QwenPaw Permalink
Agentic Harness — custom Agentic RL datasets and a QwenPaw Flash training and evaluation loop.
XINJI MAI / RESEARCH研究
Posts by CollectionPosts by Collection
portfolio
CuES Permalink
Curiosity-guided exploration grounds synthetic tasks in executable environments for agentic reinforcement learning.
AgentEvolver Permalink
Combines self-questioning, self-navigating, and self-attributing to help agents learn from their own experience.
Agentic RL Scaling Law Permalink
Asynchronous reinforcement learning for tool use and mathematical reasoning.
CMLM-ZhongJing Permalink
An open-source language model for Traditional Chinese Medicine, with a framework for model training and evaluation.
UMBEnet Permalink
Unifies multimodal emotion signals through a brain-inspired emotional central mechanism.
Hi-EF Permalink
Introduces a benchmark and multimodal framework for forecasting emotions in human interactions.
OUS Permalink
Connects scene context with facial features to support more context-aware emotion recognition.
publications
Coarse-To-Fine Key Frame Selection for Dynamic Facial Expression Recognition
Published in 2023 3rd International Conference on Networking Systems of AI (INSAI), 198–205, 2023
C2F-KFS selects expression peaks through coarse spatial and fine temporal feature extraction to improve dynamic facial expression recognition.
Recommended citation: Shaoqi Yan, Yan Wang, Xinji Mai, Qing Zhao, Boyang Wang, Wei He, Wenqiang Zhang. "Coarse-To-Fine Key Frame Selection for Dynamic Facial Expression Recognition." INSAI 2023. https://doi.org/10.1109/INSAI60116.2023.00043
Download Paper
Empower smart cities with sampling-wise dynamic facial expression recognition via frame-sequence contrastive learning
Published in Computer Communications, Volume 216, 130–139, 2024
SW-FSCL combines dual-stream dynamic sampling, key-frame weighting and frame-sequence contrastive learning to reduce neutral and noisy frames in expression recognition.
Recommended citation: Shaoqi Yan, Yan Wang, Xinji Mai, Qing Zhao, Wei Song, Jun Huang, Zeng Tao, Haoran Wang, Shuyong Gao, Wenqiang Zhang. "Empower smart cities with sampling-wise dynamic facial expression recognition via frame-sequence contrastive learning." Computer Communications 2024. https://doi.org/10.1016/j.comcom.2023.12.032
Download Paper
From Efficient Multimodal Models to World Models: A Survey
Published in arXiv preprint arXiv:2407.00118, 2024
Surveys efficient multimodal models, their training strategies, and the path toward world models.
Recommended citation: Xinji Mai, Zeng Tao, Junxiong Lin, Haoran Wang, Yang Chang, Yanlan Kang, Yan Wang, Wenqiang Zhang. "From Efficient Multimodal Models to World Models: A Survey." arXiv:2407.00118 (2024).
Download Paper
All Rivers Run into the Sea: Unified Modality Brain-Inspired Emotional Central Mechanism
Published in ACM International Conference on Multimedia (ACM MM), 2024
Unifies multimodal emotion signals through a brain-inspired emotional central mechanism.
Recommended citation: Xinji Mai, Junxiong Lin, Haoran Wang, Zeng Tao, Yan Wang, Shaoqi Yan, et al. "All Rivers Run into the Sea: Unified Modality Brain-Inspired Emotional Central Mechanism." ACM MM 2024, 632-641.
Download Paper
Observe finer to select better: Learning key frame extraction via semantic coherence for dynamic facial expression recognition in the wild
Published in Information Sciences, Volume 689, 121415, 2025
KFE-SC selects expressive video frames from coarse to fine by combining spatial semantic coherence with aligned short- and long-term temporal features.
Recommended citation: Shaoqi Yan, Yan Wang, Xinji Mai, Zeng Tao, Wei Song, Qing Zhao, Boyang Wang, Haoran Wang, Shuyong Gao, Wenqiang Zhang. "Observe finer to select better: Learning key frame extraction via semantic coherence for dynamic facial expression recognition in the wild." Information Sciences 2025. https://doi.org/10.1016/j.ins.2024.121415
Download Paper
OUS: Bridging Scene Context and Facial Features to Overcome the Rigid Cognitive Problem
Published in AAAI Conference on Artificial Intelligence (AAAI), 2025
Connects scene context with facial features to support more context-aware emotion recognition.
Recommended citation: Xinji Mai, Haoran Wang, Zeng Tao, Junxiong Lin, Shaoqi Yan, Yan Wang, et al. "OUS: Bridging Scene Context and Facial Features to Overcome the Rigid Cognitive Problem." AAAI 2025, 6054-6062.
Download Paper
D2SP: Dynamic Dual-Stage Purification Framework for Dual Noise Mitigation in Vision-based Affective Recognition
Published in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 19218–19229, 2025
D2SP combines coarse sample pruning with fine label correction to mitigate low-quality inputs and annotation noise in dynamic facial expression recognition.
