pith. sign in

arxiv: 2506.05952 · v4 · submitted 2025-06-06 · 💻 cs.CV · cs.AI

MOGO: Residual Quantized Hierarchical Causal Transformer for High-Quality and Real-Time 3D Human Motion Generation

Dongjie Fu , Tengjiao Sun , Pengcheng Fang , Xiaohao Cai , Hansung Kim This is my paper

Pith reviewed 2026-05-19 10:30 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords text-to-motion generation3D human motionautoregressive transformerresidual vector quantizationreal-time motion synthesisstreaming generation
0
0 comments X

The pith

MOGO generates high-quality 3D human motions from text in a single forward pass for real-time use.

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

The paper presents MOGO as an autoregressive system that turns text descriptions into 3D human motion sequences. It uses a motion scale-adaptive residual vector quantization step to turn movements into compact layered tokens, then feeds those into a hierarchical causal transformer that produces all layers at once. The design targets both visual realism and speed, adding a text alignment step to keep the output faithful to the input words. Experiments on standard datasets show it matches or exceeds prior transformer methods in motion quality while cutting latency enough for streaming and immediate responses. If the approach holds, it could shift motion synthesis toward practical interactive applications where users expect instant, coherent results.

Core claim

MOGO is a one-pass autoregressive framework built from MoSA-VQ, which hierarchically discretizes motion sequences using learnable scaling for compact representations, and RQHC-Transformer, which produces multi-layer motion tokens in one forward pass, combined with text condition alignment to preserve semantic control, delivering competitive fidelity on HumanML3D, KIT-ML, and CMP while improving real-time performance and zero-shot generalization.

What carries the argument

The residual quantized hierarchical causal transformer (RQHC-Transformer) that generates all layers of motion tokens in a single forward pass to reduce latency.

If this is right

  • Motions can be generated and streamed continuously as tokens arrive rather than waiting for a full sequence.
  • The model supports zero-shot application to new text prompts outside the training distribution.
  • Inference runs fast enough for real-time responsiveness on standard hardware.
  • Overall generation quality remains at or above current state-of-the-art transformer baselines.

Where Pith is reading between the lines

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

  • The same one-pass layered token idea might transfer to other sequence generation tasks such as music or speech synthesis.
  • Interactive tools could let users type short action phrases and immediately see animated results without noticeable delay.
  • Longer or more complex motions may become feasible if the hierarchical structure scales without extra passes.

Load-bearing premise

The hierarchical single-pass token generation plus text alignment keeps motion natural and faithful without adding visible errors or losing detail.

What would settle it

Quantitative results showing lower motion quality scores or no reduction in inference time compared with existing transformer methods when tested on the same HumanML3D or KIT-ML splits.

Figures

Figures reproduced from arXiv: 2506.05952 by Dongjie Fu, Hansung Kim, Pengcheng Fang, Tengjiao Sun, Xiaohao Cai.

Figure 1
Figure 1. Figure 1: Overview of the proposed MOGO framework. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative comparison of generated motions produced by different models. Compared to prior methods Momask, [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Motion Extension during Inference [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Genration Results a) Semantic Diversity and Text-Condition Alignment.: Across the 4 × 4 grid layout, we showcase 16 distinct prompts covering a diverse set of actions, such as: • Simple physical reactions: “The person was pushed but did not fall.” • Complex multi-phase motions: “A man is walking forward then steps over a stair then continues walking forward.” • Locomotion with intention shifts: “walked for… view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of motion sequences generated with and without Text Condition Alignment (TCA). TCA improves temporal [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of attention patterns with and without the PnQ (Prompt and Quantization) condition [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
read the original abstract

