Pith. sign in

REVIEW 2 major objections 5 minor 62 references

ARMS unifies solo and two-person human motion in one causal streaming model so people can approach, interact, and disengage from text without restarting generation.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A single causal diffusion model with an anchor–relational motion representation generates streaming solo and two-person motion and smooth solo–social transitions from incremental text.

T0 review reviewed 2026-07-11 challenge →

load-bearing objection Solid methods paper that actually unifies solo and two-person streaming motion with a clean asymmetric representation; transition gains look real, scope is honest. the 2 major comments →

arxiv 2607.05733 v1 pith:2D7AKPAN submitted 2026-07-07 cs.CV

ARMS: Anchor-Relational Motion Streaming for Seamless Solo-Social Motion Transitions

Huakun Liu , Qing Yu , Kent Fujiwara , Hideaki Uchiyama , Kiyoshi Kiyokawa This is my paper
classification cs.CV
keywords human motion generationhuman-human interactiontext-to-motionstreaming generationcausal diffusionautoregressive modelssolo-social transitionsanchor-relational representation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Realistic text-driven motion needs people who act alone, enter interactions, and later separate, all in one continuous stream. Most systems instead emit fixed-length clips under a fixed number of agents, so solo–social switches create jumps and long rollouts drift. ARMS claims a single causal process can do both regimes: a dynamics-asymmetric Anchor–Relational state keeps each person’s time evolution separate from partner alignment via a relative-translation term that is off for solo and for the Anchor, and on only for the Relational partner during interaction. A causal latent diffusion model then refines motion segment by segment from past context only, with mode-aware gates that turn cross-agent attention on or off. The result, on the paper’s evidence, is smoother transitions and coherent social geometry while remaining competitive on standard interaction benchmarks—including under true streaming inference.

Core claim

Solo motion and human–human interaction can be generated inside one causal generative process by a dynamics-asymmetric Anchor–Relational representation plus mode-aware relational gating: temporal evolution stays per-person and extensible, while inter-person geometry is controlled by a partner-referenced relative translation that is inactive for solo and for the Anchor and active only for the Relational stream during interaction.

What carries the argument

Dynamics-asymmetric Anchor–Relational representation: Anchor integrates global root motion with relative displacement fixed at zero; Relational stores partner-referenced Δ so its global pose is Anchor plus Δ; mode-aware relational gating then activates or masks cross-agent attention so the same causal diffusion transformer supports solo or interaction without dual-integration drift.

Load-bearing premise

That hard on/off social coupling, with the second agent started from noise or re-encoded relative dynamics inside the interaction distances seen in training, is enough for seamless joins and splits without soft blending, scene context, or contact physics.

What would settle it

Run many long solo→interaction→solo streams where agents must meet from far outside the training distance range or flip modes rapidly; if transition jerk and inter-person alignment/penetration stay near ground truth, the seamless-unification claim holds; large discontinuities or collapsed relative geometry falsify it.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Evolving text can drive continuous scenes that switch agent count without resetting the latent state.
  • Solo↔interaction and interaction↔interaction boundaries can reach jerk levels near ground truth rather than large discontinuities.
  • Streaming (incremental) generation can stay competitive with full-window interaction quality on standard benchmarks.
  • One shared causal autoencoder and model can be trained jointly on single-person and two-person data.
  • Long-horizon spatial consistency improves because only the Anchor branch integrates the global path.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Soft blending of relational attention across mode switches, which the authors flag as future work, would be a direct test of whether residual hard-gate abruptness can be removed without losing causality.
  • Pairwise relational branches plus higher-order structure could extend the same idea to small multi-agent groups once interaction-distance handling is explicit.
  • Practical pipelines can run independent solo streams until agents enter a plausible range, then merge under interaction mode—matching the paper’s own usage note for far starts.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. ARMS proposes a single causal generative process for text-driven streaming motion that unifies solo behavior and two-person interaction under changing agent configurations. The core technical idea is a dynamics-asymmetric Anchor–Relational representation: the Anchor agent uses canonical incremental dynamics with relational displacement fixed at zero (Eq. 5), while the Relational agent stores a partner-referenced planar relative translation so that its global root is reconstructed as P^(R)_t = P^(A)_t + Δ^(R)_t (Eq. 6). Motions are compressed by a shared causal temporal VAE; a causal relational diffusion transformer then refines segment-wise noised latents under mode-aware relational gating masks Γ (Eqs. 10–11) that enable or mask cross-agent attention. Inference uses staggered segment refinement, variable-length history conditioning on prompt changes, and explicit solo↔interaction activation of the Relational branch (Algorithms 1–3). Empirically, full-window and streaming variants are competitive or best on InterHuman and InterX interaction metrics (Tables 1–2), HumanML3D streaming results are strong (supp.), and constructed multi-segment transition evaluations show substantially lower jerk (PJ/AUJ) than adapted InterMask at solo–interaction and interaction–interaction boundaries (Table 3), with ablations on representation, segment size, latent dim, K, and δ (Table 4).