Recommended citation: Haoran Wang, Xinji Mai, Zeng Tao, Xuan Tong, Junxiong Lin, Yan Wang, Jiawen Yu, Shaoqi Yan, Ziheng Zhou, Wenqiang Zhang. "D2SP: Dynamic Dual-Stage Purification Framework for Dual Noise Mitigation in Vision-based Affective Recognition." CVPR 2025. https://openaccess.thecvf.com/content/CVPR2025/html/Wang_D2SP_Dynamic_Dual-Stage_Purification_Framework_for_Dual_Noise_Mitigation_in_CVPR_2025_paper.html
Download Paper
Agentic RL Scaling Law: Spontaneous Code Execution for Mathematical Problem Solving
Published in Advances in Neural Information Processing Systems (NeurIPS), 2025
Studies how reinforcement learning teaches models to use code for mathematical reasoning, and how this ability scales.
Recommended citation: Xinji Mai, Haotian Xu, Xing W, Weinong Wang, Yingying Zhang, Wenqiang Zhang. "Agentic RL Scaling Law: Spontaneous Code Execution for Mathematical Problem Solving." NeurIPS 2025. arXiv:2505.07773.
Download Paper
When Small Models Team Up: A Weak-Expert Ensemble Surpassing LLMs for Automated Intellectual-Property Audits
Published in Preprint, 2025
A multi-agent audit framework combines specialized small models for automated intellectual-property audits.
Recommended citation: Xinji Mai, et al. "When Small Models Team Up: A Weak-Expert Ensemble Surpassing LLMs for Automated Intellectual-Property Audits." Preprint (2025).
Download Paper
AgentEvolver: Towards Efficient Self-Evolving Agent System
Published in arXiv preprint arXiv:2511.10395, 2025
Combines self-questioning, self-navigating, and self-attributing to help agents learn from their own experience.
Recommended citation: Yunpeng Zhai, Shuchang Tao, Cheng Chen, Anni Zou, Ziqian Chen, Qingxu Fu, Xinji Mai, et al. "AgentEvolver: Towards Efficient Self-Evolving Agent System." arXiv:2511.10395 (2025).
Download Paper
CuES: A Curiosity-driven and Environment-grounded Synthesis Framework for Agentic RL
Published in arXiv preprint arXiv:2512.01311, 2025
Curiosity-guided exploration grounds synthetic tasks in executable environments for agentic reinforcement learning.
Recommended citation: Xinji Mai, Yunpeng Zhai, Ziqian Chen, Cheng Chen, Anni Zou, Shuchang Tao, Zhaoyang Liu, Bolin Ding. "CuES: A Curiosity-driven and Environment-grounded Synthesis Framework for Agentic RL." arXiv:2512.01311 (2025).
Download Paper
Hi-EF: Benchmarking Emotion Forecasting in Human-interaction
Published in AAAI Conference on Artificial Intelligence (AAAI), 2026
Introduces a benchmark and multimodal framework for forecasting emotions in human interactions.
Recommended citation: Haoran Wang, Xinji Mai, Zeng Tao, Junxiong Lin, Xuan Tong, Ivy Pan, Shaoqi Yan, Yan Wang, Shuyong Gao. "Hi-EF: Benchmarking Emotion Forecasting in Human-interaction." AAAI 2026, 2110-2118.
Download Paper
Customizing Video Portraits via Identity-ActionDecoupling
Published in arXiv preprint arXiv:2606.22347, 2026
IaD separates identity and facial action features to generate identity-consistent videos with text-controlled expressions, without subject-specific fine-tuning.
Recommended citation: Junxiong Lin, Haoran Wang, Xinji Mai, Zeng Tao, Xuan Tong, Ivy Pan, Wenqiang Zhang. "Customizing Video Portraits via Identity-ActionDecoupling." arXiv. https://arxiv.org/abs/2606.22347
Download Paper
Interest Entanglement: The Hidden Barrier to Blind Super-Resolution Optimization
Published in arXiv preprint arXiv:2606.22353, 2026
SFR decouples fidelity and perceptual objectives through shared representations, with InfoSqueeze filtering redundant features for blind super-resolution.
Recommended citation: Junxiong Lin, Xinji Mai, Qianyu Guo, Haoran Wang, Zeng Tao, Xuan Tong, Ivy Pan, Wenqiang Zhang. "Interest Entanglement: The Hidden Barrier to Blind Super-Resolution Optimization." arXiv. https://arxiv.org/abs/2606.22353
Download Paper
Cowork-Gym: Synthesizing Interactive Work Environments at Scale
Published in arXiv preprint arXiv:2610.17670, 2026
Scalable synthesis of interactive work environments for agent post-training.
Recommended citation: Anni Zou, Cheng Chen, Jiaji Deng, Weikang Zhou, Xinji Mai, Yue Wu, Jianing Lin, Lipeng Xie, Lixiong Qin, Qingyue Zhao, Que Shen, Shuchang Tao, Xuchen Pan, Yaliang Li, Yunpeng Zhai, Yushuo Chen, Zhihai Wang, Zhuo Zhang, Zhaoyang Liu, Bolin Ding, Jingren Zhou. "Cowork-Gym: Synthesizing Interactive Work Environments at Scale." arXiv:2610.17670 (2026).
Download Paper