Recent advances in transformer-based text-to-motion generation have led to impressive progress in synthesizing high-quality human motion. Nevertheless, jointly achieving high fidelity, streaming capability, real-time responsiveness, and scalability remains a fundamental challenge. In this paper, we propose MOGO (Motion Generation with One-pass), a novel autoregressive framework tailored for efficient and real-time 3D motion generation. MOGO comprises two key components: (1) MoSA-VQ, a motion scale-adaptive residual vector quantization module that hierarchically discretizes motion sequences with learnable scaling to produce compact yet expressive representations; and (2) RQHC-Transformer, a residual quantized hierarchical causal transformer that generates multi-layer motion tokens in a single forward pass, significantly reducing inference latency. To enhance semantic fidelity, we further introduce a text condition alignment mechanism that improves motion decoding under textual control. Extensive experiments on benchmark datasets including HumanML3D, KIT-ML, and CMP demonstrate that MOGO achieves competitive or superior generation quality compared to state-of-the-art transformer-based methods, while offering substantial improvements in real-time performance, streaming generation, and generalization under zero-shot settings.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes MOGO, an autoregressive text-to-motion framework with two main components: MoSA-VQ, a motion scale-adaptive residual vector quantization module that produces hierarchical discrete representations, and RQHC-Transformer, a residual quantized hierarchical causal transformer that generates all residual token layers in a single forward pass. A text condition alignment mechanism is added to improve semantic fidelity. The central claim is that MOGO matches or exceeds state-of-the-art transformer methods on HumanML3D, KIT-ML, and CMP in generation quality while delivering substantial gains in inference speed, streaming capability, and zero-shot generalization.

Significance. If the causality and fidelity claims hold under rigorous evaluation, the work would be significant for real-time and streaming applications in animation, VR, and robotics. The combination of residual hierarchical quantization with single-pass causal generation addresses a practical efficiency bottleneck in autoregressive motion models and could influence subsequent designs for efficient sequential generation.

major comments (2)
  1. [§4.2] §4.2 (RQHC-Transformer): The single-forward-pass generation of multi-layer residual tokens requires explicit confirmation that the causal mask is applied independently per residual level rather than only at the sequence level. If shared attention mixes information across quantization layers before the final decoder, lower-layer residuals can condition on higher-layer future tokens, violating the autoregressive assumption and risking long-horizon artifacts (e.g., foot-skating or velocity discontinuities) that FID/R-Precision on short clips may miss. A concrete diagnostic—such as per-layer attention visualization or long-sequence coherence metrics under zero-shot prompts—should be added.
  2. [§5] §5 (Experiments): The abstract and results claim competitive/superior quality plus real-time gains, yet the provided evaluation summary lacks per-metric tables, ablation on the hierarchical single-pass design versus sequential residual generation, and error analysis on long-horizon coherence. Without these, it is impossible to verify that the claimed improvements are not artifacts of short-clip metrics or dataset-specific tuning.
minor comments (2)
  1. [§3.1] Notation for residual layers and scaling factors in MoSA-VQ should be unified between text and equations to avoid ambiguity in the hierarchical discretization process.
  2. [Figure 3] Figure 3 (architecture diagram) would benefit from explicit arrows or masks indicating the causal constraints across residual levels.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. The comments have helped us strengthen the presentation of the causality properties in RQHC-Transformer and the experimental validation. We address each point below and have revised the manuscript accordingly.

read point-by-point responses

  1. Referee: [§4.2] §4.2 (RQHC-Transformer): The single-forward-pass generation of multi-layer residual tokens requires explicit confirmation that the causal mask is applied independently per residual level rather than only at the sequence level. If shared attention mixes information across quantization layers before the final decoder, lower-layer residuals can condition on higher-layer future tokens, violating the autoregressive assumption and risking long-horizon artifacts (e.g., foot-skating or velocity discontinuities) that FID/R-Precision on short clips may miss. A concrete diagnostic—such as per-layer attention visualization or long-sequence coherence metrics under zero-shot prompts—should be added.

