Pith. sign in

REVIEW 3 major objections 3 minor 58 references

MAPF-World: Action World Model for Multi-Agent Path Finding

T0 review · 3 major / 3 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read MAPF-World claims that a decentralized pathfinding policy which predicts future states and joint actions before each decision plans more coordinately and generalizes better than much larger reactive solvers, using far less data.

desk verdict The submission is not internally coherent: the abstract announces MAPF-World, an autoregressive action world model for multi-agent path finding, but the supplied full text is an unrelated paper on video-based 3D motion generation (VimoRAG). read the letter →

arxiv 2508.12087 v2 pith:VQKOELMO submitted 2025-08-16 cs.AI cs.MA

Zhanjiang Yang , Yang Shen , Yueming Li , Meng Li , Lijun Sun This is my paper · ORCID
classification cs.AIcs.MA
keywords multi-agentpathfindingworldmodelautoregressiveactionfuturestatepredictiondecentralizedplanningzero-shotgeneralizationreal-worldmapgeneration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Multi-agent path finding (MAPF) asks how to move many agents from starts to goals without collisions, and decentralized learnable solvers scale well but act reactively: each choice uses only the current local view, so coordination over long horizons suffers. This paper proposes MAPF-World, an autoregressive action world model that couples situation understanding with action generation by predicting future states and future actions, then conditioning decisions on those predicted futures. The paper claims this look-ahead outperforms state-of-the-art learnable solvers on MAPF benchmarks, including zero-shot transfer to out-of-distribution maps, while training with 96.5% smaller model size and 92% less data. A secondary contribution is an automatic map generator grounded in real-world layouts for training and evaluating solvers. If correct, the work suggests that foresight, not raw scale, is what decentralized multi-agent planners lack.

What carries the argument

The central object is an 'action world model': an autoregressive model, trained alongside the decentralized MAPF policy, that unifies situation understanding and action generation by predicting future states and future actions step by step, then feeding those predicted futures back into the decision. It carries the argument because it is the mechanism intended to replace reactive, observation-only decision-making with look-ahead: the policy conditions each action on where the model predicts the joint configuration of agents and environment will be, and the paper attributes better coordination and long-horizon performance to this explicit modeling of spatial features and temporal dependencies.

What would settle it

Run MAPF-World on long-horizon, out-of-distribution maps with the future-prediction branch disabled or supplied with random futures instead of its own predictions. If success rates and collision counts stay essentially the same, then look-ahead is not the load-bearing mechanism and the world-model explanation of the gains is unsupported.

Watch

Extended reading notes

Core claim

The central claim is that decentralized MAPF solvers degrade in complex, long-term scenarios because they are reactive policies with weak accounts of environmental temporal dynamics and inter-agent dependencies. MAPF-World answers by unifying situation understanding and action generation in one autoregressive model that explicitly predicts future states and future actions, so each decision is informed by where the joint configuration of agents and environment is heading rather than by immediate local observations alone. The paper attributes the improved situational awareness, coordination, and far-sightedness of its solver to this explicit modeling of spatial features and temporal dependencies. In experiments, MAPF-World is reported to outperform state-of-the-art learnable solvers, to generalize zero-shot to out-of-distribution cases, and to do so with a 96.5% smaller model and 92% less training data, aided by a new real-world-grounded map generator.

Load-bearing premise

The approach assumes that the future states and actions the model predicts faithfully reflect how the multi-agent environment actually evolves; if those predictions come from the same decision process that chooses the actions, the claimed benefit of looking ahead has no support.

