Pith. sign in

REVIEW 2 major objections 5 minor 3 cited by

A motion LLM that critiques its own drafts lands closer to the prompt.

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 ↗

T0 review · deepseek-v4-flash

2026-08-03 17:02 UTC pith:2VMVLI4M

load-bearing objection Interleaved generation, assessment, and refinement is a real, well-ablated step forward for motion LLMs; the main caveat is that the assessment bridge is trained from text captions rather than from the motion itself. the 2 major comments →

arxiv 2512.10730 v2 pith:2VMVLI4M submitted 2025-12-11 cs.CV

IRG-MotionLLM: Interleaving Motion Generation, Assessment and Refinement for Text-to-Motion Generation

Yuan-Ming Li , Qize Yang , Nan Lei , Shenghao Fu , Ling-An Zeng , Jian-Fang Hu , Xihan Wei , Wei-Shi Zheng This is my paper
classification cs.CV
keywords text-to-motion generationmotion-aware large language modelsinterleaved reasoningself-assessmentrefinementreinforcement learningGRPOHumanML3D
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.

This paper tries to establish that a single motion-aware language model can generate better text-aligned 3D human motion if it is trained to interleave generation with self-assessment and refinement, rather than treating understanding and generation as separate tasks. The authors propose IRMoGen, a paradigm in which the model plans, generates a motion, evaluates how well it matches the goal text, writes refinement instructions, and then regenerates, looping until it judges the motion satisfactory. They build IRG-MotionLLM with a three-stage training scheme and an automated data engine that synthesizes interleaved reasoning annotations from existing motion-text datasets. The reported experiments show consistent gains in text-motion alignment across datasets and evaluators, and suggest that reinforcement learning further encourages longer, more corrective reasoning loops.

Core claim

The central claim is that motion assessment and refinement tasks act as bridges that let understanding knowledge flow into generation. When the model is trained on these 'improving' tasks alongside basic motion captioning and generation, and then explicitly trained to interleave generation, assessment, and refinement in a single reasoning trajectory, the final generated motion aligns better with the goal text than the initial generation and than motions from baseline models. The paper also reports that this interleaved ability emerges even at Stage-1 before explicit chain-of-thought training, and that GRPO-based reinforcement learning pushes the model to perform more rounds of refinement, ea

What carries the argument

IRMoGen, the Interleaved Reasoning for Motion Generation paradigm, is the central object: a structured loop in which the model alternates between producing motion tokens, writing an alignment assessment and refinement instruction, and generating a revised motion, guided by a chain-of-thought template with plan tokens. The three-stage training scheme (task initialization, CoT learning with an 'Ignore Incorrect' loss mask for intermediate motions, and GRPO reinforcement with text-motion and motion-motion alignment rewards) carries the argument, supported by a data engine that samples negative motions at varying alignment levels and uses an LLM to synthesize evaluation and refinement annotation

Load-bearing premise

The model's assessment and refinement supervision comes from an LLM judging text descriptions of motions, not from actual motion content, so the model may learn to critique captions rather than the real generated motion.

What would settle it

Construct a test set where the text description of a generated motion is intentionally misleading (e.g., caption says 'walking forward' but the actual motion walks backward), and check whether IRG-MotionLLM's assessment flags the true motion error; if it consistently passes misleading captions, the text-based annotation proxy has failed.

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

If this is right

  • If correct, unified motion LLMs can self-correct through internal text-motion dialogue, removing the need for external modules or human intervention during inference.
  • Training on assessment and refinement tasks also improves the model's motion-to-text captioning, suggesting bidirectional knowledge flow between understanding and generation.
  • Reinforcement learning that rewards text-motion alignment lengthens the reasoning trajectory, and more rounds of refinement monotonically improve alignment.
  • The trained model can be repurposed as a text-motion reward model, boosting a separate motion generator's performance under an RLAIF framework.
  • The gains hold across two datasets (HumanML3D, KIT-ML) and two evaluators, indicating the effect is not tied to one benchmark's quirks.

Where Pith is reading between the lines

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

  • Because the synthetic assessment and refinement annotations are produced from text captions of negative motions rather than from the motions themselves, the model's critique skill is only as faithful as the LLM that wrote those captions; an editor might test whether the loop still helps when assessments are grounded in actual motion features.
  • The interleaved self-correction loop resembles recent 'thinking with generated images' ideas, suggesting the paradigm may transfer to other continuous generative modalities such as audio, video, or robotic trajectories where a model can evaluate and refine its own output in a closed loop.
  • The paper's focus is text-motion alignment; an obvious extension is to add physical plausibility or smoothness assessment into the loop, which could fix unnatural intermediate poses rather than only semantic mismatches.
  • Since longer reasoning trajectories emerge only after reinforcement learning, one can hypothesize that scaling up the base LLM and training data (larger motion datasets, more diverse captions) would further amplify the benefit of interleaving.

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. The paper proposes IRG-MotionLLM, a motion-aware LLM that interleaves text-to-motion generation with assessment and refinement in an iterative reasoning loop (IRMoGen). Training is done in three stages: (1) supervised fine-tuning on eight atomic tasks, including assessment and refinement; (2) supervised learning on IRMoGen-CoT trajectories generated by an automated GPT-4o-based data engine; (3) GRPO reinforcement learning with format, text-motion alignment, and motion-motion alignment rewards. Experiments on HumanML3D and KIT-ML report consistent gains in R-Precision and MM-Dist over the base model and prior unified motion LLMs, supported by per-stage ablations, cross-evaluator validation, and robustness tests.

