arxiv: 2603.24936 · v2 · submitted 2026-03-26 · 💻 cs.CV · cs.AI

Recognition: unknown

TIGFlow-GRPO: Trajectory Forecasting via Interaction-Aware Flow Matching and Reward-Guided Optimization

Hao Meng, Jianguo Wei, Wenhuan Lu, Xuepeng Jing, Zhizhi Yu

Authors on Pith no claims yet

Pith reviewed 2026-05-15 00:52 UTC · model grok-4.3

classification 💻 cs.CV cs.AI

keywords trajectory forecastingflow matchinginteraction graphreward optimizationsocial compliancephysical feasibilitycrowd simulation

0 comments

The pith

A two-stage model first encodes interactions with graphs then aligns flow predictions to social and physical rules via reward optimization.

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

The paper introduces TIGFlow-GRPO to fix the gap where pure supervised flow matching produces trajectories that ignore social norms and scene constraints. Stage one builds a conditional flow matching predictor around a Trajectory-Interaction-Graph module that captures agent-agent and agent-scene relations. Stage two converts deterministic flow rollout into stochastic SDE sampling and applies GRPO using a composite reward that scores view-aware compliance and map-aware feasibility. A sympathetic reader cares because the result is forecasts that stay accurate farther into the future and respect real behavioral constraints, which matters for safe crowd surveillance and autonomous navigation.

Core claim

TIGFlow-GRPO shows that reformulating flow-based trajectory generation as stochastic ODE-to-SDE sampling and steering the samples with GRPO against a composite reward for social compliance and physical feasibility produces multimodal predictions that are more accurate, more stable over long horizons, and more behaviorally plausible than supervised baselines.

What carries the argument

The Trajectory-Interaction-Graph (TIG) module that strengthens conditional features for agent and scene relations, together with Flow-GRPO post-training that uses SDE rollout and reward evaluation to align outputs with behavioral rules.

If this is right

Forecasting accuracy rises on the ETH/UCY and SDD benchmarks.
Long-horizon stability of generated trajectories increases.
Predicted paths become more socially compliant with surrounding agents.
Trajectories respect physical scene constraints more closely.

Where Pith is reading between the lines

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

The same two-stage pattern of interaction encoding followed by reward alignment could transfer to other constrained generative tasks such as robot motion planning.
Varying the reward weights might expose which norms matter most for compliance in different environments.
Replacing the current SDE rollout with other exploration mechanisms could test whether the benefit is specific to stochastic flow sampling.

Load-bearing premise

The composite reward correctly measures social norms and physical feasibility without bias, and the SDE rollout supplies exploration that GRPO can usefully optimize.

What would settle it

If the GRPO stage produces no measurable gain in social-compliance or physical-feasibility scores relative to the first-stage TIGFlow model on held-out sequences from the same ETH/UCY or SDD splits, the value of the reward-guided alignment step would be falsified.

Figures

Figures reproduced from arXiv: 2603.24936 by Hao Meng, Jianguo Wei, Wenhuan Lu, Xuepeng Jing, Zhizhi Yu.

**Figure 1.** Figure 1: Representative examples of social interaction patterns and scene semantics. From left to right, the teaser shows group [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** Figure 2: Overview of TIGFlow-GRPO. A spatio-temporal encoder produces context tokens from historical trajectories and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Flow-GRPO reward modeling and policy optimization. Predicted trajectories are organized into [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Figure 4: Qualitative results on the ETH/UCY benchmark. Compared with DD-MDN, TIGFlow-GRPO produces predictions that [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

read the original abstract

Human trajectory forecasting is important for intelligent multimedia systems operating in visually complex environments, such as autonomous driving and crowd surveillance. Although Conditional Flow Matching (CFM) has shown strong ability in modeling trajectory distributions from spatio-temporal observations, existing approaches still focus primarily on supervised fitting, which may leave social norms and scene constraints insufficiently reflected in generated trajectories. To address this issue, we propose TIGFlow-GRPO, a two-stage generative approach that aligns flow-based trajectory generation with behavioral rules. In the first stage, we build a CFM-based predictor with a Trajectory-Interaction-Graph (TIG) module to model fine-grained visual-spatial interactions and strengthen context encoding. This stage captures both agent-agent and agent-scene relations more effectively, providing more informative conditional features for subsequent alignment. In the second stage, we perform Flow-GRPO post-training, where deterministic flow rollout is reformulated as stochastic ODE-to-SDE sampling to enable trajectory exploration, and a composite reward combines view-aware social compliance with map-aware physical feasibility. By evaluating trajectories explored through SDE rollout, GRPO progressively steers multimodal predictions toward behaviorally plausible futures. Experiments on the ETH/UCY and SDD datasets show that TIGFlow-GRPOimproves forecasting accuracy and long-horizon stability while generatingtrajectories that are more socially compliant and physically feasible.These results suggest that the proposed approach provides an effective way to connectflow-based trajectory modeling with behavior-aware alignment in dynamic multimedia environments.