Editorial extensions

If this is right

  • On the paper's benchmarks, MAPF-World outperforms prior state-of-the-art learnable solvers while using 96.5% fewer parameters and 92% less training data.
  • The solver exhibits zero-shot generalization to out-of-distribution maps, meaning its performance transfers to layouts unlike those seen in training.
  • The new automatic map generator, grounded in real-world layouts, provides a more practical training and evaluation protocol than purely synthetic benchmark maps.
  • Future-state and future-action prediction is presented as the mechanism that fixes the reactive-policy limitation of prior decentralized learnable solvers.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the result is correct, it implies that the bottleneck for learned MAPF solvers is not capacity but foresight: deliberately predictive objectives could substitute for the scale that foundation-model-based solvers rely on. This is my extrapolation, not the paper's claim.
  • The same predict-then-act recipe transfers naturally to other decentralized sequential-decision settings, such as warehouse robot fleets, traffic intersections, and social navigation, where reactive policies also degrade over long horizons. This is an editorial inference.
  • The abstract does not show that predicted futures are checked against true environment transitions; if the future-prediction head merely re-expresses the decision policy, the reported gains could come from the auxiliary loss's regularization rather than from genuine look-ahead. This is an editorial caveat, not something the paper claims.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The manuscript as submitted consists of an abstract announcing MAPF-World, an autoregressive action world model for multi-agent path finding, followed by a full text that is an unrelated paper titled "VimoRAG: Video-based Retrieval-augmented 3D Motion Generation for Motion Language Models" by a different set of authors. The abstract claims that MAPF-World outperforms state-of-the-art learnable MAPF solvers with superior zero-shot generalization, a 96.5% smaller model, and 92% less training data. However, the body contains no description of the MAPF-World architecture, its world-model prediction mechanism, the automatic map generator, training data, baselines, or evaluation protocol; all sections, equations, tables, and experiments belong to VimoRAG. The central claims are therefore entirely unsupported by the provided manuscript text.

Significance. If the MAPF-World results were actually established, the paper would be significant: it would demonstrate that world-model-based lookahead improves decentralized MAPF decision-making at dramatically lower model and data cost, and it would introduce a grounded map generator for more realistic benchmarks. The current submission, however, provides no method or evidence for any of these claims. There are no parameter-free derivations, machine-checked proofs, reproducible code, or falsifiable predictions pertaining to MAPF in the manuscript; the only concrete experimental content concerns 3D motion generation and is irrelevant to the stated topic. The claim of significance is therefore unverifiable from the submitted artifact.

major comments (3)
  1. [Abstract, first/last paragraphs] The full text of the submission is the VimoRAG paper, not the MAPF-World paper promised by the title and abstract. No section or equation describes the MAPF-World model, its autoregressive future-state/action prediction, the automatic map generator, the training protocol, or the MAPF benchmarks. Consequently, the abstract's central claims of superiority over state-of-the-art learnable solvers, zero-shot out-of-distribution generalization, and 96.5%/92% efficiency gains are completely unverifiable from the submitted text. This is not a local gap but the absence of the entire artifact needed to evaluate the contribution.
  2. [Abstract, first/last paragraphs] The abstract asserts "Extensive experiments demonstrate that MAPF-World outperforms state-of-the-art learnable solvers" and reports specific efficiency numbers, but the manuscript contains no MAPF experiments, no baseline comparisons, no dataset descriptions, and no evaluation metrics. The provided Tables 1–8 and Figures 1–13 all belong to VimoRAG and concern text-to-motion generation, so no numerical evidence bears on the MAPF claims. The paper therefore fails to satisfy the basic requirement that its central empirical assertions be accompanied by methods and results.
  3. [Abstract, paragraph 2] The proposal hinges on "explicitly modeling environmental dynamics... through future state and actions prediction," but the manuscript provides no definition of the prediction target, no loss function, and no evidence that the predicted futures are grounded in actual environment transitions as opposed to generated by the decision policy itself. Without such material, the claimed mechanism of "informed, coordinated, and far-sighted decision-making" cannot be assessed even in principle.
minor comments (3)
  1. [Title and metadata] The title, author list, and abstract of the manuscript do not match the title, authors, or subject matter of the body; this suggests a submission error that must be corrected before any review can proceed.
  2. [References] The abstract cites MAPF-related concepts and benchmarks, but the reference list contains only citations for motion generation and video retrieval; no MAPF-related literature is cited anywhere in the manuscript.
  3. [Section 5 (Discussion)] The Limitations subsection addresses latency of video-augmented motion generation and is unrelated to the MAPF-World claims; it does not acknowledge the absence of experimental validation for the abstract's assertions.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; supplied full text is an unrelated paper, so MAPF-World claims are unverified rather than circular.