Significance. If the findings hold, the paper offers a practical way to make motion-understanding tasks improve motion generation within a single LLM, a timely and non-obvious contribution. The evidence is unusually extensive: stage-wise ablations (Table 1), initial-vs-final generation comparisons (Table 3), cross-evaluator results (Table 5), perturbation robustness (Table 9), and a transfer experiment using IRG-MotionLLM as a reward model (Table 10). The release of code and data is a further strength. The main reservations are the unvalidated text-only annotation proxy used to train assessment/refinement and the overlap between the Stage-3 reward and the benchmark evaluator; both are load-bearing for the central claim and need additional experiments, but they are fixable within the scope of a revision.

major comments (2)
  1. [Sec. 4, Fig. 9, Sec. 10.1] The data engine constructs assessment and refinement supervision from text only. GPT-4o receives the goal text and the 'Text Description for Generated Motions' — the caption paired with the negative motion — not motion tokens, coordinates, or any motion-derived representation. The prompt tells the model to 'assume that you are watching the motion,' but the actual input is a caption. No validation is reported that the GPT-4o evaluations and refinement instructions are faithful to the true motion content; Sec. 10.1 filters only invalid responses. Because these labels train the improving tasks in Stage-1 and the full IRMoGen-CoT trajectories in Stage-2, the central claim that assessment/refinement bridges motion understanding and generation rests on an unvalidated proxy. If the captions are not reliable descriptions of the negative motions, the model may learn to critique/edit captions rath
  2. [Sec. 3.3, Eq. (2); Sec. 11.2] The Stage-3 text-motion alignment reward in Eq. (2) uses the paired text and motion encoders from the HumanML3D evaluator [13] — the same encoders used to compute R-Precision and MM-Dist in the benchmark (Sec. 11.2). The Stage-3 gains in Tables 1, 3, and 4 may therefore partly reflect direct optimization of the evaluation metrics. The cross-evaluator results in Table 5 and the discussion in Sec. 12.4 mitigate this concern, but MARDM [49] is still a text-motion embedding-distance metric very similar in form to the reward, so the issue is reduced but not eliminated. A cleaner test would be to train Stage-3 with a held-out evaluator (e.g., MARDM) as the reward, or to ablate the r_tm term, and then compare on the original benchmark. This would separate 'unleashing reasoning capacity' from 'tuning the evaluator.'
minor comments (5)
  1. [Figs. 1-2] The text contains non-English placeholder strings '无法显示该图片。' interspersed in the figure captions. These are presumably rendering artifacts and should be removed before submission.
  2. [Table 4] The base model row 'MotionLLM*' reports Top-1 R-Precision 0.496, while the published MotionLLM row reports 0.515. The paper should explain this discrepancy (evaluation protocol, checkpoint differences) so readers can assess the improvement from the training scheme rather than from a weaker base checkpoint.
  3. [Sec. 3.2] There is a typo/formatting error: 'we note{m 1, .., mn−1}asintermediate motions' should read 'we denote {m_1, ..., m_{n-1}} as intermediate motions.'
  4. [Sec. 10.1 / Sec. 9.2] The hyperparameters p, N_neg, N_max, G, and beta_kl are set without sensitivity analysis. Given that several are non-standard (e.g., p=90%), at least one ablation or a brief justification would help.
  5. [App. Eq. (4)] The notation in Eq. (4) is inconsistent: 'πbold' should be 'π_{θ_old}' and the ratio/objective would benefit from the standard clipping notation to avoid ambiguity.

Circularity Check

1 steps flagged

Stage-3 GRPO reward duplicates the HumanML3D evaluation metric, so part of the reported Stage-3 alignment gain is metric-tuning; Stage-1/2 and cross-evaluator results provide independent grounding.

specific steps
  1. fitted input called prediction [Sec. 3.3 (Eq. 2), Sec. 11.2 (Eq. 6), Sec. 12.4]
    "rtm(t,m_final,i) = −∥E_t(t)−E_m(m_final,i)∥2, (2) ... E_t and E_m are chosen from paired text and motion encoders in [13]. ... Multimodal Distance (MM-Dist) ... MM-Dist= 1/N Σ ||f_pred,i − f_text,i|| ... However, as we use the official evaluator [13] as the reward model, an extreme question may be raised: Does GRPO-based RL tuning in Stage-3 unleash the potential capacity or simply hack the evaluator?"

    The Stage-3 Text-Motion Alignment Reward in Eq. 2 is the negative of the sample-level MM-Dist computed from the same paired [13] text/motion encoders used by the official HumanML3D evaluator, and R-Precision ranks motions by those same distances. Maximizing this reward during GRPO therefore directly optimizes the reported MM-Dist/R-Precision metrics, so the Stage-3 improvement on the HumanML3D benchmark is at least partly enforced by construction rather than by an independent alignment signal. The paper itself acknowledges this 'hack the evaluator' concern and mitigates it with the MARDM [49] cross-evaluator and the non-RL Stage-1/2 results, which is why the circularity is partial rather than total.

full rationale

The derivation chain is largely self-contained: the IRMoGen paradigm, the three-stage training scheme, and the data engine are defined from external text-motion datasets and pre-trained encoders, and no load-bearing self-citation chain is used (the MotionLLM base is an external work; the authors' own related citation is not load-bearing). Stage-1 and Stage-2 gains come from supervised tasks and CoT trajectories whose intermediate motions are sampled by a separate TMR-based [54] ranking, not from the evaluation metric, and the final-motion targets are ground-truth motions. The main identifiable circular step is in Stage-3: the RL reward (Eq. 2) is the negative of the same [13] encoder distance that defines HumanML3D MM-Dist (Eq. 6) and R-Precision, so the main-benchmark Stage-3 alignment gain is partly metric-optimization. The paper explicitly raises this concern and supplies an independent MARDM [49] evaluator plus non-RL stage improvements, preventing the entire central claim from collapsing into the metric. The GPT-4o caption-based annotation pipeline is a fidelity/correctness risk (annotations are generated from text descriptions, not motion tokens), but it is not a circular reduction by construction and is therefore not scored as circularity here.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

No new physical entities, forces, dimensions, or conserved quantities are introduced. IRMoGen is a training paradigm, not an invented entity. The free parameters listed are the design choices that shape the data engine, inference loop, and RL stage; the most consequential unverified assumption is that GPT-4o text-based annotations faithfully teach motion-grounded assessment.