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.

Desk Editor's Note private letter to a colleague

TIGFlow-GRPO adds a TIG module to CFM and then applies Flow-GRPO with SDE sampling plus a composite reward, but the abstract supplies no numbers or ablations so the gains remain unverified.

read the letter

The key point is that TIGFlow-GRPO is a two-stage method: a conditional flow matching predictor enhanced with a Trajectory-Interaction-Graph module for modeling interactions, followed by Flow-GRPO that uses SDE sampling and a composite reward to align outputs with social and physical rules. The abstract claims better forecasting accuracy and long-horizon stability on ETH/UCY and SDD, plus more compliant trajectories. The new element is the TIG module inside CFM for fine-grained visual-spatial interactions and the specific Flow-GRPO reformulation that turns deterministic rollout into stochastic sampling for GRPO. This combination isn't a direct copy of prior work. It does well at outlining how the first stage captures agent-agent and agent-scene relations to provide better conditional features, and at explaining the SDE rollout for exploration in the alignment stage. The main soft spot is the lack of detail on results. No numbers, baselines, or ablations are mentioned, so it's hard to assess if the gains come from the method or other factors. The stress-test concern about unvalidated reward biases and untested SDE efficacy holds up here, as there's no sensitivity analysis or comparison to other fine-tuning approaches. This makes the central claim rest on unshown evidence. This is for people in CV and robotics focused on trajectory forecasting in dynamic environments. Readers working on generative models with RL post-training might find the pipeline worth looking at for ideas, even if they need to verify the experiments themselves. I would recommend sending it for peer review. The idea is solid enough and grounded in established techniques to warrant referee input, though the authors will probably need to add substantial empirical details and ablations.

Referee Report

2 major / 2 minor

Summary. The paper proposes TIGFlow-GRPO, a two-stage generative approach for human trajectory forecasting. The first stage uses Conditional Flow Matching (CFM) augmented by a Trajectory-Interaction-Graph (TIG) module to encode agent-agent and agent-scene interactions. The second stage reformulates deterministic CFM rollout as stochastic SDE sampling and applies Flow-GRPO optimization driven by a composite reward that combines view-aware social compliance with map-aware physical feasibility. The central claim is that this yields improved forecasting accuracy, long-horizon stability, and more socially compliant and physically feasible trajectories on the ETH/UCY and SDD benchmarks.

Significance. If the empirical results hold after proper validation, the work would usefully extend flow-matching methods by adding a post-training alignment stage that incorporates behavioral constraints, addressing a known limitation of purely supervised generative models in dynamic scenes. The TIG module and the SDE-to-GRPO pipeline constitute a concrete technical contribution that could be adopted in autonomous driving and surveillance pipelines.

major comments (2)

[Experiments] Experiments section: the headline claim that TIGFlow-GRPO improves accuracy and compliance rests on the Flow-GRPO stage, yet the manuscript supplies no ablation that isolates the contribution of individual reward terms (social-compliance vs. physical-feasibility) or that varies SDE noise scale. Without these controls it is impossible to rule out reward hacking or dataset-specific artifacts as the source of any observed gains.
[Abstract] Abstract and Experiments: no quantitative numbers, ADE/FDE values, baseline comparisons, error bars, or statistical significance tests are reported for the ETH/UCY and SDD results, leaving the magnitude and reliability of the claimed improvements unsupported in the provided text.

minor comments (2)

Abstract contains two typographical errors: 'TIGFlow-GRPOimproves' should read 'TIGFlow-GRPO improves' and 'generatingtrajectories' should read 'generating trajectories'.
The composite reward function should be defined explicitly (including weight selection procedure) in the main text rather than left at the level of the abstract description.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the major comments point by point below and will incorporate the suggested changes in the revised manuscript to strengthen the empirical validation.

read point-by-point responses

Referee: [Experiments] Experiments section: the headline claim that TIGFlow-GRPO improves accuracy and compliance rests on the Flow-GRPO stage, yet the manuscript supplies no ablation that isolates the contribution of individual reward terms (social-compliance vs. physical-feasibility) or that varies SDE noise scale. Without these controls it is impossible to rule out reward hacking or dataset-specific artifacts as the source of any observed gains.