Significance. If the results hold, the paper addresses a practically important and previously underexplored gap: incremental text-driven generation that can enter and leave social interaction without resetting the generative state. The dynamics-asymmetric representation is a clean, reusable design that reconciles long-horizon causal rollout with stable inter-person geometry, and the mode-aware gating shows that one architecture can serve both solo and duo regimes. Strengths include joint training on HumanML3D and InterHuman under a shared representation, competitive standard interaction metrics under streaming inference, explicit transition metrics (jerk) against a nontrivial adapted baseline, ablations that isolate the representation and segment schedule, and released code. Limitations (hard binary gating, ≤2 agents, interaction-distance range of the training data, no scene/physics) are stated and appropriately scoped; they do not negate the contribution for the claimed setting.

major comments (2)
  1. §4.4 / Table 3: The central transition claim rests on custom multi-segment sequences and an adapted InterMask baseline (history-conditioned unmasked prefix; second agent masked then activated). The adaptation is necessary and the large PJ/AUJ gap is informative, but the protocol is not standardized. The manuscript should more fully specify sequence construction (how solo/interaction clips are chosen and concatenated, prompt timing, initialization of the second agent for the baseline), and ideally report at least one additional streaming interaction baseline (e.g., Interact2Ar or HINT under comparable history conditioning) so that the smoothness gain is not tied to a single adaptation.
  2. §3.4 Inference and Limitations: Relational-branch initialization at interaction onset is stated as “from noise unless otherwise stated,” with optional re-encoding of relative dynamics. Table 3’s near-GT solo↔interaction jerk is load-bearing for the seamless-transition claim; the paper should report (or ablate) noise vs. re-encoded initialization and any failure modes when agents start outside the InterHuman interaction-distance range, so readers can judge how much of the smoothness comes from the representation/gating versus favorable init and distance assumptions.
minor comments (5)
  1. Clarify invertibility and exact conversion between the 272-dim HumanML3D variant, InterHuman, and the Anchor–Relational state (Eq. 4) so that metric conversion “back to native representation” is fully reproducible.
  2. Fig. 2 and §3.4: Make the segment-wise noise schedule and staggered refinement offset δ more explicit in the main figure or caption; the inference description is clear in the supplement but dense in the main text.
  3. Table 1 streaming vs full-window: briefly discuss the R-Precision drop under streaming (compounding uncertainty) so readers do not over-read the full-window numbers as the primary streaming result.
  4. Limitations already note hard gating and multi-agent scaling; a short forward pointer in §3.4 to soft blending or pairwise multi-agent extensions would help set expectations without expanding scope.
  5. Minor polish: consistent hyphenation (solo–social / solo-social), and ensure all arXiv-only baselines (HINT, Interact2Ar) are cited with enough detail for readers to locate the concurrent work.