full rationale

The abstract under review describes MAPF-World, an autoregressive action world model for multi-agent path finding, and claims superior zero-shot generalization and reduced model/data requirements. However, the supplied full text is a different manuscript, VimoRAG, on video-based retrieval-augmented 3D motion generation, with different authors, title, and subject matter. None of MAPF-World's equations, architecture, training procedures, map generator, or benchmark comparisons appear in the provided material. Consequently, there is no derivation chain in which a claimed prediction or result can be shown to reduce to its own inputs by construction. No quoted reduction from the MAPF-World content is possible, and no fitted parameter is renamed as a prediction. The VimoRAG text itself evaluates against external benchmarks and does not exhibit a self-definitional or self-citation-dependent core. The appropriate finding is therefore that no circularity is identifiable from the supplied evidence; the MAPF-World claims are unverified because the supporting artifact is absent, not because the argument is circular.

Assumptions & free parameters 0 free parameters · 2 assumptions · 0 invented entities

From the abstract alone, no free parameters can be identified; the central claims rest on two untested domain assumptions: that future-prediction improves MAPF decisions, and that the synthetic map generator is representative. The supplied full text is a different paper, so no further parameters or axioms can be audited.

assumptions (2)
  • domain assumption Predicting future states and actions improves multi-agent decision quality
    Central premise asserted in the abstract; no proof or ablation is available in the supplied full text because the full text is an unrelated paper.
  • domain assumption The automatic map generator produces layouts that are representative of real-world scenarios and suitable for training and evaluation
    The abstract claims the generator is grounded in real-world scenarios but gives no validation details.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MAPF-World: Action World Model for Multi-Agent Path Finding." pith.science (2026). https://pith.science/paper/VQKOELMO

@misc{pith2026250812087,
  author       = {Pith},
  title        = {Pith review of: MAPF-World: Action World Model for Multi-Agent Path Finding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VQKOELMO}},
  note         = {Machine review of arXiv:2508.12087}
}
read the original abstract

Multi-agent path finding (MAPF) is the problem of planning conflict-free paths from the designated start locations to goal positions for multiple agents. It underlies a variety of real-world tasks, including multi-robot coordination, robot-assisted logistics, and social navigation. Recent decentralized learnable solvers have shown great promise for large-scale MAPF, especially when leveraging foundation models and large datasets. However, these agents are reactive policy models and exhibit limited modeling of environmental temporal dynamics and inter-agent dependencies, resulting in performance degradation in complex, long-term planning scenarios. To address these limitations, we propose MAPF-World, an autoregressive action world model for MAPF that unifies situation understanding and action generation, guiding decisions beyond immediate local observations. It improves situational awareness by explicitly modeling environmental dynamics, including spatial features and temporal dependencies, through future state and actions prediction. By incorporating these predicted futures, MAPF-World enables more informed, coordinated, and far-sighted decision-making, especially in complex multi-agent settings. Furthermore, we augment MAPF benchmarks by introducing an automatic map generator grounded in real-world scenarios, capturing practical map layouts for training and evaluating MAPF solvers. Extensive experiments demonstrate that MAPF-World outperforms state-of-the-art learnable solvers, showcasing superior zero-shot generalization to out-of-distribution cases. Notably, MAPF-World is trained with a 96.5% smaller model size and 92% reduced data.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

58 extracted references · 36 canonical work pages

  1. [1]

    Generating human motion from textual descriptions with discrete representations

    Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In Proceedings of the CVPR, pages 14730–14740, 2023

    work page 2023

  2. [2]