    Authors: We appreciate the referee’s emphasis on rigorously verifying the autoregressive property across residual layers. In RQHC-Transformer the input sequence is formed by interleaving tokens from all residual levels at each time step, and a block-diagonal causal mask is applied so that each residual level attends only to its own past tokens and to lower-level tokens from the same or earlier time steps. Higher-level tokens at the current time step are masked from lower-level predictions within the same forward pass. This design prevents any future-token leakage across layers while still enabling the hierarchical conditioning that makes single-pass generation possible. We have expanded Section 4.2 with a precise description of the per-level masking, an accompanying diagram, and per-layer attention maps in the supplementary material. We have also added zero-shot long-sequence coherence metrics (velocity discontinuity and foot-skating rates) to the experimental results. revision: yes

  2. Referee: [§5] §5 (Experiments): The abstract and results claim competitive/superior quality plus real-time gains, yet the provided evaluation summary lacks per-metric tables, ablation on the hierarchical single-pass design versus sequential residual generation, and error analysis on long-horizon coherence. Without these, it is impossible to verify that the claimed improvements are not artifacts of short-clip metrics or dataset-specific tuning.

    Authors: We agree that more granular reporting is necessary to substantiate the claims. The original manuscript contained aggregate metrics and a limited ablation; however, it did not include exhaustive per-metric tables, a direct head-to-head comparison against sequential residual generation, or quantitative long-horizon coherence analysis. In the revised version we have (i) expanded Table 1 to report all individual metrics (FID, R-Precision, MM-Dist, etc.) across the three datasets, (ii) added a dedicated ablation subsection comparing the single-pass hierarchical architecture against a sequential residual baseline under identical training budgets, and (iii) introduced long-horizon coherence metrics together with qualitative examples on extended zero-shot prompts. These additions confirm that the reported gains are not confined to short clips. revision: yes

Circularity Check

0 steps flagged

No significant circularity; claims rest on experimental benchmarks rather than self-referential definitions or fitted predictions.

full rationale

The paper presents MOGO as a novel autoregressive framework with two components—MoSA-VQ for scale-adaptive residual vector quantization and RQHC-Transformer for single-pass hierarchical token generation—plus a text alignment mechanism. No equations, derivations, or parameter-fitting steps are described in the provided text that would reduce any claimed prediction or result to its own inputs by construction. The quality and efficiency claims are positioned as outcomes of extensive experiments on HumanML3D, KIT-ML, and CMP datasets, which constitute external validation rather than tautological re-labeling of fitted quantities. No load-bearing self-citations, uniqueness theorems, or ansatzes imported from prior author work are invoked in the abstract or summary to justify core architectural choices. The derivation chain is therefore self-contained against the reported benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented physical entities; the new modules are algorithmic components rather than postulated entities with independent evidence.

pith-pipeline@v0.9.0 · 5752 in / 1094 out tokens · 41827 ms · 2026-05-19T10:30:59.997350+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

44 extracted references · 44 canonical work pages · 3 internal anchors

  1. [1]

    Human motion generation: A survey,

    W. Zhu, X. Ma, D. Ro, H. Ci, J. Zhang, J. Shi, F. Gao, Q. Tian, and Y . Wang, “Human motion generation: A survey,”IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 46, no. 4, pp. 2430– 2449, 2023. I, II

    work page 2023

  2. [2]

    Momask: Generative masked modeling of 3d human motions,

    C. Guo, Y . Mu, M. G. Javed, S. Wang, and L. Cheng, “Momask: Generative masked modeling of 3d human motions,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 1900–1910. I, II, III-B, I

    work page 2024

  3. [3]

    Generating human motion from textual descriptions with discrete representations,

    J. Zhang, Y . Zhang, X. Cun, Y . Zhang, H. Zhao, H. Lu, X. Shen, and Y . Shan, “Generating human motion from textual descriptions with discrete representations,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2023, pp. 14 730–14 740. I, II, I

    work page 2023

  4. [4]

    Mmm: Generative masked motion model,

    E. Pinyoanuntapong, P. Wang, M. Lee, and C. Chen, “Mmm: Generative masked motion model,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 1546–1555. I, II, I

    work page 2024

  5. [5]