Circularity Check

0 steps flagged

No significant circularity: empirical streaming motion method validated on external datasets and independent metrics, not by construction from its own inputs.

full rationale

ARMS is a methods paper whose load-bearing claims are architectural design plus empirical comparison, not a first-principles derivation that could collapse into its definitions. The Anchor–Relational state (Eqs. 4–6) and mode gate Γ (Eqs. 10–11) are explicit modeling choices that make solo and interaction share one causal process; they do not define the reported transition or retrieval metrics. Training uses HumanML3D/InterHuman (and InterX for transfer); evaluation uses standard external protocols (R-Precision, FID, MM Dist, Diversity) and jerk-based transition measures against adapted baselines (InterMask, InterGen-style ablations). Dataset conversion into the shared representation is stated as deterministic and invertible, with metrics computed after conversion back to native formats—preprocessing, not a tautological rewrite of the training objective. Citations of prior causal diffusion / MotionStreamer-style components supply architectural scaffolding; the central solo–social streaming claim is tested by new experiments (Tables 1–4, qualitative streams), not by importing a uniqueness theorem or renaming a fitted constant as a prediction. No step reduces Eq. X to Eq. Y by construction or fits a parameter then re-labels a near-identical quantity as the main result.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

Central claims rest on standard diffusion/VAE practice plus paper-specific design choices (asymmetric dynamics, hard gating, segment noise schedule) and hyperparameters chosen for training/inference. No new physical entities; invented constructs are architectural. Free parameters are the usual ML knobs that affect reported metrics.

free parameters (5)
  • segment size S = 5
    Controls joint refinement window; default S=5; ablated at 1/3/10; affects FID and jerk.
  • latent dimension d = 64
    VAE latent width; default 64; ablated 32/128.
  • max context length L = 75
    Maximum latent context per agent; set to 75 (300 frames).
  • diffusion steps K and refinement offset δ = K=50, δ=5
    Inference schedule; default K=50, δ=5; ablated; trade quality vs speed/smoothness.
  • CFG scale and text-drop probability = 2.5 / 0.1
    Classifier-free guidance scale 2.5 at inference; 0.1 drop in training.
axioms (4)
  • domain assumption Causal temporal VAE compression with downsampling factor 4 yields a latent space suitable for autoregressive diffusion without destroying motion semantics.
    Adopted from MotionStreamer-style causal compression (§3.3); standard in recent streaming motion work.
  • ad hoc to paper Partner-referenced planar relative translation on the Relational branch plus zero relational term on the Anchor preserves inter-person geometry while allowing unbounded causal rollout.
    Core design of dynamics-asymmetric representation (§3.2); motivated by InterGen/DuetGen tradeoffs but not independently proven.
  • ad hoc to paper Hard mode-aware relational gating (Γ) is sufficient to switch between solo and interaction without soft blending for the evaluated transitions.
    Defined in Eqs. (10)–(11); Limitations note binary switching may still be abrupt.
  • domain assumption HumanML3D and InterHuman (and InterX after conversion) are adequate proxies for streaming solo–social behavior when sequences are concatenated for evaluation.
    Training and constructed streaming tests (§4.1, §4.4).
invented entities (2)
  • Dynamics-asymmetric Anchor–Relational motion representation no independent evidence
    purpose: Unify solo and interaction in one state space with partner-relative Δxz only on the Relational stream.
    Central representation invention (§3.2); independent_evidence false beyond empirical ablations.
  • Mode-aware relational gating masks in causal relational diffusion no independent evidence
    purpose: Activate/mask cross-agent attention for interaction vs solo within one denoiser.
    Enables single-model configuration switching (§3.4); validated by transition metrics, not external theory.

reviewed 2026-07-11 · how reviews work

0 comments
Cite this review

Pith. "Pith review of ARMS: Anchor-Relational Motion Streaming for Seamless Solo-Social Motion Transitions." pith.science (2026). https://pith.science/paper/2D7AKPAN