free parameters (4)
  • Maximum refinement rounds N_max = 4
    Inference stopping constraint (Sec. 9.2); affects performance as shown in Fig. 11, chosen by hand rather than fitted to data.
  • Negative sampling fraction p = 90% HumanML3D, 87.5% KIT-ML
    Data-engine rank cutoff for sampling negative motion-text pairs (Sec. 10.1); chosen by hand and affects difficulty of assessment/refinement training.
  • Number of negative samples N_neg = 4
    Number of intermediate motions per trajectory (Sec. 10.1); chosen by hand; determines Stage-2 trajectory length and training signal.
  • GRPO group size G and KL weight beta_kl = G=4, beta_kl=0.01
    RL hyperparameters in Table 6; standard values, not tuned against test set.
axioms (6)
  • domain assumption Motion VQ-VAE tokens preserve enough detail for assessment and refinement.
    Sec. 3.1 and Sec. 13; the authors admit the vanilla VQ-VAE may lose fine-grained motion details. If tokenization drops fine action cues, the assessment/refinement loop cannot recover them.
  • domain assumption HumanML3D/KIT-ML paired captions accurately describe their motions.
    The data engine and all metrics rely on caption-motion correspondence (Sec. 4, Sec. 11). Noisy captions would corrupt both training annotations and evaluation.
  • ad hoc to paper GPT-4o-generated goal analyses, evaluations, and refinement instructions are sufficiently faithful for training.
    Sec. 4 and Figs. 8-9; the pipeline filters invalid responses but does not verify annotation correctness against actual motion content.
  • ad hoc to paper Embedding-distance ranking with a pre-trained motion encoder yields meaningful progressive alignment levels.
    Sec. 4; negative motions are sampled from the top-p fraction of farthest embeddings and sorted as improving. The ordering is not human-verified.
  • domain assumption Official HumanML3D text/motion encoders are valid proxies for text-motion alignment, and the MARDM evaluator is an independent check.
    Eqs. (2)-(3) and Sec. 12.4. If both the primary and cross evaluators share biases, Stage-3 gains could be metric-specific rather than perceptual.
  • standard math GRPO is a valid optimization procedure for this tokenized motion LLM.
    Sec. 8.4; standard algorithm from [62], no new mathematical machinery claimed.

pith-pipeline@v1.3.0-alltime-deepseek · 36073 in / 14361 out tokens · 143618 ms · 2026-08-03T17:02:27.393894+00:00 · methodology

0 comments
read the original abstract

Recent advances in motion-aware large language models have shown remarkable promise for jointly learning motion understanding and generation knowledge. However, these models typically treat understanding and generation separately, limiting the mutual benefits that could arise from interactive feedback between tasks. In this work, we reveal that motion assessment and refinement tasks can act as crucial bridges to enable knowledge flow from motion understanding to generation. Specifically, we propose Interleaved Reasoning for Motion Generation (IRMoGen), a novel paradigm that tightly couples motion generation with assessment and refinement through iterative text-motion dialogue. To realize this, we introduce IRG-MotionLLM, the first model that seamlessly interleaves motion generation, assessment, and refinement to improve the alignment between generated motion and goal text. IRG-MotionLLM is developed progressively with a novel three-stage training scheme, initializing and subsequently enhancing native IRMoGen capabilities. To facilitate this development, we construct an automated data engine to synthesize interleaved reasoning annotations from existing text-motion datasets. Extensive experiments demonstrate the properties brought by IRMoGen training, and the advanced cross-benchmark and cross-evaluator performance of IRG-MotionLLM. Code and models are available at https://github.com/HumanMLLM/IRG-MotionLLM.

Figures

Figures reproduced from arXiv: 2512.10730 by Jian-Fang Hu, Ling-An Zeng, Nan Lei, Qize Yang, Shenghao Fu, Wei-Shi Zheng, Xihan Wei, Yuan-Ming Li.

Figure 1
Figure 1. Figure 1: Illustration of our proposed IRG-MotionLLM.Given a goal text and reasoning instruction as input, our IRG-MotionLLM is able to perform Text-Motion Interleaved Reasoning until a satisfactory motion is generated. The reasoning process includes multiple moves, i.e., an initial analysis on the goal, followed by multiple rounds of motion Generation, Assessment, and Refinement. The model can also adaptively make … view at source ↗
Figure 2
Figure 2. Figure 2: A three-stage training scheme is proposed to build our IRG-MotionLLM. In the first stage (Upper), we endow the model with meta IRMoGen abilities via eight related tasks. In the second stage (Lower-Left), we define an IRMoGen-CoT template and train a model to explicitly couple motion understanding and generation knowledge and perform native interleaved reasoning. In the third stage (Lower-Right), we employ … view at source ↗
Figure 3
Figure 3. Figure 3: The IRMoGen-CoT Template. IRMoGen-CoT Learning (Stage-2): Enabled by the data engine introduced in Sec. 4, we organize dynamic reason￾ing trajectories to train Fs1 to perform native interleaved reasoning, and obtain our Stage-2 model Fs2 (shown in the lower-left part of [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Negative text-motion pairs selection and annotation [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization results.Native interleaved reasoning enables IRG-MotionLLM to: (1) recognize the misalignment between pre￾viously generated motion and the goal text and provide proper refinement instructions; (2) refine the motion based on previous reasoning. Such characteristics help our method more accurately follow the details of the goal text than existing methods. The human figures are colored from ligh… view at source ↗
Figure 7
Figure 7. Figure 7: The inference pipeline for performing IRMoGen on our Stage-1 model. Our Stage-1 model can perform interleaved reasoning across generation, assessment and refinement tasks by sequentially using prompts of different tasks together with the outputs from previous steps. To determine when to stop, extra scripts or manual interventions are needed to check the refinement instruction. D.M.G: Direct Motion Generati… view at source ↗
Figure 8
Figure 8. Figure 8: The prompts for generating annotations of Goal Analysis. pointed up and hands on the chest to mimic wings, combined with a pecking motion involving the head and upper torso bending fo [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: The prompts for generating annotations of Text-Motion Alignment Assessment. [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The task instructions and response templates. The fields between “<” and “>” indicate the placeholders. IRG-MotionLLM Stage-1 Stage-2 Stage-3 General Hyper-perameters Batch Size 128 32 32 # GPUs (Nvidia L20s) 32 32 32 Learning Rate 1e-5 1e-5 1e-6 Epochs / Steps (H3D) 10 epochs 10 epochs 900 steps Epochs / Steps (KIT) 4 epochs 4 epochs 300 steps GRPO Hyper-perameters KL weight (βkl) - - 0.01 Group size (G)… view at source ↗
Figure 11
Figure 11. Figure 11: Ablation studies on the maximum rounds of refine￾ment on the HumanML3D dataset. Allowing model to perform more rounds of refinement brings improvement on the alignment between generated results and the goal text. 12.4. Stage-3 Training: Unleashing the capacity or Hacking the evaluator? In our Stage-3 training stage, we conduct GRPO-based RL tuning to further enhance the generation performance of our metho… view at source ↗
Figure 12
Figure 12. Figure 12: Inference pipeline for robustness evaluation. Given the reasoning instruction and a goal text, we first obtain the native reasoning trajectory of our IRG-MotionLLM. After that, we replace the first generated motion in the native reasoning trajectory to a random sampled perturbation motion, then feed the perturbed initial trajectory (containing only the goal analysis and perturbation motion) together with … view at source ↗
Figure 13
Figure 13. Figure 13: Visualization of the IRMoGen-CoT data for Stage-2 training [PITH_FULL_IMAGE:figures/full_fig_p023_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: More native interleaved reasoning trajectories of IRG-MotionLLM. Our method performs multi-round interleaved rea￾soning to capture the mis-alignment of the previous generated motion and finally generate a satisfactory motion. The planning steps are removed. Zoom in for best view [PITH_FULL_IMAGE:figures/full_fig_p024_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative results on robustness evaluation. Our IRG-MotionLLM can also perform accurate motion assessment on ran￾domly sampled perturbation motions, and refine the motion until it has reached the goal. Initial Motion First-Round Refined Motion Second-Round Refined Motion Third-Round Refined Motion Goal Text: a man crouches, stands back up, scratches his head, and crouches again. Bend down and Sit on the… view at source ↗
Figure 16
Figure 16. Figure 16: Our IRG-MotionLLM may fail on highly complex goal text. Although our IRG-MotionLLM produces a three-round refined motion which is much more aligned with the goal than previous generated motions, it still misses the action of “scratching his head”. However, we note that generating perfect motion to such complex goal text is also challenging for existing methods [15, 49, 73] [PITH_FULL_IMAGE:figures/full_f… view at source ↗