    Motiondiffuse: Text-driven human motion generation with diffusion model.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(6):4115–4128, 2024

    Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(6):4115–4128, 2024

    2024

  3. [3]

    Motiongpt: Finetuned llms are general-purpose motion generators

    Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Motiongpt: Finetuned llms are general-purpose motion generators. InProceedings of the AAAI, pages 7368–7376, 2024

    work page 2024

  4. [4]

    Motiongpt: Human motion as a foreign language.Proceedings of the NeurIPS, pages 20067–20079, 2023

    Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language.Proceedings of the NeurIPS, pages 20067–20079, 2023

    work page 2023

  5. [5]

    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 CVPR, pages 5152–5161, 2022

    work page 2022

  6. [6]

    Remodiffuse: Retrieval-augmented motion diffusion model

    Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. InProceedings of the ICCV, pages 364–373, 2023

    work page 2023

  7. [7]

    Vimo: Generating motions from casual videos.arXiv preprint arXiv:2408.06614, 2024

    Liangdong Qiu, Chengxing Yu, Yanran Li, Zhao Wang, Haibin Huang, Chongyang Ma, Di Zhang, Pengfei Wan, and Xiaoguang Han. Vimo: Generating motions from casual videos.arXiv preprint arXiv:2408.06614, 2024

    arXiv 2024

  8. [8]

    Neural discrete representation learning.Proceedings of the NeurIPS, 30, 2017

    Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning.Proceedings of the NeurIPS, 30, 2017

    work page 2017

Show all 58 references
  1. [9]

    Phi-3 technical report: A highly capable language model locally on your phone.arXiv preprint arXiv:2404.14219, 2024

    Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone.arXiv preprint arXiv:2404.14219, 2024

    2024 arXiv

  2. [10]

    Videogpt+: Integrating image and video encoders for enhanced video understanding.arXiv preprint arXiv:2406.09418, 2024

    Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Khan. Videogpt+: Integrating image and video encoders for enhanced video understanding.arXiv preprint arXiv:2406.09418, 2024

    2024 arXiv

  3. [11]

    Motion-x: a large-scale 3d expressive whole-body human motion dataset

    Jing Lin, Ailing Zeng, Shunlin Lu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, and Lei Zhang. Motion-x: a large-scale 3d expressive whole-body human motion dataset. InProceedings of the NeurIPS, pages 25268–25280, 2023

    2023

  4. [12]

    Human motion generation: A survey.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(4):2430–2449, 2024

    Wentao Zhu, Xiaoxuan Ma, Dongwoo Ro, Hai Ci, Jinlu Zhang, Jiaxin Shi, Feng Gao, Qi Tian, and Yizhou Wang. Human motion generation: A survey.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(4):2430–2449, 2024

    2024

  5. [13]

    Momask: Generative masked modeling of 3d human motions

    Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. InProceedings of the CVPR, pages 1900–1910, 2024

    1900

  6. [14]

    Omg: Towards open-vocabulary motion generation via mixture of controllers

    Han Liang, Jiacheng Bao, Ruichi Zhang, Sihan Ren, Yuecheng Xu, Sibei Yang, Xin Chen, Jingyi Yu, and Lan Xu. Omg: Towards open-vocabulary motion generation via mixture of controllers. InProceedings of the CVPR, pages 482–493, 2024

    2024

  7. [15]

    Motionclip: Exposing human motion generation to clip space

    Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. InProceedings of the ECCV, pages 358–374, 2022

    2022

  8. [16]

    Plan, posture and go: Towards open-vocabulary text-to-motion generation

    Jinpeng Liu, Wenxun Dai, Chunyu Wang, Yiji Cheng, Yansong Tang, and Xin Tong. Plan, posture and go: Towards open-vocabulary text-to-motion generation. InProceedings of the ECCV, pages 445–463, 2025

    2025

  9. [17]