@misc{pith2026260705733,
  author       = {Pith},
  title        = {Pith review of: ARMS: Anchor-Relational Motion Streaming for Seamless Solo-Social Motion Transitions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2D7AKPAN}},
  note         = {Machine review of arXiv:2607.05733}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Generating temporally continuous and socially coherent human motion from text remains a fundamental challenge, particularly in realistic streams where people act alone, enter interactions, and later disengage. Most existing methods generate fixed-length motion clips under static agent configurations, which makes them brittle to solo-social transitions and unsuitable for incremental generation over long horizons. We propose ARMS, an Anchor-Relational Motion Streaming framework that unifies solo motion and human-human interaction within a single causal generative process. ARMS introduces a dynamics-asymmetric representation that decouples per-person temporal evolution from inter-person alignment via a partner-referenced relative-translation term, enabling seamless switching of social coupling without sacrificing long-horizon stability or spatial consistency between agents. On top of a causal latent space, a causal relational diffusion model progressively refines motion segment by segment using only past context, capturing both intra-person temporal dependencies and inter-person relations. Mode-aware relational gating activates or masks cross-agent connections, allowing the same model to support both solo and interaction generation. Experiments show that ARMS improves transition smoothness and social coherence compared to interaction-centric baselines, while also achieving competitive results on human-human interaction benchmarks.

Figures

Figures reproduced from arXiv: 2607.05733 by Hideaki Uchiyama, Huakun Liu, Kent Fujiwara, Kiyoshi Kiyokawa, Qing Yu.

Figure 1
Figure 1. Figure 1: ARMS generates long-horizon motion streams from evolving textual instruc￾tions, enabling seamless transitions between solo behavior and human–human interac￾tion within a single generative process. In this example, two individuals approach each other, engage in multiple interactions (e.g., greeting and fighting), and then disengage to continue with independent actions, without resetting or reinitializing ge… view at source ↗
Figure 2
Figure 2. Figure 2: We encode motion into causal latent streams for the Anchor and Rela￾tional agents, concatenate them, and apply symmetric segment-wise noising. A text￾conditioned Causal Relational Diffusion Transformer denoises the sequence using segment-wise causal attention and mode-aware relational gating masks, enabling cross￾agent visibility for interaction and masking it for solo motion. Per-frame state definition. W… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of long-duration interaction synthesis. Compared to InterMask, our method generates smoother temporal transitions, avoids abrupt motion discontinuities, and preserves coherent interaction dynamics across extended sequences. “A person walks forward” “One pivots and lifts their right leg to strike” “One pivots and lifts their right leg to strike” “They show respect to each other by bow… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative results of streaming motion generation. From left to right, we show: (1) solo motion transitioning into human–human interaction, (2) interaction evolving into a different interaction, and (3) interaction resolving back into solo motion. All sequences are generated in a streaming manner without resetting the latent state. 4.3 Motion Streaming: Qualitative Evaluation We present qualitative result… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative results of dynamic social motion streaming generation [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

62 extracted references · 62 canonical work pages · 4 internal anchors

  1. [1]

    Journal of Machine Learning Research26(209), 1–80 (2025) 10

    Albergo, M., Boffi, N.M., Vanden-Eijnden, E.: Stochastic interpolants: A unifying framework for flows and diffusions. Journal of Machine Learning Research26(209), 1–80 (2025) 10

    work page 2025

  2. [2]

    In: The Eleventh International Conference on Learning Representations (2023) 10

    Albergo, M.S., Vanden-Eijnden, E.: Building normalizing flows with stochastic in- terpolants. In: The Eleventh International Conference on Learning Representations (2023) 10

    work page 2023

  3. [3]

    ACM Transactions on Graphics (TOG)42(4), 1–20 (2023) 3

    Alexanderson,S.,Nagy,R.,Beskow,J.,Henter,G.E.:Listen,denoise,action!audio- driven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG)42(4), 1–20 (2023) 3

    work page 2023

  4. [4]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Barquero, G., Escalera, S., Palmero, C.: Seamless human motion composition with blended positional encodings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 457–469 (2024) 4, 13

    work page 2024

  5. [5]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Cai, Z., Jiang, J., Qing, Z., Guo, X., Zhang, M., Lin, Z., Mei, H., Wei, C., Wang, R., Yin, W., et al.: Digital life project: Autonomous 3d characters with social intelligence. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 582–592 (2024) 1, 2, 10

    work page 2024

  6. [6]