Authors: We agree that isolating the contribution of each reward term and varying the SDE noise scale is necessary to substantiate the claims and mitigate concerns about reward hacking. In the revised version we will add dedicated ablation tables and figures that separately disable the social-compliance reward, the physical-feasibility reward, and sweep the SDE noise scale, reporting the resulting ADE/FDE and compliance metrics on both ETH/UCY and SDD. revision: yes
Referee: [Abstract] Abstract and Experiments: no quantitative numbers, ADE/FDE values, baseline comparisons, error bars, or statistical significance tests are reported for the ETH/UCY and SDD results, leaving the magnitude and reliability of the claimed improvements unsupported in the provided text.

Authors: We acknowledge that the current abstract is purely qualitative. We will revise the abstract to report the key ADE/FDE improvements, list the main baselines, and reference the error bars and statistical tests already computed in the Experiments section. We will also ensure the Experiments section explicitly highlights these quantitative results, error bars, and significance tests in the main text and tables for immediate visibility. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical two-stage pipeline with independent training and post-training stages

full rationale

The paper presents TIGFlow-GRPO as a two-stage empirical method: first-stage supervised CFM training with a TIG module for interaction modeling, followed by Flow-GRPO post-training that converts deterministic rollout to stochastic SDE sampling and optimizes via a composite reward. No equations or steps reduce predictions or results by construction to fitted parameters, self-definitions, or self-citations. Central claims rest on experimental outcomes on ETH/UCY and SDD datasets rather than mathematical equivalence to inputs. The derivation chain is self-contained against external benchmarks and does not invoke load-bearing self-citations or ansatzes that collapse to the target result.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

Based solely on the abstract, the approach rests on standard assumptions of flow matching and RL fine-tuning plus newly introduced components whose details are not provided.

free parameters (1)

reward weights
Composite reward combines view-aware social compliance with map-aware physical feasibility; weights are not specified and must be chosen or fitted.

axioms (1)

domain assumption SDE rollout enables effective exploration of multimodal futures without introducing artifacts
Abstract states deterministic flow rollout is reformulated as stochastic ODE-to-SDE sampling to enable trajectory exploration.

invented entities (1)

Trajectory-Interaction-Graph (TIG) module no independent evidence
purpose: Model fine-grained visual-spatial interactions for context encoding
New module introduced in first stage to strengthen agent-agent and agent-scene relations

pith-pipeline@v0.9.0 · 5579 in / 1301 out tokens · 39851 ms · 2026-05-15T00:52:15.657575+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

50 extracted references · 12 canonical work pages · 6 internal anchors

[1]

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

2025
[2]

Inhwan Bae, Jean Oh, and Hae-Gon Jeon. 2023. Eigentrajectory: Low-rank descriptors for multi-modal trajectory forecasting. InProceedings of the IEEE/CVF International Conference on Computer Vision. 10017–10029

2023
[3]

Inhwan Bae, Young-Jae Park, and Hae-Gon Jeon. 2024. Singulartrajectory: Uni- versal trajectory predictor using diffusion model. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 17890–17901

2024
[4]

Mohammadhossein Bahari, Saeed Saadatnejad, Amirhossein Askari Farsangi, Seyed-Mohsen Moosavi-Dezfooli, and Alexandre Alahi. 2025. Certified human trajectory prediction. InProceedings of the Computer Vision and Pattern Recogni- tion Conference. 12301–12311

2025
[5]

Prafulla Dhariwal and Alexander Nichol. 2021. Diffusion models beat gans on image synthesis.Advances in neural information processing systems34 (2021), 8780–8794

2021
[6]

Carles Domingo-Enrich, Michal Drozdzal, Brian Karrer, and Ricky TQ Chen
[7]

Adjoint matching: Fine-tuning flow and diffusion generative models with memoryless stochastic optimal control.arXiv preprint arXiv:2409.08861(2024)

work page arXiv 2024
[8]