discussion (0)

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

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. MoGeFlow: Flowing Through Motion Codebook Geometry for Text-to-Motion Generation

    cs.GR 2026-06 unverdicted novelty 6.0

    MoGeFlow learns text-conditioned flows over PartVQ group-specific code embeddings to generate motions, achieving SOTA R-Precision on HumanML3D and KIT-ML while preserving discrete token validity.

  2. PhysiGen: Integrating Collision-Aware Physical Constraints for High-Fidelity Human-Human Interaction Generation

    cs.CV 2026-05 unverdicted novelty 6.0

    PhysiGen reduces interpenetration in text-driven 3D human interaction generation by simplifying meshes to geometric primitives for fast collision detection and guiding optimization with collision regions.

  3. MotionHiFlow: Text-to-motion via hierarchical flow matching

    cs.CV 2026-04 unverdicted novelty 6.0

    MotionHiFlow generates text-aligned 3D human motions using hierarchical flow matching across temporal scales, cross-scale transitions, a Text-Motion Diffusion Transformer, and a topology-aware Motion VAE, achieving st...

Reference graph

Works this paper leans on

109 extracted references · 35 linked inside Pith · cited by 3 Pith papers

  1. [1]

    interleaving reasoning: Next-generation reasoning systems for agi

    Wenxuan huang. interleaving reasoning: Next-generation reasoning systems for agi. github repository, 2025. 3

    2025

  2. [2]

    introducing openai o3 and o4-mini

    Openai. introducing openai o3 and o4-mini. openai blog, 2025b. 3

  3. [3]

    Being-m0

    Bin Cao, Sipeng Zheng, Ye Wang, Lujie Xia, Qianshan Wei, Qin Jin, Jing Liu, and Zongqing Lu. Being-m0. 5: A real-time controllable vision-language-motion model.arXiv preprint arXiv:2508.07863, 2025. 3, 6, 7

    Pith/arXiv arXiv 2025

  4. [4]

    Executing your commands via motion diffusion in latent space

    Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18000–18010, 2023. 3, 8, 6

    2023

  5. [5]

    Thinking with gen- erated images.arXiv preprint arXiv:2505.22525, 2025

    Ethan Chern, Zhulin Hu, Steffi Chern, Siqi Kou, Jiadi Su, Yan Ma, Zhijie Deng, and Pengfei Liu. Thinking with gen- erated images.arXiv preprint arXiv:2505.22525, 2025. 1, 3

    Pith/arXiv arXiv 2025

  6. [6]

    Emerging properties in unified multimodal pretraining.arXiv preprint arXiv:2505.14683, 2025

    Chaorui Deng, Deyao Zhu, Kunchang Li, Chenhui Gou, Feng Li, Zeyu Wang, Shu Zhong, Weihao Yu, Xiaonan Nie, Ziang Song, et al. Emerging properties in unified multimodal pretraining.arXiv preprint arXiv:2505.14683, 2025. 3

    Pith/arXiv arXiv 2025

  7. [7]

    Go to zero: Towards zero-shot motion generation with million-scale data

    Ke Fan, Shunlin Lu, Minyue Dai, Runyi Yu, Lixing Xiao, Zhiyang Dou, Junting Dong, Lizhuang Ma, and Jingbo Wang. Go to zero: Towards zero-shot motion generation with million-scale data. InProceedings of the IEEE/CVF In- ternational Conference on Computer Vision, pages 13336– 13348, 2025. 3, 9

    2025

  8. [8]

    Got: Unleashing reasoning capability of multimodal large language model for visual generation and editing.arXiv preprint arXiv:2503.10639, 2025

    Rongyao Fang, Chengqi Duan, Kun Wang, Linjiang Huang, Hao Li, Shilin Yan, Hao Tian, Xingyu Zeng, Rui Zhao, Jifeng Dai, et al. Got: Unleashing reasoning capability of multimodal large language model for visual generation and editing.arXiv preprint arXiv:2503.10639, 2025. 3

    Pith/arXiv arXiv 2025

  9. [9]

    Love-r1: Advancing long video understanding with an adaptive zoom-in mechanism via multi-step reasoning.arXiv preprint arXiv:2509.24786,

    Shenghao Fu, Qize Yang, Yuan-Ming Li, Xihan Wei, Xi- aohua Xie, and Wei-Shi Zheng. Love-r1: Advancing long video understanding with an adaptive zoom-in mechanism via multi-step reasoning.arXiv preprint arXiv:2509.24786,

    arXiv

  10. [10]