    Attt2m: Text-driven hu- man motion generation with multi-perspective attention mechanism,

    C. Zhong, L. Hu, Z. Zhang, and S. Xia, “Attt2m: Text-driven hu- man motion generation with multi-perspective attention mechanism,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 509–519. I, II, III-B, I

    work page 2023

  6. [6]

    Human Motion Diffusion Model

    G. Tevet, S. Raab, B. Gordon, Y . Shafir, D. Cohen-Or, and A. H. Bermano, “Human motion diffusion model,” arXiv preprint arXiv:2209.14916, 2022. I, II

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  7. [7]

    High-Resolution Image Synthesis with Latent Diffusion Models

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR). New Orleans, LA, USA: IEEE, June 2022, pp. 10 684– 10 695, arXiv:2112.10752. I

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  8. [8]

    Executing your commands via motion diffusion in latent space,

    X. Chen, B. Jiang, W. Liu, Z. Huang, B. Fu, T. Chen, and G. Yu, “Executing your commands via motion diffusion in latent space,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 18 000–18 010. I, II

    work page 2023

  9. [9]

    Language mod- els are few-shot learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell et al., “Language mod- els are few-shot learners,” Advances in neural information processing systems, vol. 33, pp. 1877–1901, 2020. I

    work page 1901

  10. [10]

    Exploring the limits of transfer learning with a unified text-to-text transformer,

    C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y . Zhou, W. Li, and P. J. Liu, “Exploring the limits of transfer learning with a unified text-to-text transformer,” Journal of machine learning research, vol. 21, no. 140, pp. 1–67, 2020. I

    work page 2020

  11. [11]

    Language2pose: Natural language grounded pose forecasting,

    C. Ahuja and L.-P. Morency, “Language2pose: Natural language grounded pose forecasting,” in 2019 International conference on 3D vision (3DV). IEEE, 2019, pp. 719–728. II JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 8

    work page 2019

  12. [12]

    Synthesis of compositional animations from textual descriptions,

    A. Ghosh, N. Cheema, C. Oguz, C. Theobalt, and P. Slusallek, “Synthesis of compositional animations from textual descriptions,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 1396–1406. II

    work page 2021

  13. [13]

    Deep video generation, prediction and completion of human action sequences,

    H. Cai, C. Bai, Y .-W. Tai, and C.-K. Tang, “Deep video generation, prediction and completion of human action sequences,” in Proceedings of the European conference on computer vision (ECCV) , 2018, pp. 366–

    work page 2018

  14. [14]

    Learning diverse stochastic human-action generators by learning smooth latent transitions,

    Z. Wang, P. Yu, Y . Zhao, R. Zhang, Y . Zhou, J. Yuan, and C. Chen, “Learning diverse stochastic human-action generators by learning smooth latent transitions,” in Proceedings of the AAAI conference on artificial intelligence, vol. 34, no. 07, 2020, pp. 12 281–12 288. II

    work page 2020

  15. [15]

    Action2video: Generating videos of human 3d actions,

    C. Guo, X. Zuo, S. Wang, X. Liu, S. Zou, M. Gong, and L. Cheng, “Action2video: Generating videos of human 3d actions,” International Journal of Computer Vision , vol. 130, no. 2, pp. 285–315, 2022. II

    work page 2022

  16. [16]

    Action-conditioned 3d human motion synthesis with transformer vae,

    M. Petrovich, M. J. Black, and G. Varol, “Action-conditioned 3d human motion synthesis with transformer vae,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 10 985–10 995. II

    work page 2021

  17. [17]

    Generat- ing diverse and natural 3d human motions from text,

    C. Guo, S. Zou, X. Zuo, S. Wang, W. Ji, X. Li, and L. Cheng, “Generat- ing diverse and natural 3d human motions from text,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 5152–5161. II, IV-A1, IV-A1, IV-A2, VII

    work page 2022

  18. [18]