    Textual decomposition then sub-motion-space scattering for open-vocabulary motion generation.arXiv preprint arXiv:2411.04079, 2024

    Ke Fan, Jiangning Zhang, Ran Yi, Jingyu Gong, Yabiao Wang, Yating Wang, Xin Tan, Chengjie Wang, and Lizhuang Ma. Textual decomposition then sub-motion-space scattering for open-vocabulary motion generation.arXiv preprint arXiv:2411.04079, 2024

    2024 arXiv

  10. [18]

    Being comes from not-being: Open-vocabulary text-to-motion generation with wordless training

    Junfan Lin, Jianlong Chang, Lingbo Liu, Guanbin Li, Liang Lin, Qi Tian, and Chang-wen Chen. Being comes from not-being: Open-vocabulary text-to-motion generation with wordless training. InProceedings of the CVPR, pages 23222–23231, 2023

    2023

  11. [19]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. Attention is all you need. InProceedings of the NeurIPS, 2017. 11

    2017

  12. [20]

    Human motion diffusion model

    Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. InProceedings of the ICLR, 2023

    2023

  13. [21]

    Denoising diffusion probabilistic models.Proceedings of the NeurIPS, pages 6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Proceedings of the NeurIPS, pages 6840–6851, 2020

    2020

  14. [22]

    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 CVPR, pages 18000–18010, 2023

    2023

  15. [23]

    Openagi: When llm meets domain experts

    Yingqiang Ge, Wenyue Hua, Kai Mei, jianchao ji, Juntao Tan, Shuyuan Xu, Zelong Li, and Yongfeng Zhang. Openagi: When llm meets domain experts. InProceedings of the NeurIPS, pages 5539–5568, 2023

    2023

  16. [24]

    Motionllm: Understanding human behaviors from human motions and videos.arXiv preprint arXiv:2405.20340, 2024

    Ling-Hao Chen, Shunlin Lu, Ailing Zeng, Hao Zhang, Benyou Wang, Ruimao Zhang, and Lei Zhang. Motionllm: Understanding human behaviors from human motions and videos.arXiv preprint arXiv:2405.20340, 2024

    2024 arXiv

  17. [25]

    Motiongpt-2: A general-purpose motion-language model for motion generation and understanding.arXiv preprint arXiv:2410.21747, 2024

    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

    arXiv

  18. [26]

    Motionbank: A large-scale video motion benchmark with disentangled rule-based annotations.arXiv preprint arXiv:2410.13790, 2024

    Liang Xu, Shaoyang Hua, Zili Lin, Yifan Liu, Feipeng Ma, Yichao Yan, Xin Jin, Xiaokang Yang, and Wenjun Zeng. Motionbank: A large-scale video motion benchmark with disentangled rule-based annotations.arXiv preprint arXiv:2410.13790, 2024

    arXiv

  19. [27]

    Large motion model for unified multi-modal motion generation

    Mingyuan Zhang, Daisheng Jin, Chenyang Gu, Fangzhou Hong, Zhongang Cai, Jingfang Huang, Chongzhi Zhang, Xinying Guo, Lei Yang, Ying He, et al. Large motion model for unified multi-modal motion generation. InProceedings of the ECCV, pages 397–421, 2024

    2024

  20. [28]

    The action similarity labeling challenge.IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(3):615–621, 2011

    Orit Kliper-Gross, Tal Hassner, and Lior Wolf. The action similarity labeling challenge.IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(3):615–621, 2011

    2011

  21. [29]

    Hmdb: a large video database for human motion recognition

    Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. InProceedings of the ICCV, pages 2556–2563, 2011

    2011

  22. [30]

    The kinetics human action video dataset.arXiv preprint arXiv:1705.06950, 2017

    Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset.arXiv preprint arXiv:1705.06950, 2017

    2017 arXiv

  23. [31]

    From actemes to action: A strongly-supervised representa- tion for detailed action understanding