    Reasoning robustness of llms to adversar- ial typographical errors

    Esther Gan, Yiran Zhao, Liying Cheng, Mao Yancan, Anirudh Goyal, Kenji Kawaguchi, Min-Yen Kan, and Michael Shieh. Reasoning robustness of llms to adversar- ial typographical errors. InProceedings of the 2024 Confer- ence on Empirical Methods in Natural Language Processing, pages 10449–10459, 2024. 7

    2024

  11. [11]

    Thinkmorph: Emergent properties in multimodal interleaved chain-of-thought reasoning.arXiv preprint arXiv:2510.27492, 2025

    Jiawei Gu, Yunzhuo Hao, Huichen Will Wang, Linjie Li, Michael Qizhe Shieh, Yejin Choi, Ranjay Krishna, and Yu Cheng. Thinkmorph: Emergent properties in multimodal interleaved chain-of-thought reasoning.arXiv preprint arXiv:2510.27492, 2025. 3

    arXiv 2025

  12. [12]

    Ac- tion2motion: Conditioned generation of 3d human motions

    Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Ac- tion2motion: Conditioned generation of 3d human motions. InProceedings of the 28th ACM International Conference on Multimedia, pages 2021–2029, 2020. 3

    2021

  13. [13]

    Generating diverse and natural 3d human motions from text

    Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5152–5161, 2022. 1, 2, 5, 6, 3, 7, 9

    2022

  14. [14]

    Tm2t: Stochastic and tokenized modeling for the reciprocal genera- tion of 3d human motions and texts

    Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal genera- tion of 3d human motions and texts. InEuropean Conference on Computer Vision, pages 580–597. Springer, 2022. 7

    2022

  15. [15]

    Momask: Generative masked model- ing of 3d human motions

    Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked model- ing of 3d human motions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900–1910, 2024. 1, 3, 8, 6, 9, 12

    1900

  16. [16]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025. 2, 3, 5

    Pith/arXiv arXiv 2025

  17. [17]

    Atom: Aligning text-to-motion model at event-level with gpt-4vision reward

    Haonan Han, Xiangzuo Wu, Huan Liao, Zunnan Xu, Zhongyuan Hu, Ronghui Li, Yachao Zhang, and Xiu Li. Atom: Aligning text-to-motion model at event-level with gpt-4vision reward. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 22746–22755,

  18. [18]

    Semanticboost: Elevating motion gen- eration with augmented textual cues.arXiv preprint arXiv:2310.20323, 2023

    Xin He, Shaoli Huang, Xiaohang Zhan, Chao Weng, and Ying Shan. Semanticboost: Elevating motion gen- eration with augmented textual cues.arXiv preprint arXiv:2310.20323, 2023. 3

    Pith/arXiv arXiv 2023

  19. [19]

    Egolm: Multi-modal language model of egocentric motions

    Fangzhou Hong, Vladimir Guzov, Hyo Jin Kim, Yuting Ye, Richard Newcombe, Ziwei Liu, and Lingni Ma. Egolm: Multi-modal language model of egocentric motions. InPro- ceedings of the Computer Vision and Pattern Recognition Conference, pages 5344–5354, 2025. 3

    2025

  20. [20]

    Motionverse: A unified multimodal framework for motion comprehension, generation and edit- ing.arXiv preprint arXiv:2509.23635, 2025

    Ruibing Hou, Mingshuang Luo, Hongyu Pan, Hong Chang, and Shiguang Shan. Motionverse: A unified multimodal framework for motion comprehension, generation and edit- ing.arXiv preprint arXiv:2509.23635, 2025. 3

    arXiv 2025

  21. [21]

    Language is not all you need: Aligning perception with language mod- els.Advances in Neural Information Processing Systems, 36:72096–72109, 2023

    Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Barun Patra, et al. Language is not all you need: Aligning perception with language mod- els.Advances in Neural Information Processing Systems, 36:72096–72109, 2023. 1

    2023

  22. [22]

    Interleaving rea- soning for better text-to-image generation.arXiv preprint arXiv:2509.06945, 2025

    Wenxuan Huang, Shuang Chen, Zheyong Xie, Shaosheng Cao, Shixiang Tang, Yufan Shen, Qingyu Yin, Wenbo Hu, Xiaoman Wang, Yuntian Tang, et al. Interleaving rea- soning for better text-to-image generation.arXiv preprint arXiv:2509.06945, 2025. 1, 3

    Pith/arXiv arXiv 2025

  23. [23]

    Vision-r1: Incentivizing reasoning capability in multimodal large language models.arXiv preprint arXiv:2503.06749,

    Wenxuan Huang, Bohan Jia, Zijie Zhai, Shaosheng Cao, Zheyu Ye, Fei Zhao, Zhe Xu, Yao Hu, and Shaohui Lin. Vision-r1: Incentivizing reasoning capability in multimodal large language models.arXiv preprint arXiv:2503.06749,

    Pith/arXiv arXiv

  24. [24]

    Gpt-4o system card

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perel- man, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli- hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 5, 9

    Pith/arXiv arXiv 2024

  25. [25]

    Multimodal generative ai with autoregressive llms for human motion understanding and generation: A way for- ward.arXiv preprint arXiv:2506.03191, 2025

    Muhammad Islam, Tao Huang, Euijoon Ahn, and Usman Naseem. Multimodal generative ai with autoregressive llms for human motion understanding and generation: A way for- ward.arXiv preprint arXiv:2506.03191, 2025. 1

    Pith/arXiv arXiv 2025

  26. [26]

    Openai o1 system card.arXiv preprint arXiv:2412.16720, 2024

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richard- son, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card.arXiv preprint arXiv:2412.16720, 2024. 3

    Pith/arXiv arXiv 2024

  27. [27]

    Motiongpt: Human motion as a foreign lan- guage.Advances in Neural Information Processing Systems, 36:20067–20079, 2023

    Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign lan- guage.Advances in Neural Information Processing Systems, 36:20067–20079, 2023. 1, 3, 6, 7, 8, 4, 9

    2023

  28. [28]

    Motionchain: Conversational motion controllers via multimodal prompts

    Biao Jiang, Xin Chen, Chi Zhang, Fukun Yin, Zhuoyuan Li, Gang Yu, and Jiayuan Fan. Motionchain: Conversational motion controllers via multimodal prompts. InEuropean Conference on Computer Vision, pages 54–74. Springer,