Zhiwei Dong, Ran Ding, Wei Li, Peng Zhang, Guobin Tang, and Jia Guo. 2025. Leveraging sd map to augment hd map-based trajectory prediction. InProceedings of the Computer Vision and Pattern Recognition Conference. 17219–17228

2025
[9]

Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. 2024. Scaling rectified flow transformers for high-resolution image synthesis. InForty- first international conference on machine learning

2024
[10]

Jiajun Fan, Shuaike Shen, Chaoran Cheng, Yuxin Chen, Chumeng Liang, and Ge Liu. 2025. Online reward-weighted fine-tuning of flow matching with wasserstein regularization. InThe Thirteenth International Conference on Learning Represen- tations

2025
[11]

Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. 2023. Reinforcement learning for fine-tuning text-to-image diffusion models. InThirty- seventh Conference on Neural Information Processing Systems (NeurIPS) 2023. Neural Information Processing Systems Foundation

2023
[12]

Zilin Fang, David Hsu, Gim Hee Lee, and Gim Hee Lee. 2025. Neuralized Markov Random Field for Interaction-Aware Stochastic Human Trajectory Prediction.. InICLR

2025
[13]

Yuxiang Fu, Qi Yan, Lele Wang, Ke Li, and Renjie Liao. 2025. Moflow: One-step flow matching for human trajectory forecasting via implicit maximum likelihood estimation based distillation. InProceedings of the Computer Vision and Pattern Recognition Conference. 17282–17293

2025
[14]

Tianci Gao, Yuzhen Zhang, Hang Guo, and Pei Lv. 2025. SocialMP: Learning Social Aware Motion Patterns via Additive Fusion for Pedestrian Trajectory Prediction. InProceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence. 90–98

2025
[15]

Itai Gat, Tal Remez, Neta Shaul, Felix Kreuk, Ricky TQ Chen, Gabriel Synnaeve, Yossi Adi, and Yaron Lipman. 2024. Discrete flow matching.Advances in Neural Information Processing Systems37 (2024), 133345–133385

2024
[16]

Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yongming Rao, Jie Zhou, and Jiwen Lu. 2022. Stochastic trajectory prediction via motion indeterminacy diffusion. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 17113–17122

2022
[17]

Agrim Gupta, Justin Johnson, Li Fei-Fei, Silvio Savarese, and Alexandre Alahi
[18]

InProceedings of the IEEE conference on computer vision and pattern recognition

Social gan: Socially acceptable trajectories with generative adversarial networks. InProceedings of the IEEE conference on computer vision and pattern recognition. 2255–2264
[19]

Manuel Hetzel, Hannes Reichert, Konrad Doll, and Bernhard Sick. 2024. Reli- able probabilistic human trajectory prediction for autonomous applications. In European Conference on Computer Vision. Springer, 135–152

2024
[20]

Manuel Hetzel, Kerim Turacan, Hannes Reichert, Konrad Doll, and Bernhard Sick. 2026. DD-MDN: Human Trajectory Forecasting with Diffusion-Based Dual Mixture Density Networks and Uncertainty Self-Calibration.arXiv preprint arXiv:2602.11214(2026)

work page arXiv 2026
[21]

Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models.Advances in neural information processing systems33 (2020), 6840–6851

2020
[22]

Jaewoo Jeong, Seohee Lee, Daehee Park, Giwon Lee, and Kuk-Jin Yoon. 2025. Multi-modal knowledge distillation-based human trajectory forecasting. InPro- ceedings of the Computer Vision and Pattern Recognition Conference. 24222–24233

2025
[23]

Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. 2023. Motiondiffuser: Controllable multi-agent motion prediction using diffusion. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 9644–9653

2023
[24]

Alon Lerner, Yiorgos Chrysanthou, and Dani Lischinski. 2007. Crowds by exam- ple. InComputer graphics forum, Vol. 26. Wiley Online Library, 655–664

2007
[25]

Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le
[26]

Flow matching for generative modeling.arXiv preprint arXiv:2210.02747 (2022)

work page internal anchor Pith review Pith/arXiv arXiv 2022
[27]

Jie Liu, Gongye Liu, Jiajun Liang, Yangguang Li, Jiaheng Liu, Xintao Wang, Pengfei Wan, Di Zhang, and Wanli Ouyang. 2025. Flow-grpo: Training flow matching models via online rl.arXiv preprint arXiv:2505.05470(2025)

work page internal anchor Pith review Pith/arXiv arXiv 2025
[28]