    In: The Thirteenth International Conference on Learning Representations (2025) 2

    Cen, Z., Pi, H., Peng, S., Shuai, Q., Shen, Y., Bao, H., Zhou, X., Hu, R.: Ready- to-react: Online reaction policy for two-character interaction generation. In: The Thirteenth International Conference on Learning Representations (2025) 2

    work page 2025

  7. [7]

    Advances in Neural Information Processing Systems37, 24081–24125 (2024) 7, 9, 23

    Chen, B., Martí Monsó, D., Du, Y., Simchowitz, M., Tedrake, R., Sitzmann, V.: Diffusion forcing: Next-token prediction meets full-sequence diffusion. Advances in Neural Information Processing Systems37, 24081–24125 (2024) 7, 9, 23

    work page 2024

  8. [8]

    In: The Thirteenth International Conference on Learning Representations (2025) 1, 2, 3

    Chen, J., Hu, P., Chang, X., Shi, Z., Kampffmeyer, M., Liang, X.: Sitcom-crafter: A plot-driven human motion generation system in 3d scenes. In: The Thirteenth International Conference on Learning Representations (2025) 1, 2, 3

    work page 2025

  9. [9]

    ACM Transactions on Graphics (TOG)40(4), 1–13 (2021) 3

    Chen, K., Tan, Z., Lei, J., Zhang, S.H., Guo, Y.C., Zhang, W., Hu, S.M.: Chore- omaster: choreography-oriented music-driven dance synthesis. ACM Transactions on Graphics (TOG)40(4), 1–13 (2021) 3

    work page 2021

  10. [10]

    In: ACM SIGGRAPH 2024 Conference Papers

    Chen, R., Shi, M., Huang, S., Tan, P., Komura, T., Chen, X.: Taming diffusion probabilistic models for character control. In: ACM SIGGRAPH 2024 Conference Papers. pp. 1–10 (2024) 4

    work page 2024

  11. [11]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 18000–18010 (2023) 27

    work page 2023

  12. [12]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Dabral, R., Mughal, M.H., Golyanik, V., Theobalt, C.: Mofusion: A framework for denoising-diffusion-based motion synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9760–9770 (2023) 3

    work page 2023

  13. [13]

    In: Proceedings of the Special Interest Group on Com- puterGraphicsandInteractiveTechniquesConferenceConferencePapers.pp.1–11 (2025) 3, 4, 5

    Ghosh, A., Zhou, B., Dabral, R., Wang, J., Golyanik, V., Theobalt, C., Slusallek, P., Guo, C.: Duetgen: Music driven two-person dance generation via hierarchi- cal masked modeling. In: Proceedings of the Special Interest Group on Com- puterGraphicsandInteractiveTechniquesConferenceConferencePapers.pp.1–11 (2025) 3, 4, 5

    work page 2025

  14. [14]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Guo, C., Mu, Y., Javed, M.G., Wang, S., Cheng, L.: Momask: Generative masked modeling of 3d human motions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1900–1910 (2024) 27

    work page 1900

  15. [15]

    In: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition

    Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition. pp. 5152–5161 (2022) 3, 4, 9, 10, 11 18 H. Liu et al

    work page 2022

  16. [16]

    In: European Conference on Computer Vision

    Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized model- ing for the reciprocal generation of 3d human motions and texts. In: European Conference on Computer Vision. pp. 580–597. Springer (2022) 3

    work page 2022

  17. [17]

    In: Proceedings of the 28th ACM international conference on multimedia

    Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: Proceedings of the 28th ACM international conference on multimedia. pp. 2021–2029 (2020) 3

    work page 2021

  18. [18]

    In: The Thirteenth International Conference on Learning Representations (2025) 3, 10, 11, 13, 28, 29

    Javed, M.G., Li, X., et al.: Intermask: 3d human interaction generation via collabo- rative masked modeling. In: The Thirteenth International Conference on Learning Representations (2025) 3, 10, 11, 13, 28, 29

    work page 2025

  19. [19]