  29. [29]

    Causal motion tokenizer for streaming motion generation

    Biao Jiang, Xin Chen, Ailing Zeng, Xinru Sun, Fukun Yin, Xianfang Zeng, Xuanyang Zhang, Gang Yu, and Tao Chen. Causal motion tokenizer for streaming motion generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2024–2034, 2025. 3

    2024

  30. [30]

    T2i-r1: Reinforcing image generation with col- laborative semantic-level and token-level cot.arXiv preprint arXiv:2505.00703, 2025

    Dongzhi Jiang, Ziyu Guo, Renrui Zhang, Zhuofan Zong, Hao Li, Le Zhuo, Shilin Yan, Pheng-Ann Heng, and Hong- sheng Li. T2i-r1: Reinforcing image generation with col- laborative semantic-level and token-level cot.arXiv preprint arXiv:2505.00703, 2025. 3, 2

    Pith/arXiv arXiv 2025

  31. [31]

    Unitoken: Harmonizing multimodal understanding and generation through unified vi- sual encoding

    Yang Jiao, Haibo Qiu, Zequn Jie, Shaoxiang Chen, Jingjing Chen, Lin Ma, and Yu-Gang Jiang. Unitoken: Harmonizing multimodal understanding and generation through unified vi- sual encoding. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 3600–3610, 2025. 7

    2025

  32. [32]

    Motion generation: A survey of generative approaches and benchmarks.arXiv preprint arXiv:2507.05419, 2025

    Aliasghar Khani, Arianna Rampini, Bruno Roy, Larasika Nadela, Noa Kaplan, Evan Atherton, Derek Cheung, and Jacky Bibliowicz. Motion generation: A survey of generative approaches and benchmarks.arXiv preprint arXiv:2507.05419, 2025. 1

    Pith/arXiv arXiv 2025

  33. [33]

    Man Ho Lam, Chaozheng Wang, Jen-tse Huang, and Michael R. Lyu. Codecrash: Exposing llm fragility to mis- leading natural language in code reasoning.In Advances in Neural Information Processing Systems, 38, 2025. 7

    2025

  34. [34]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. InInterna- tional conference on machine learning, pages 12888–12900. PMLR, 2022. 1

    2022

  35. [35]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InIn- ternational conference on machine learning, pages 19730– 19742. PMLR, 2023. 1

    2023

  36. [36]

    Lamp: Language-motion pretraining for motion generation, retrieval, and captioning

    Zhe Li, Weihao Yuan, Yisheng He, Lingteng Qiu, Shen- hao Zhu, Xiaodong Gu, Weichao Shen, Yuan Dong, Zi- long Dong, and Laurence T Yang. Lamp: Language-motion pretraining for motion generation, retrieval, and captioning. arXiv preprint arXiv:2410.07093, 2024. 3, 6, 7

    Pith/arXiv arXiv 2024

  37. [37]

    Re- momask: Retrieval-augmented masked motion generation

    Zhengdao Li, Siheng Wang, Zeyu Zhang, and Hao Tang. Re- momask: Retrieval-augmented masked motion generation. arXiv preprint arXiv:2508.02605, 2025. 3

    arXiv 2025

  38. [38]

    Mixture-of- transformers: A sparse and scalable architecture for multi- modal foundation models.Transactions on Machine Learn- ing Research, 2025

    Weixin Liang, LILI YU, Liang Luo, Srini Iyer, Ning Dong, Chunting Zhou, Gargi Ghosh, Mike Lewis, Wen tau Yih, Luke Zettlemoyer, and Xi Victoria Lin. Mixture-of- transformers: A sparse and scalable architecture for multi- modal foundation models.Transactions on Machine Learn- ing Research, 2025. 9

    2025

  39. [39]

    Rmd: A simple baseline for more general human motion generation via training- free retrieval-augmented motion diffuse.arXiv preprint arXiv:2412.04343, 2024

    Zhouyingcheng Liao, Mingyuan Zhang, Wenjia Wang, Lei Yang, and Taku Komura. Rmd: A simple baseline for more general human motion generation via training- free retrieval-augmented motion diffuse.arXiv preprint arXiv:2412.04343, 2024. 3

    Pith/arXiv arXiv 2024

  40. [40]

    Rouge: A package for automatic evaluation of summaries

    Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. InText summarization branches out, pages 74–81, 2004. 6, 4, 5

    2004

  41. [41]

    The quest for generalizable motion generation: Data, model, and evaluation.arXiv preprint arXiv:2510.26794,

    Jing Lin, Ruisi Wang, Junzhe Lu, Ziqi Huang, Guorui Song, Ailing Zeng, Xian Liu, Chen Wei, Wanqi Yin, Qingping Sun, et al. The quest for generalizable motion generation: Data, model, and evaluation.arXiv preprint arXiv:2510.26794,

    arXiv

  42. [42]

    Panoptic captioning: An equivalence bridge for image and text

    Kun-Yu Lin, Hongjun Wang, Weining Ren, and Kai Han. Panoptic captioning: An equivalence bridge for image and text. 2025. 1

    2025

  43. [43]

    Evaluating text-to-visual generation with image-to-text gen- eration

    Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text gen- eration. InEuropean Conference on Computer Vision, pages 366–384. Springer, 2024. 8

    2024

  44. [44]

    Motionrl: Align text-to-motion generation to human preferences with multi-reward reinforcement learning.arXiv preprint arXiv:2410.06513, 2024

    Xiaoyang Liu, Yunyao Mao, Wengang Zhou, and Houqiang Li. Motionrl: Align text-to-motion generation to human preferences with multi-reward reinforcement learning.arXiv preprint arXiv:2410.06513, 2024. 5, 2

    Pith/arXiv arXiv 2024

  45. [45]

    Scamo: Exploring the scaling law in au- toregressive motion generation model

    Shunlin Lu, Jingbo Wang, Zeyu Lu, Ling-Hao Chen, Wenxun Dai, Junting Dong, Zhiyang Dou, Bo Dai, and Ruimao Zhang. Scamo: Exploring the scaling law in au- toregressive motion generation model. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 27872–27882, 2025. 3

    2025

  46. [46]