Jiayi Liu, Jiaming Zhou, Ke Ye, Kun-Yu Lin, Allan Wang, and Junwei Liang. 2025. EgoTraj-Bench: Towards Robust Trajectory Prediction Under Ego-view Noisy Observations.arXiv preprint arXiv:2510.00405(2025)

work page arXiv 2025
[29]

Xingchao Liu, Chengyue Gong, and Qiang Liu. 2022. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003(2022)

work page internal anchor Pith review Pith/arXiv arXiv 2022
[30]

Karttikeya Mangalam, Yang An, Harshayu Girase, and Jitendra Malik. 2021. From goals, waypoints & paths to long term human trajectory forecasting. In Proceedings of the IEEE/CVF international conference on computer vision. 15233– 15242

2021
[31]

Weibo Mao, Chenxin Xu, Qi Zhu, Siheng Chen, and Yanfeng Wang. 2023. Leapfrog diffusion model for stochastic trajectory prediction. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 5517–5526

2023
[32]

Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback.Advances in neural information processing systems35 (2022), 27730–27744

2022
[33]

Stefano Pellegrini, Andreas Ess, Konrad Schindler, and Luc Van Gool. 2009. You’ll never walk alone: Modeling social behavior for multi-target tracking. In2009 IEEE 12th international conference on computer vision. IEEE, 261–268

2009
[34]

Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, and Chuang. 2024. Movie gen: A cast of media foundation models.arXiv preprint arXiv:2410.13720(2024)

work page internal anchor Pith review Pith/arXiv arXiv 2024
[35]

Alexandre Robicquet, Amir Sadeghian, Alexandre Alahi, and Silvio Savarese
[36]

InEuropean conference on computer vision

Learning social etiquette: Human trajectory understanding in crowded scenes. InEuropean conference on computer vision. Springer, 549–565
[37]

Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. 2020. Trajectron++: Dynamically-feasible trajectory forecasting with heterogeneous data. InEuropean conference on computer vision. Springer, 683–700

2020
[38]

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

work page internal anchor Pith review Pith/arXiv arXiv 2024
[39]

Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. 2020. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456(2020)

work page internal anchor Pith review Pith/arXiv arXiv 2020
[40]

Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. 2020. Learning to summarize with human feedback.Advances in neural information processing systems33 (2020), 3008–3021

2020
[41]

Xiaohui Sun, Ruitong Xiao, Jianye Mo, Bowen Wu, Qun Yu, and Baoxun Wang
[42]

F5r-tts: Improving flow-matching based text-to-speech with group relative policy optimization.arXiv preprint arXiv:2504.02407(2025)

work page arXiv 2025
[43]

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need.Advances in neural information processing systems30 (2017)

2017
[44]

Liwen Xiao, Zhiyu Pan, Zhicheng Wang, Zhiguo Cao, and Wei Li. 2025. SRefiner: Soft-Braid Attention for Multi-Agent Trajectory Refinement. InProceedings of the IEEE/CVF International Conference on Computer Vision. 960–969

2025
[45]

Chenxin Xu, Maosen Li, Zhenyang Ni, Ya Zhang, and Siheng Chen. 2022. Group- net: Multiscale hypergraph neural networks for trajectory prediction with rela- tional reasoning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 6498–6507

2022
[46]

Chenxin Xu, Robby T Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. 2023. Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 1410–1420

2023
[47]

Pei Xu, Jean-Bernard Hayet, and Ioannis Karamouzas. 2022. Socialvae: Human trajectory prediction using timewise latents. InEuropean Conference on Computer Vision. Springer, 511–528

2022
[48]

Yi Xu, Lichen Wang, Yizhou Wang, and Yun Fu. 2022. Adaptive trajectory prediction via transferable gnn. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 6520–6531

2022
[49]

Shuchen Xue, Chongjian Ge, Shilong Zhang, Yichen Li, and Zhi-Ming Ma. 2025. Advantage weighted matching: Aligning rl with pretraining in diffusion models. arXiv preprint arXiv:2509.25050(2025)

work page arXiv 2025
[50]

Junjie Zheng, Chunbo Hao, Guobin Ma, Xiaoyu Zhang, Gongyu Chen, Chaofan Ding, Zihao Chen, and Lei Xie. 2025. YingMusic-Singer: Zero-shot Singing Voice Synthesis and Editing with Annotation-free Melody Guidance.arXiv preprint arXiv:2512.04779(2025)

work page arXiv 2025