    Advances in Neural Information Processing Systems36, 20067–20079 (2023) 3, 27

    Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems36, 20067–20079 (2023) 3, 27

    work page 2023

  20. [20]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13401–13412 (2021) 3

    work page 2021

  21. [21]

    International Journal of Computer Vision132(9), 3463–3483 (2024) 2, 3, 4, 5, 9, 10, 11, 14, 28, 30

    Liang,H.,Zhang,W.,Li,W.,Yu,J.,Xu,L.:Intergen:Diffusion-basedmulti-human motion generation under complex interactions. International Journal of Computer Vision132(9), 3463–3483 (2024) 2, 3, 4, 5, 9, 10, 11, 14, 28, 30

    work page 2024

  22. [22]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Lim, D., Bae, J., Hwang, I., Lee, S., Lee, H., Kim, Y.M.: Event-driven storytelling with multiple lifelike humans in a 3d scene. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11654–11664 (2025) 1, 2, 3

    work page 2025

  23. [23]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Liu, H., Zhu, Z., Becherini, G., Peng, Y., Su, M., Zhou, Y., Zhe, X., Iwamoto, N., Zheng, B., Black, M.J.: Emage: Towards unified holistic co-speech gesture generation via expressive masked audio gesture modeling. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1144–1154 (2024) 3

    work page 2024

  24. [24]

    arXiv preprint arXiv:2601.20383 (2026) 4, 10, 11, 15

    Liu, M., Di, Y., Wang, G., Qu, Y., Zhu, D., Li, Y., Ji, X.: Hint: Hierarchical inter- action modeling for autoregressive multi-human motion generation. arXiv preprint arXiv:2601.20383 (2026) 4, 10, 11, 15

    work page arXiv 2026

  25. [25]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.R.: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11461– 11471 (2023) 24

    work page 2023

  26. [26]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5442–5451 (2019) 9

    work page 2019

  27. [27]

    Absolute Coordinates Make Motion Generation Easy

    Meng, Z., Han, Z., Peng, X., Xie, Y., Jiang, H.: Absolute coordinates make motion generation easy. arXiv preprint arXiv:2505.19377 (2025) 9, 10

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  28. [28]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Mullen, J.F., Kothandaraman, D., Bera, A., Manocha, D.: Placing human anima- tions into 3d scenes by learning interaction-and geometry-driven keyframes. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 300–310 (2023) 3

    work page 2023

  29. [29]

    In: Proceedings of the IEEE/CVF International Conference on Com- puter Vision

    Ota, S., Yu, Q., Fujiwara, K., Ikehata, S., Sato, I.: Pino: Person-interaction noise optimization for long-duration and customizable motion generation of arbitrary- sized groups. In: Proceedings of the IEEE/CVF International Conference on Com- puter Vision. pp. 10676–10685 (2025) 15

    work page 2025

  30. [30]

    arXiv e-prints pp

    Ouyang, R., Li, H., Zhang, Z., Wang, X., Zhu, Z., Huang, G., Wang, X.: Motion-r1: Chain-of-thought reasoning and reinforcement learning for human motion genera- tion. arXiv e-prints pp. arXiv–2506 (2025) 3 ARMS: Anchor–Relational Motion Streaming 19

    work page 2025

  31. [31]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single im- age. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10975–10985 (2019) 10, 26

    work page 2019

  32. [32]

    Peebles,W.,Xie,S.:Scalablediffusionmodelswithtransformers.In:Proceedingsof the IEEE/CVF international conference on computer vision. pp. 4195–4205 (2023) 8

    work page 2023

  33. [33]

    In: Proceedings of the IEEE/CVF international con- ference on computer vision

    Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion syn- thesis with transformer vae. In: Proceedings of the IEEE/CVF international con- ference on computer vision. pp. 10985–10995 (2021) 3

    work page 2021

  34. [34]

    In: European conference on computer vision

    Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: European conference on computer vision. pp. 480–

    work page

  35. [35]

    In: ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecog- nition

    Petrovich, M., Litany, O., Iqbal, U., Black, M.J., Varol, G., Bin Peng, X., Rempe, D.: Multi-track timeline control for text-driven 3d human motion generation. In: ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecog- nition. pp. 1911–1921 (2024) 4

    work page 1911

  36. [36]