    Flame: Free-form language-based motion synthesis & editing,

    J. Kim, J. Kim, and S. Choi, “Flame: Free-form language-based motion synthesis & editing,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 7, 2023, pp. 8255–8263. II

    work page 2023

  19. [19]

    Motiondiffuse: Text-driven human motion generation with diffusion model,

    M. Zhang, Z. Cai, L. Pan, F. Hong, X. Guo, L. Yang, and Z. Liu, “Motiondiffuse: Text-driven human motion generation with diffusion model,” IEEE transactions on pattern analysis and machine intelligence, vol. 46, no. 6, pp. 4115–4128, 2024. II, I

    work page 2024

  20. [20]

    Remodiffuse: Retrieval-augmented motion diffusion model,

    M. Zhang, X. Guo, L. Pan, Z. Cai, F. Hong, H. Li, L. Yang, and Z. Liu, “Remodiffuse: Retrieval-augmented motion diffusion model,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 364–373. II

    work page 2023

  21. [21]

    Priority-centric human motion generation in discrete latent space,

    H. Kong, K. Gong, D. Lian, M. B. Mi, and X. Wang, “Priority-centric human motion generation in discrete latent space,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 14 806–14 816. II

    work page 2023

  22. [22]

    Motiongpt: Human motion as a foreign language,

    B. Jiang, X. Chen, W. Liu, J. Yu, G. Yu, and T. Chen, “Motiongpt: Human motion as a foreign language,” Advances in Neural Information Processing Systems, vol. 36, pp. 20 067–20 079, 2023. II, I

    work page 2023

  23. [23]

    Hierarchical transformers are more efficient language models,

    P. Nawrot, S. Tworkowski, M. Tyrolski, Ł. Kaiser, Y . Wu, C. Szegedy, and H. Michalewski, “Hierarchical transformers are more efficient language models,” in Findings of the Association for Computational Linguistics: NAACL 2022 , 2022, pp. 1559–1571. II

    work page 2022

  24. [24]

    Hier- archical transformers for long document classification,

    R. Pappagari, P. ˙Zelasko, J. Villalba, Y . Carmiel, and N. Dehak, “Hier- archical transformers for long document classification,” in 2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU) . Singapore: IEEE, December 2019, pp. 838–844, arXiv:1910.10781. II

    work page arXiv 2019

  25. [25]

    Cogview2: Faster and better text-to-image generation via hierarchical transformers,

    M. Ding, W. Zheng, W. Hong, and J. Tang, “Cogview2: Faster and better text-to-image generation via hierarchical transformers,” Advances in Neural Information Processing Systems , vol. 35, pp. 16 890–16 902,

    work page

  26. [26]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 10 012–10 022. II

    work page 2021

  27. [27]

    Scaling vision transformers to gigapixel images via hierarchical self-supervised learning,

    R. J. Chen, C. Chen, Y . Li, T. Y . Chen, A. D. Trister, R. G. Krishnan, and F. Mahmood, “Scaling vision transformers to gigapixel images via hierarchical self-supervised learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 16 144–16 155. II

    work page 2022

  28. [28]

    Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts,

    C. Guo, X. Zuo, S. Wang, and L. Cheng, “Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts,” in European Conference on Computer Vision . Springer, 2022, pp. 580–597. II

    work page 2022

  29. [29]

    Fg-t2m: Fine- grained text-driven human motion generation via diffusion model,

    Y . Wang, Z. Leng, F. W. Li, S.-C. Wu, and X. Liang, “Fg-t2m: Fine- grained text-driven human motion generation via diffusion model,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 22 035–22 044. I

    work page 2023

  30. [30]

    Motion anything: Any to motion generation,

    Z. Zhang, Y . Wang, W. Mao, D. Li, R. Zhao, B. Wu, Z. Song, B. Zhuang, I. Reid, and R. Hartley, “Motion anything: Any to motion generation,” arXiv preprint arXiv:2503.06955 , 2025. I

    work page arXiv 2025

  31. [31]