    M3gpt: An ad- vanced multimodal, multitask framework for motion com- prehension and generation.Advances in Neural Information Processing Systems, 37:28051–28077, 2024

    Mingshuang Luo, Ruibing Hou, Zhuo Li, Hong Chang, Zimo Liu, Yaowei Wang, and Shiguang Shan. M3gpt: An ad- vanced multimodal, multitask framework for motion com- prehension and generation.Advances in Neural Information Processing Systems, 37:28051–28077, 2024. 3

    2024

  47. [47]

    Troje, Ger- ard Pons-Moll, and Michael J

    Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Ger- ard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. InThe IEEE International Conference on Computer Vision (ICCV), 2019. 3

    2019

  48. [48]

    Learning generalizable human mo- tion generator with reinforcement learning.arXiv preprint arXiv:2405.15541, 2024

    Yunyao Mao, Xiaoyang Liu, Wengang Zhou, Zhenbo Lu, and Houqiang Li. Learning generalizable human mo- tion generator with reinforcement learning.arXiv preprint arXiv:2405.15541, 2024. 5, 6, 7

    Pith/arXiv arXiv 2024

  49. [49]

    Rethinking diffusion for text-driven human motion generation.arXiv preprint arXiv:2411.16575, 2024

    Zichong Meng, Yiming Xie, Xiaogang Peng, Zeyu Han, and Huaizu Jiang. Rethinking diffusion for text-driven human motion generation.arXiv preprint arXiv:2411.16575, 2024. 2, 3, 6, 8, 4, 7, 9, 12

    Pith/arXiv arXiv 2024

  50. [50]

    Absolute coordinates make motion generation easy.arXiv preprint arXiv:2505.19377, 2025

    Zichong Meng, Zeyu Han, Xiaogang Peng, Yiming Xie, and Huaizu Jiang. Absolute coordinates make motion generation easy.arXiv preprint arXiv:2505.19377, 2025. 3

    Pith/arXiv arXiv 2025

  51. [51]

    Motion-r1: Chain-of-thought reasoning and reinforcement learning for human motion generation.arXiv preprint arXiv:2506.10353, 2025

    Runqi Ouyang, Haoyun Li, Zhenyuan Zhang, Xiaofeng Wang, Zheng Zhu, Guan Huang, and Xingang Wang. Motion-r1: Chain-of-thought reasoning and reinforcement learning for human motion generation.arXiv preprint arXiv:2506.10353, 2025. 3, 5, 8, 2, 6, 7

    arXiv 2025

  52. [52]

    Bleu: a method for automatic evaluation of machine translation

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. InProceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318,

  53. [53]

    Modipo: text-to-motion alignment via ai-feedback-driven direct preference optimiza- tion.arXiv preprint arXiv:2405.03803, 2024

    Massimiliano Pappa, Luca Collorone, Giovanni Ficarra, In- dro Spinelli, and Fabio Galasso. Modipo: text-to-motion alignment via ai-feedback-driven direct preference optimiza- tion.arXiv preprint arXiv:2405.03803, 2024. 2, 8

    Pith/arXiv arXiv 2024

  54. [54]

    Black, and G ¨ul Varol

    Mathis Petrovich, Michael J. Black, and G ¨ul Varol. TMR: Text-to-motion retrieval using contrastive 3D human motion synthesis. InInternational Conference on Computer Vision (ICCV), 2023. 5

    2023

  55. [55]

    Bamm: Bidirectional autoregressive motion model

    Ekkasit Pinyoanuntapong, Muhammad Usama Saleem, Pu Wang, Minwoo Lee, Srijan Das, and Chen Chen. Bamm: Bidirectional autoregressive motion model. InEuropean Conference on Computer Vision, pages 172–190. Springer,

  56. [56]

    Mmm: Generative masked motion model

    Ekkasit Pinyoanuntapong, Pu Wang, Minwoo Lee, and Chen Chen. Mmm: Generative masked motion model. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1546–1555, 2024. 3, 8, 6

    2024

  57. [57]

    The kit motion-language dataset.Big data, 4(4):236–252,

    Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset.Big data, 4(4):236–252,

  58. [58]

    Uni-cot: Towards unified chain-of-thought reasoning across text and vision.arXiv preprint arXiv:2508.05606,

    Luozheng Qin, Jia Gong, Yuqing Sun, Tianjiao Li, Meng- ping Yang, Xiaomeng Yang, Chao Qu, Zhiyu Tan, and Hao Li. Uni-cot: Towards unified chain-of-thought reasoning across text and vision.arXiv preprint arXiv:2508.05606,

    arXiv

  59. [59]

    Direct preference optimization: Your language model is secretly a reward model.Advances in neural information processing systems, 36:53728–53741, 2023

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christo- pher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model.Advances in neural information processing systems, 36:53728–53741, 2023. 2

    2023

  60. [60]

    BREAK-THE-CHAIN: Adversarial Prompting in Code Generation, 2025

    Jaechul Roh, Varun Gandhi, Shivani Anilkumar, and Arin Garg. BREAK-THE-CHAIN: Adversarial Prompting in Code Generation, 2025. UMass Amherst CS685 Advanced NLP Project. 7

    2025

  61. [61]

    Proximal policy optimization algo- rithms.arXiv preprint arXiv:1707.06347, 2017

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Rad- ford, and Oleg Klimov. Proximal policy optimization algo- rithms.arXiv preprint arXiv:1707.06347, 2017. 2

    Pith/arXiv arXiv 2017

  62. [62]

    Deepseekmath: Pushing the limits of math- ematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of math- ematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024. 2, 5

    Pith/arXiv arXiv 2024

  63. [63]

    Unlocking pretrained llms for motion-related multimodal generation: A fine-tuning approach to unify diffusion and next-token prediction.arXiv preprint arXiv:2503.06119,

    Shinichi Tanaka, Zhao Wang, Yoichi Kato, and Jun Ohya. Unlocking pretrained llms for motion-related multimodal generation: A fine-tuning approach to unify diffusion and next-token prediction.arXiv preprint arXiv:2503.06119,

    Pith/arXiv arXiv

  64. [64]

    Gemma 2: Improving open language models at a practical size.arXiv preprint arXiv:2408.00118,

    Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupati- raju, L´eonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ram´e, et al. Gemma 2: Improving open language models at a practical size.arXiv preprint arXiv:2408.00118,

