users specify arbitrary goals.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"The paper introduces ARDY, an autoregressive diffusion model for interactive 3D human motion generation that supports online text prompting and flexible, long-horizon kinematic constraints (root waypoints/trajectories, full-body keyframes, end-effector positions/rotations) without test-time optimization or RL policies. The core technical contributions are: (1) a hybrid motion representation combining explicit global root features with a learned latent body embedding, (2) a two-stage interleaved transformer denoiser that first predicts clean root motion and then predicts latent body motion conditioned on that root, and (3) a variable-length history context and out-of-window future constraint conditioning mechanism. The method is evaluated on the public HumanML3D benchmark against offline (MaskControl) and online (DiP) baselines, and on the proprietary Bones Rigplay dataset via ablation studies. An interactive demo demonstrates real-time control via text, mouse, and keyboard inputs at 33–63 ms latency.","tokens_in":25779,"tokens_out":1432,"duration_ms":261247,"significance":"The paper addresses a genuine and important gap in the motion generation literature: existing offline methods offer rich controllability but are too slow for interactive use, while existing online methods are fast but sacrifice either text conditioning, kinematic control, or long-horizon planning. The combination of capabilities in Table 1—real-time generation, online text prompting, diverse spatial control types, native control without optimization or RL, and long history/future context—is unmatched by prior work. The two-stage denoiser and hybrid representation are well-motivated and validated by ablation (Table 2). The release of code and models, along with the interactive demo, strengthens reproducibility and practical impact. The evaluation on HumanML3D against both offline and autoregressive baselines is appropriate and the results are competitive or superior. The work is a solid contribution to the interactive motion generation field.","major_comments":[{"comment":"The central claim that ARDY 'natively learns controllable generation' (Sec. 1, Sec. 3.3) is primarily validated on the proprietary Bones Rigplay dataset where spatial constraints are 'sampled directly from the ground-truth test set' (Sec. 5.1). This means every constraint is guaranteed to be physically plausible and consistent with natural human motion. The only robustness test mentioned—adding 'slight random perturbations to the global translation and heading of a subset of sampled constraints' (Sec. 5.1)—reports no quantitative results for perturbed vs. unperturbed settings, and body-level constraints (joint positions/rotations, full-body keyframes) are never perturbed. On the public HumanML3D benchmark (Table 4), the raw (non-optimized) joint position error is 4.15 cm, substantially worse than the ~2.5 cm on Bones Rigplay (Table 2). This gap raises the question of whether the strong '","section":"§5.1, Constraint Sampling and Table 2/Table 3 evaluation"},{"comment":"The two-stage architecture is presented as a key design decision, and the ablation in Table 2 shows improvements over the one-stage baseline. However, the one-stage baseline still achieves competitive foot skating (0.264 vs. 0.264 m/s text-only; 0.248 vs. 0.250 m/s constraints) and R-precision (65.84 vs. 65.47 text-only). The performance gap is concentrated in constraint accuracy (e.g., waypoint error 0.164 m vs. 0.024 m). The paper should more precisely characterize when the two-stage design matters most (e.g., for which constraint types) and whether the one-stage baseline could be improved with better constraint conditioning, to clarify whether the two-stage decomposition is fundamentally necessary or simply a useful heuristic that happens to help with certain constraint types.","section":"§3.4, Two-Stage Denoiser and Table 2 ablation"}],"minor_comments":[{"comment":"Table 2: The one-stage baseline achieves identical foot skating (0.264 m/s) and slightly higher R-precision (65.84 vs. 65.47) in text-only generation. The paper should acknowledge this more directly rather than focusing only on the improvements in constraint-conditioned metrics.","section":null},{"comment":"§3.5, Eq. (11): The loss combines four terms with equal weight (1:1:1:1). No justification is given for this weighting. A brief ablation or discussion of sensitivity to loss weights would help reproducibility.","section":null},{"comment":"§5.1: The Bones Rigplay dataset is proprietary and not publicly available. While the HumanML3D evaluation partially addresses this, the main ablation analysis (Tables 2–3) cannot be independently reproduced. The paper should note this limitation more explicitly.","section":null},{"comment":"§6.2, Table 4: The paper states that using the original HumanML3D evaluator models disadvantages ARDY on FID and R-precision. However, the R-precision gap between ARDY (0.729) and MaskControl* (0.760) is non-trivial. A brief discussion of why the retargeting difference affects R-precision specifically would strengthen the comparison.","section":null},{"comment":"Fig. 3: The architecture diagram is informative but dense. A clearer separation of the root transformer and body transformer data flows, perhaps with color-coded arrows, would improve readability for readers unfamiliar with the two-stage diffusion paradigm.","section":null},{"comment":"§4.1: The replan buffer mechanism is described as optional, with the 4-step model using no buffer and the 10-step model using a single buffer frame. It would be helpful to report the actual end-to-end latency experienced by the user (including buffer wait time) for both configurations, not just the model inference latency.","section":null},{"comment":"Table 3, Diffusion steps section: The 4-step model achieves FID 0.034 and the 10-step model achieves 0.027. The paper claims 'performance is still acceptable for most applications when going as low as four steps' (Sec. 3.5). Quantifying what 'acceptable' means in terms of perceptual quality would be helpful, especially given the interactive demo uses the 4-step model.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The paper is a strong contribution with a well-executed system. The main concern is the reliance on proprietary data for the core ablation analysis, which limits independent verification. The HumanML3D evaluation partially mitigates this but uses a different tokenizer (vanilla AE vs. FSQ) and different generation horizon settings, making it difficult to directly compare the ablation findings to the public benchmark results. The authors should be asked to at minimum report perturbed-constraint results on HumanML3D to demonstrate robustness of the 'native control' claim on public data. The interactive demo and code release are significant positives."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"ARDY is a real-time autoregressive motion diffusion model that handles online text prompts and flexible kinematic constraints—root waypoints, full-body keyframes, end-effector positions/rotations—without test-time optimization or RL policies. That combination is genuinely new and the system seems to work well enough for a deployable demo at 33ms latency with 4 diffusion steps. The two main ideas are a hybrid representation (explicit global root + latent body tokens from an FSQ tokenizer) and a two-stage denoiser that predicts clean root first, then conditions body prediction on it. Both are cleanly motivated and the ablations in Table 2 show real gains over explicit-only representations, global-root decoders, and a one-stage baseline. The out-of-horizon goal handling (conditioning on constraints beyond the generation window) is a practical strength that prior autoregressive work like DiP can't do. The HumanML3D comparison against DiP (Table 5) is convincing: ARDY maintains ~2.5–2.9 cm joint error on out-of-horizon goals where DiP degrades to 17.6 cm. The stress-test concern about constraints being sampled from ground truth is valid but somewhat overstated. The perturbation they mention is only on root translation/heading, and they don't quantify it separately, which is a gap. But the HumanML3D results use the same ground-truth-sampled protocol as MaskControl and DiP, so the comparison is fair. The real issue is that 4.15 cm raw error on HumanML3D vs 2.5 cm on Bones Rigplay suggests the strong native numbers are partly dataset-dependent. The proprietary Bones Rigplay dataset (700 hours) is the backbone of the ablation analysis and isn't publicly available, though code and models are promised. The one-stage ablation (Table 2) still gets competitive foot skating and R-precision, so the two-stage advantage is quantitative, not a qualitative wall. These are real but minor soft spots. This paper is for researchers in character animation, interactive motion generation, and possibly humanoid robotics. It deserves a serious referee. Accept for review.","headline":"ARDY: Autoregressive Diffusion with Hybrid Representation for Interactive Human Motion Generation","tokens_in":25755,"tokens_out":500,"would_cite":true,"duration_ms":102460,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Real-time motion generation obeys text and spatial goals simultaneously","keywords":[],"falsifier":"A monolithic one-stage denoiser that jointly predicts root and body from the same hybrid representation, trained and evaluated under identical conditions, would match or exceed the two-stage architecture on both motion quality (FID, foot skating) and constraint adherence (joint error, trajectory error). If such a baseline existed, the central