    Weiyu Zhang, Menglong Zhu, and Derpanis G. From actemes to action: A strongly-supervised representa- tion for detailed action understanding. InProceedings of the ICCV, pages 2248–2255, 2013

    2013

  24. [32]

    UCF101: A dataset of 101 human actions classes from videos in the wild.CoRR, abs/1212.0402, 2012

    Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild.CoRR, abs/1212.0402, 2012

    2012 arXiv

  25. [33]

    Ntu rgb+ d: A large scale dataset for 3d human activity analysis

    Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. InProceedings of the CVPR, pages 1010–1019, 2016

    2016

  26. [34]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024

    2024 arXiv

  27. [35]

    Humanomni: A large vision-speech language model for human-centric video understanding.arXiv preprint arXiv:2501.15111, 2025

    Jiaxing Zhao, Qize Yang, Yixing Peng, Detao Bai, Shimin Yao, Boyuan Sun, Xiang Chen, Shenghao Fu, Xihan Wei, Liefeng Bo, et al. Humanomni: A large vision-speech language model for human-centric video understanding.arXiv preprint arXiv:2501.15111, 2025

    2025 arXiv

  28. [36]

    Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7157–7173, 2022

    Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7157–7173, 2022

    2022

  29. [37]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InProceedings of the ICML, pages 8748–8763, 2021

    2021

  30. [38]

    Motionbert: A unified perspective on learning human motion representations

    Wentao Zhu, Xiaoxuan Ma, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. InProceedings of the ICCV, pages 15085–15099, 2023

    2023

  31. [39]

    Internvideo: General video foundation models via generative and discriminative learning.arXiv preprint arXiv:2212.03191, 2022

    Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning.arXiv preprint arXiv:2212.03191, 2022. 12

    2022 arXiv

  32. [40]

    Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning.Neurocomputing, 508:293–304, 2022

    Huaishao Luo, Lei Ji, Ming Zhong, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning.Neurocomputing, 508:293–304, 2022

    2022

  33. [41]

    Exploring text-to-motion generation with human preference

    Jenny Sheng, Matthieu Lin, Andrew Zhao, Kevin Pruvost, Yu-Hui Wen, Yangguang Li, Gao Huang, and Yong-Jin Liu. Exploring text-to-motion generation with human preference. InProceedings of the CVPR, pages 1888–1899, 2024

    2024

  34. [42]

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

    Massimiliano Pappa, Luca Collorone, Indro Spinelli, and Fabio Galasso. Modipo: text-to-motion alignment via ai-feedback-driven direct preference optimization.arXiv preprint arXiv:2405.03803, 2024

    2024 arXiv

  35. [43]

    Direct preference optimization: your language model is secretly a reward model

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. Direct preference optimization: your language model is secretly a reward model. InProceedings of the NeurIPS, pages 53728–53741, 2023

    2023

  36. [44]

    Direct preference optimization of video large multimodal models from language model reward.arXiv preprint arXiv:2404.01258, 2024

    Ruohong Zhang, Liangke Gui, Zhiqing Sun, Yihao Feng, Keyang Xu, Yuanhan Zhang, Di Fu, Chunyuan Li, Alexander Hauptmann, Yonatan Bisk, et al. Direct preference optimization of video large multimodal models from language model reward.arXiv preprint arXiv:2404.01258, 2024

    2024 arXiv

  37. [45]

    Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685, 2021

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685, 2021

    2021 arXiv

  38. [46]

    Motionlab: Unified human motion generation and editing via the motion-condition-motion paradigm.arXiv preprint arXiv:2502.02358, 2025

    Ziyan Guo, Zeyu Hu, Na Zhao, and De Wen Soh. Motionlab: Unified human motion generation and editing via the motion-condition-motion paradigm.arXiv preprint arXiv:2502.02358, 2025

    2025 arXiv

  39. [47]

    Motionlcm: Real- time controllable motion generation via latent consistency model