    Pith/arXiv arXiv

  65. [65]

    Human motion dif- fusion model.arXiv preprint arXiv:2209.14916, 2022

    Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion dif- fusion model.arXiv preprint arXiv:2209.14916, 2022. 1, 3, 8, 6

    Pith/arXiv arXiv 2022

  66. [66]

    Cider: Consensus-based image description evalua- tion

    Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evalua- tion. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4566–4575, 2015. 6, 4, 5

    2015

  67. [67]

    Aligning human mo- tion generation with human perceptions.arXiv preprint arXiv:2407.02272, 2024

    Haoru Wang, Wentao Zhu, Luyi Miao, Yishu Xu, Feng Gao, Qi Tian, and Yizhou Wang. Aligning human mo- tion generation with human perceptions.arXiv preprint arXiv:2407.02272, 2024. 9

    Pith/arXiv arXiv 2024

  68. [68]

    Image as a foreign language: Beit pretraining for vision and vision- language tasks

    Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhil- iang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mo- hammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for vision and vision- language tasks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19175– 19186, 2023. 1

    2023

  69. [69]

    Motiongpt-2: A general-purpose motion- language model for motion generation and understanding

    Yuan Wang, Di Huang, Yaqi Zhang, Wanli Ouyang, Jile Jiao, Xuetao Feng, Yan Zhou, Pengfei Wan, Shixiang Tang, and Dan Xu. Motiongpt-2: A general-purpose motion- language model for motion generation and understanding. arXiv preprint arXiv:2410.21747, 2024. 1, 3, 8, 6, 7

    Pith/arXiv arXiv 2024

  70. [70]

    Scaling large motion models with million-level human motions.arXiv preprint arXiv:2410.03311, 2024

    Ye Wang, Sipeng Zheng, Bin Cao, Qianshan Wei, Weishuai Zeng, Qin Jin, and Zongqing Lu. Scaling large motion models with million-level human motions.arXiv preprint arXiv:2410.03311, 2024. 3, 6, 7

    Pith/arXiv arXiv 2024

  71. [71]

    Mg-motionllm: A unified framework for motion comprehension and gener- ation across multiple granularities

    Bizhu Wu, Jinheng Xie, Keming Shen, Zhe Kong, Jianfeng Ren, Ruibin Bai, Rong Qu, and Linlin Shen. Mg-motionllm: A unified framework for motion comprehension and gener- ation across multiple granularities. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 27849–27858, 2025. 1, 3, 6, 7, 8

    2025

  72. [72]

    Omnigen2: Exploration to advanced multimodal generation.arXiv preprint arXiv:2506.18871, 2025

    Chenyuan Wu, Pengfei Zheng, Ruiran Yan, Shitao Xiao, Xin Luo, Yueze Wang, Wanli Li, Xiyan Jiang, Yexin Liu, Junjie Zhou, et al. Omnigen2: Exploration to advanced multimodal generation.arXiv preprint arXiv:2506.18871, 2025. 3

    Pith/arXiv arXiv 2025

  73. [73]

    Motion-agent: A conversational framework for human motion generation with llms.arXiv preprint arXiv:2405.17013, 2024

    Qi Wu, Yubo Zhao, Yifan Wang, Xinhang Liu, Yu-Wing Tai, and Chi-Keung Tang. Motion-agent: A conversational framework for human motion generation with llms.arXiv preprint arXiv:2405.17013, 2024. 2, 3, 6, 7, 8, 1, 4, 5, 9, 12

    Pith/arXiv arXiv 2024

  74. [74]

    Mote: Learning motion-text diffusion model for multiple generation tasks.arXiv preprint arXiv:2411.19786,

    Yiming Wu, Wei Ji, Kecheng Zheng, Zicheng Wang, and Dong Xu. Mote: Learning motion-text diffusion model for multiple generation tasks.arXiv preprint arXiv:2411.19786,

    Pith/arXiv arXiv

  75. [75]

    Vimorag: Video-based retrieval-augmented 3d mo- tion generation for motion language models.arXiv preprint arXiv:2508.12081, 2025

    Haidong Xu, Guangwei Xu, Zhedong Zheng, Xiatian Zhu, Wei Ji, Xiangtai Li, Ruijie Guo, Meishan Zhang, Hao Fei, et al. Vimorag: Video-based retrieval-augmented 3d mo- tion generation for motion language models.arXiv preprint arXiv:2508.12081, 2025. 8, 6

    arXiv 2025

  76. [76]

    Cross-modal retrieval for motion and text via droptriple loss

    Sheng Yan, Yang Liu, Haoqiang Wang, Xin Du, Mengyuan Liu, and Hong Liu. Cross-modal retrieval for motion and text via droptriple loss. InProceedings of the 5th ACM Interna- tional Conference on Multimedia in Asia, pages 1–7, 2023. 3

    2023

  77. [77]

    Motionscript: Nat- ural language descriptions for expressive 3d human motions

    Payam Jome Yazdian, Eric Liu, Rachel Lagasse, Hamid Mo- hammadi, Li Cheng, and Angelica Lim. Motionscript: Nat- ural language descriptions for expressive 3d human motions. arXiv preprint arXiv:2312.12634, 2023. 3

    arXiv 2023

  78. [78]

    Exploring vision transformers for 3d human motion-language models with motion patches

    Qing Yu, Mikihiro Tanaka, and Kent Fujiwara. Exploring vision transformers for 3d human motion-language models with motion patches. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 937–946, 2024. 9

    2024

  79. [79]

    Remogpt: Part-level retrieval-augmented motion-language models

    Qing Yu, Mikihiro Tanaka, and Kent Fujiwara. Remogpt: Part-level retrieval-augmented motion-language models. In Proceedings of the AAAI Conference on Artificial Intelli- gence, pages 9635–9643, 2025. 3

    2025

  80. [80]

    Mogents: Motion generation based on spatial-temporal joint modeling.Neural Information Processing Systems (NeurIPS), 2024

    Weihao Yuan, Weichao Shen, Yisheng HE, Yuan Dong, Xi- aodong Gu, Zilong Dong, Liefeng Bo, and Qixing Huang. Mogents: Motion generation based on spatial-temporal joint modeling.Neural Information Processing Systems (NeurIPS), 2024. 6, 7

    2024

Showing first 80 references.