    In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion

    Ruiz-Ponce, P., Barquero, G., Palmero, C., Escalera, S., García-Rodríguez, J.: in2in: Leveraging individual information to generate human interactions. In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion. pp. 1941–1951 (2024) 3, 10

    work page 1941

  37. [37]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Ruiz-Ponce, P., Barquero, G., Palmero, C., Escalera, S., García-Rodríguez, J.: Mix- ermdm: Learnable composition of human motion diffusion models. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 12380–12390 (2025) 3

    work page 2025

  38. [38]

    arXiv preprint arXiv:2512.19692 (2025) 4, 11

    Ruiz-Ponce, P., Escalera, S., García-Rodríguez, J., Deng, J., Potamias, R.A.: Inter- act2ar: Full-body human-human interaction generation via autoregressive diffusion models. arXiv preprint arXiv:2512.19692 (2025) 4, 11

    work page arXiv 2025

  39. [39]

    Text-driven Motion Generation: Overview, Challenges and Directions

    Sahili, A.R., Neji, N., Tabia, H.: Text-driven motion generation: Overview, chal- lenges and directions. arXiv preprint arXiv:2505.09379 (2025) 1

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  40. [40]

    DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter

    Sanh, V., Debut, L., Chaumond, J., Wolf, T.: Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108 (2019) 9

    work page internal anchor Pith review Pith/arXiv arXiv 1910

  41. [41]

    In: The Twelfth International Conference on Learning Represen- tations (2024) 3, 4, 10, 11, 30

    Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. In: The Twelfth International Conference on Learning Represen- tations (2024) 3, 4, 10, 11, 30

    work page 2024

  42. [42]

    ACM Transactions on Graphics (TOG)43(4), 1–14 (2024) 4

    Shi, Y., Wang, J., Jiang, X., Lin, B., Dai, B., Peng, X.B.: Interactive character con- trol with auto-regressive motion diffusion models. ACM Transactions on Graphics (TOG)43(4), 1–14 (2024) 4

    work page 2024

  43. [43]

    Neurocomputing568, 127063 (2024) 8

    Su, J., Ahmed, M., Lu, Y., Pan, S., Bo, W., Liu, Y.: Roformer: Enhanced trans- former with rotary position embedding. Neurocomputing568, 127063 (2024) 8

    work page 2024

  44. [44]

    International Journal of Computer Vision134(3), 113 (2026) 1, 2

    Sui, K., Ghosh, A., Hwang, I., Zhou, B., Wang, J., Guo, C.: A survey on human interaction motion generation. International Journal of Computer Vision134(3), 113 (2026) 1, 2

    work page 2026

  45. [45]

    In: Proceedings of the IEEE/CVF international conference on computer vi- sion

    Tanaka, M., Fujiwara, K.: Role-aware interaction generation from textual descrip- tion. In: Proceedings of the IEEE/CVF international conference on computer vi- sion. pp. 15999–16009 (2023) 3, 10

    work page 2023

  46. [46]

    In: The Thirteenth International Conference on Learning Representations (2025) 3, 4 20 H

    Tevet, G., Raab, S., Cohan, S., Reda, D., Luo, Z., Peng, X.B., Bermano, A.H., van de Panne, M.: Closd: Closing the loop between simulation and diffusion for multi-task character control. In: The Thirteenth International Conference on Learning Representations (2025) 3, 4 20 H. Liu et al

    work page 2025

  47. [47]

    Human Motion Diffusion Model

    Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022) 3

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  48. [48]

    In: The Eleventh International Conference on Learning Representations (2023) 10, 11, 27

    Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. In: The Eleventh International Conference on Learning Representations (2023) 10, 11, 27

    work page 2023

  49. [49]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Wang, Y., Wang, S., Zhang, J., Fan, K., Wu, J., Xue, Z., Liu, Y.: Timotion: Tem- poral and interactive framework for efficient human-human motion generation. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 7169–7178 (2025) 3, 10, 11, 29

    work page 2025

  50. [50]

    arXiv preprint arXiv:2510.06504 (2025) 3, 10

    Wu, Q., Dou, Z., Guo, C., Huang, Y., Feng, Q., Zhou, B., Wang, J., Liu, L.: Text2interact: High-fidelity and diverse text-to-two-person interaction generation. arXiv preprint arXiv:2510.06504 (2025) 3, 10

    work page arXiv 2025

  51. [51]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Xiao, L., Lu, S., Pi, H., Fan, K., Pan, L., Zhou, Y., Feng, Z., Zhou, X., Peng, S., Wang, J.: Motionstreamer: Streaming motion generation via diffusion-based autoregressive model in causal latent space. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10086–10096 (2025) 4, 7, 9, 22, 27, 28, 29

    work page 2025

  52. [52]

    In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Xu, L., Lv, X., Yan, Y., Jin, X., Wu, S., Xu, C., Liu, Y., Zhou, Y., Rao, F., Sheng, X., et al.: Inter-x: Towards versatile human-human interaction analysis. In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 22260–22271 (2024) 10

    work page 2024

  53. [53]

    In: Thirteenth International Conference on 3D Vision (2025) 3, 4, 30

    Yu, H., Zhang, J., Chen, C., Xiang, T., Fang, Y., Niebles, J.C., Adeli, E.: Socialgen: Modeling multi-human social interaction with language models. In: Thirteenth International Conference on 3D Vision (2025) 3, 4, 30

    work page 2025

  54. [54]

    arXiv preprint arXiv:2602.22594 (2026) 8, 9, 23

    Yu, Q., Watanabe, A., Fujiwara, K.: Causal motion diffusion models for autore- gressive motion generation. arXiv preprint arXiv:2602.22594 (2026) 8, 9, 23

    work page arXiv 2026

  55. [55]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 27

    Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 27

    work page 2023

  56. [56]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Zhang, Y., Feng, Y., Cseke, A., Saini, N., Bajandas, N., Heron, N., Black, M.J.: Primal: Physically reactive and interactive motor model for avatar learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12725–12736 (2025) 4

    work page 2025

  57. [57]

    In: The Thirteenth International Conference on Learning Representations (2025) 4

    Zhao, K., Li, G., Tang, S.: Dartcontrol: A diffusion-based autoregressive motion model for real-time text-driven motion control. In: The Thirteenth International Conference on Learning Representations (2025) 4

    work page 2025

  58. [58]

    In: Proceedings of the IEEE/CVF international con- ference on computer vision

    Zhao, K., Zhang, Y., Wang, S., Beeler, T., Tang, S.: Synthesizing diverse human motions in 3d indoor scenes. In: Proceedings of the IEEE/CVF international con- ference on computer vision. pp. 14738–14749 (2023) 3

    work page 2023

  59. [59]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Zhong, C., Hu, L., Zhang, Z., Xia, S.: Attt2m: Text-driven human motion generation with multi-perspective attention mechanism. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 509–519 (2023) 27

    work page 2023

  60. [60]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation rep- resentations in neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5745–5753 (2019) 6

    work page 2019

  61. [61]

    arXiv preprint arXiv:2506.24086 (2025) 3 ARMS: Anchor–Relational Motion Streaming 21

    Zhu, B., Jiang, B., Wang, S., Tang, S., Chen, T., Luo, L., Zheng, Y., Chen, X.: Motiongpt3: Human motion as a second modality. arXiv preprint arXiv:2506.24086 (2025) 3 ARMS: Anchor–Relational Motion Streaming 21

    work page arXiv 2025

  62. [62]

    a person walks straight forward

    Zhuo, W., Ma, F., Fan, H.: Infinidreamer: Arbitrarily long human motion genera- tion via segment score distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14688–14698 (2025) 4 ARMS: Anchor–Relational Motion Streaming for Seamless Solo-Social Motion Transitions Supplementary Material Huakun Liu1 , Qing Yu2 , Kent ...

    work page 2025

This paper was first reviewed by grok-4.5 on July 11, 2026.