    The kit motion-language dataset,

    M. Plappert, C. Mandery, and T. Asfour, “The kit motion-language dataset,” Big data, vol. 4, no. 4, pp. 236–252, 2016. IV-A1

    work page 2016

  32. [32]

    Animationgpt: An aigc tool for generating game combat motion assets,

    Y . Liao, Y . Fu, Z. Cheng, and J. Wang, “Animationgpt: An aigc tool for generating game combat motion assets,” https://github.com/fyyakaxyy/ AnimationGPT, 2024. IV-A1

    work page 2024

  33. [33]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, b. ichter, F. Xia, E. Chi, Q. V . Le, and D. Zhou, “Chain-of-thought prompting elicits reasoning in large language models,” in Advances in Neural Information Processing Systems , S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, Eds., vol. 35. Curran Associates, Inc., 2022, pp. 24 824–24 837. [On-...

    work page 2022

  34. [34]

    Large lan- guage models are zero-shot reasoners,

    T. Kojima, S. S. Gu, M. Reid, Y . Matsuo, and Y . Iwasawa, “Large lan- guage models are zero-shot reasoners,” Advances in neural information processing systems, vol. 35, pp. 22 199–22 213, 2022. A0b

    work page 2022

  35. [35]

    Self-Consistency Improves Chain of Thought Reasoning in Language Models

    X. Wang, J. Wei, D. Schuurmans, Q. Le, E. Chi, S. Narang, A. Chowdh- ery, and D. Zhou, “Self-consistency improves chain of thought reasoning in language models,” arXiv preprint arXiv:2203.11171 , 2022. A0c JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 9 APPENDIX A. Qualitative Results of Infinite Motion Generation. A people makes two squats. A...

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  36. [36]

    Lift both arms as if holding a golf club

    work page

  37. [37]

    Swing the arms forward in an arc

    work page

  38. [38]

    A person lifts both arms, swings them forward in an arc, and returns to a standing position

    Return to a relaxed standing posture • Final rewritten instruction: “A person lifts both arms, swings them forward in an arc, and returns to a standing position. ” b) Example 2. (Wild data Test): • Raw prompt: “A human is practicing expressive spinning gestures. ” • Normalized: “Spin, pause, raise arms. ” • Steps:

    work page

  39. [39]

    Spin once on the spot

    work page

  40. [40]

    Hold still for a beat JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 13

    work page 2021

  41. [41]

    A person spins once on the spot, holds still briefly, and raises both hands

    Raise both hands • Final rewritten instruction: “A person spins once on the spot, holds still briefly, and raises both hands. ” c) Example 3. (Wild data Test): • Raw prompt: “A human is rehearsing a dynamic forward motion sequence. ” • Normalized: “Walk, crouch, roll forward, stand. ” • Steps:

    work page

  42. [42]

    A person walks forward, drops into a crouch, rolls ahead, and stands upright

    Stand upright • Final rewritten instruction: “A person walks forward, drops into a crouch, rolls ahead, and stands upright. ” As shown in Figure 5, these cases show how TCA strategy reliably converts noisy, multi-clause requests into concise, executable motion scripts—enabling the decoder to render complex behaviours with accurate temporal structure witho...

    work page 2021

  43. [43]

    However, this design introduces a level of indirection that leads to redundant attention computation and diluted gradient flow

    Without PnQ: An additional layer token is inserted between the textual prompt and the quantized motion sequence, serving as an intermediate fusion node. However, this design introduces a level of indirection that leads to redundant attention computation and diluted gradient flow. As a result, the model exhibits weaker cross-modal alignment, as the semanti...

    work page

  44. [44]

    With PnQ: We instead fuse the prompt token directly with the quantization condition tokens through positional and conditional injection. This eliminates the need for an intermediate token and allows the model to establish a more direct and semantically grounded connection between the text prompt and motion representation. Consequently, this design not onl...

    work page 2021