architectural claim would collapse.","tokens_in":25276,"feed_emoji":"🏃","tokens_out":628,"duration_ms":230186,"temperature":0.7,"pith_summary":"The paper claims that interactive, real-time 3D human motion generation need not trade away controllability. By decomposing each pose into an explicit global root trajectory (which can be directly overwritten with user-specified waypoints or paths) and a compact latent body embedding (which a learned tokenizer compresses for efficient diffusion-based generation), the method achieves precise spatial control and high-fidelity body motion in a single streaming model. A two-stage denoising transformer first predicts the clean root, then predicts the latent body conditioned on that root, iterating within each diffusion step so root and body mutually constrain each other. The model accepts variable-length history (up to 8 seconds) and future constraint horizons (up to 10 seconds), enabling it to handle complex multi-step text prompts and long-horizon goals that prior autoregressive methods with short context windows cannot. Because constraints are sampled from ground-truth motion during training and injected as masked conditioning tokens, the model learns to follow them natively, without test-time optimization or reinforcement learning policies. The authors validate on both a public benchmark and a large-scale proprietary dataset, showing competitive motion quality and substantially lower constraint errors than prior offline and autoregressive baselines, including a 4-step diffusion variant achieving 33 ms generation latency.","feed_headline":"Real-time motion generation obeys text and spatial goals simultaneously","feed_subtitle":"A hybrid root-and-body representation lets one streaming diffusion model follow complex prompts and long-horizon kinematic constraints at 33","key_machinery":"Hybrid motion tokens (explicit global root + FSQ-quantized latent body), interleaved two-stage transformer denoiser (root-first, body-second), masked constraint injection with root overwriting, variable-length history and future-goal conditioning, latency-aware replanning buffer","core_discovery":"The central mechanism is the hybrid representation: explicit global root features concatenated with learned latent body tokens, processed by an interleaved two-stage diffusion denoiser. This decomposition lets the model overwrite root features directly for precise trajectory control while keeping the body representation compact enough for efficient few-step diffusion generation. Combined with variable-length history and out-of-window future constraint conditioning, this architecture natively supports online text prompting and flexible kinematic constraints in a single streaming framework, a combination no prior method achieved.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Streaming motion diffusion follows text prompts and kinematic constraints online","Hybrid root-latent representation enables real-time controllable 3D motion synthesis","Two-stage denoiser supports online text and long-horizon kinematic goals for motion","Explicit root and latent body features balance trajectory control with fast generation","Streaming framework unifies online text prompts with flexible spatial constraints"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The paper assumes that decomposing generation into a root-prediction stage followed by a body-prediction stage is necessary for jointly achieving controllability and motion fidelity. The ablation supporting this shows the one-stage baseline is competitive on some metrics, so the gap is quantitative rather than a fundamental qualitative barrier.","fun_headline_variants_meta":{"raw":{"variants":["Streaming motion diffusion follows text prompts and kinematic constraints online","Hybrid root-latent representation enables real-time controllable 3D motion synthesis","Two-stage denoiser supports online text and long-horizon kinematic goals for motion","Explicit root and latent body features balance trajectory control with fast generation","Streaming framework unifies online text prompts with flexible spatial constraints"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":718,"prompt_tokens":625,"completion_tokens":93,"prompt_tokens_details":null},"tokens_in":625,"tokens_out":93,"duration_ms":112034,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T01:57:59.129349+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"A monolithic one-stage denoiser that jointly predicts root and body from the same hybrid representation, trained and evaluated under identical conditions, would match or exceed the two-stage architecture on both motion quality (FID, foot skating) and constraint adherence (joint error, trajectory error). If such a baseline existed, the central architectural claim would collapse.","supporting_citations":[],"review_version":1}