    Wenxun Dai, Ling-Hao Chen, Jingbo Wang, Jinpeng Liu, Bo Dai, and Yansong Tang. Motionlcm: Real- time controllable motion generation via latent consistency model. InProceedings of the ECCV, pages 390–408, 2024

    2024

  40. [48]

    Motionclr: Motion generation and training-free editing via understanding attention mechanisms.arXiv e-prints, pages arXiv–2410, 2024

    Ling-Hao Chen, Wenxun Dai, Xuan Ju, Shunlin Lu, and Lei Zhang. Motionclr: Motion generation and training-free editing via understanding attention mechanisms.arXiv e-prints, pages arXiv–2410, 2024

    2024

  41. [49]

    Bipo: Bidirec- tional partial occlusion network for text-to-motion synthesis.arXiv preprint arXiv:2412.00112, 2024

    Seong-Eun Hong, Soobin Lim, Juyeong Hwang, Minwook Chang, and Hyeongyeop Kang. Bipo: Bidirec- tional partial occlusion network for text-to-motion synthesis.arXiv preprint arXiv:2412.00112, 2024

    2024

  42. [50]

    Stablemofusion: Towards robust and efficient diffusion-based motion generation framework

    Yiheng Huang, Hui Yang, Chuanchen Luo, Yuxi Wang, Shibiao Xu, Zhaoxiang Zhang, Man Zhang, and Junran Peng. Stablemofusion: Towards robust and efficient diffusion-based motion generation framework. InProceedings of the MM, pages 224–232, 2024

    2024

  43. [51]

    Mogents: Motion generation based on spatial-temporal joint modeling.Proceedings of the NeurIPS, pages 130739–130763, 2024

    Weihao Yuan, Yisheng He, Weichao Shen, Yuan Dong, Xiaodong Gu, Zilong Dong, Liefeng Bo, and Qixing Huang. Mogents: Motion generation based on spatial-temporal joint modeling.Proceedings of the NeurIPS, pages 130739–130763, 2024

    2024

  44. [52]

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

    Zhe Li, Weihao Yuan, Yisheng HE, Lingteng Qiu, Shenhao Zhu, Xiaodong Gu, Weichao Shen, Yuan Dong, Zilong Dong, and Laurence Tianruo Yang. LaMP: Language-motion pretraining for motion generation, retrieval, and captioning. InProceedings of the ICLR, 2025

    2025

  45. [53]

    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

    2024 arXiv

  46. [54]

    Quo vadis, motion generation? from large language models to large motion models.arXiv preprint arXiv:2410.03311, 2024

    Ye Wang, Sipeng Zheng, Bin Cao, Qianshan Wei, Qin Jin, and Zongqing Lu. Quo vadis, motion generation? from large language models to large motion models.arXiv preprint arXiv:2410.03311, 2024

    2024 arXiv

  47. [55]

    Scamo: Exploring the scaling law in autoregressive motion generation model.arXiv preprint arXiv:2412.14559, 2024

    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 autoregressive motion generation model.arXiv preprint arXiv:2412.14559, 2024

    2024 arXiv

  48. [56]

    Avatargpt: All-in-one framework for motion understanding planning generation and beyond

    Zixiang Zhou, Yu Wan, and Baoyuan Wang. Avatargpt: All-in-one framework for motion understanding planning generation and beyond. InProceedings of the CVPR, pages 1357–1366, 2024

    2024

  49. [57]

    Internvideo2: Scaling foundation models for multimodal video understanding

    Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Zun Wang, Yansong Shi, et al. Internvideo2: Scaling foundation models for multimodal video understanding. InProceedings of the ECCV, pages 396–416, 2025

    2025

  50. [58]

    Please describe the person’s actions in the video using a single sentence that contains a series of verbs

    Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus).arXiv preprint arXiv:1606.08415, 2016. 13 In theappendix, we present more experimental settings and results (Section A), more qualitative results (Section B), details of human-centric video database (Section C...

    2016 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.