Cowork-Gym: Synthesizing Interactive Work Environments at Scale
Synthesizes 64.9K interactive work environments through a build–execute–revise loop and stateful user simulation for agent post-training.通过构建、执行与修订闭环及有状态用户模拟，合成 6.49 万个交互式工作环境，用于智能体后训练。
XINJI MAI / PUBLICATIONS研究成果
Publications论文
Agentic reinforcement learning, self-evolving agents, and multimodal intelligence.智能体强化学习、自进化智能体与多模态智能。
Interest Entanglement: The Hidden Barrier to Blind Super-Resolution Optimization
SFR decouples fidelity and perceptual objectives through shared representations, with InfoSqueeze filtering redundant features for blind super-resolution.SFR 通过共享表征解耦保真度与感知质量优化目标，并以 InfoSqueeze 过滤冗余特征，改善盲超分辨率的目标平衡。
Customizing Video Portraits via Identity-ActionDecoupling
IaD separates identity and facial action features to generate identity-consistent videos with text-controlled expressions, without subject-specific fine-tuning.IaD 分离身份与面部动作特征，在无需针对特定人物微调的情况下，生成身份一致且表情受文本控制的人像视频。
Hi-EF: Benchmarking Emotion Forecasting in Human-interaction
Introduces a benchmark and multimodal framework for forecasting emotions in human interactions.构建人际互动中的情绪预测基准，并提出多模态预测框架。
CuES: A Curiosity-driven and Environment-grounded Synthesis Framework for Agentic RL
Curiosity-guided exploration grounds synthetic tasks in executable environments for agentic reinforcement learning.以好奇心引导环境探索，在可执行的环境中合成智能体强化学习任务。
AgentEvolver: Towards Efficient Self-Evolving Agent System
Combines self-questioning, self-navigating, and self-attributing to help agents learn from their own experience.结合自主提问、自主探索与自主归因，让智能体从自身经验中持续学习。
When Small Models Team Up: A Weak-Expert Ensemble Surpassing LLMs for Automated Intellectual-Property Audits
A multi-agent audit framework combines specialized small models for automated intellectual-property audits.将多个专门化小模型组合为多智能体审核框架，用于知识产权自动审核。
Agentic RL Scaling Law: Spontaneous Code Execution for Mathematical Problem Solving
Studies how reinforcement learning teaches models to use code for mathematical reasoning, and how this ability scales.研究强化学习如何让模型自主调用代码进行数学推理，以及这种能力的规模化规律。
D2SP: Dynamic Dual-Stage Purification Framework for Dual Noise Mitigation in Vision-based Affective Recognition
D2SP combines coarse sample pruning with fine label correction to mitigate low-quality inputs and annotation noise in dynamic facial expression recognition.D2SP 结合粗粒度样本剪枝与细粒度标签修正，缓解动态表情识别中的低质量输入和标注噪声。
OUS: Bridging Scene Context and Facial Features to Overcome the Rigid Cognitive Problem
Connects scene context with facial features to support more context-aware emotion recognition.联结场景上下文与面部特征，让情绪识别更充分地理解情境。
Observe finer to select better: Learning key frame extraction via semantic coherence for dynamic facial expression recognition in the wild
KFE-SC selects expressive video frames from coarse to fine by combining spatial semantic coherence with aligned short- and long-term temporal features.KFE-SC 结合空间语义一致性与长短期时序特征对齐，以由粗到细的方式筛选表情关键帧。
All Rivers Run into the Sea: Unified Modality Brain-Inspired Emotional Central Mechanism
Unifies multimodal emotion signals through a brain-inspired emotional central mechanism.通过脑启发的情感中枢机制，统一建模多模态情绪信号。
From Efficient Multimodal Models to World Models: A Survey
Surveys efficient multimodal models, their training strategies, and the path toward world models.梳理高效多模态模型、训练策略及其向世界模型演进的路径。
Empower smart cities with sampling-wise dynamic facial expression recognition via frame-sequence contrastive learning
SW-FSCL combines dual-stream dynamic sampling, key-frame weighting and frame-sequence contrastive learning to reduce neutral and noisy frames in expression recognition.SW-FSCL 结合双流动态采样、关键帧加权和帧序列对比学习，减轻中性帧与噪声帧对表情识别的影响。
Coarse-To-Fine Key Frame Selection for Dynamic Facial Expression Recognition
C2F-KFS selects expression peaks through coarse spatial and fine temporal feature extraction to improve dynamic facial expression recognition.C2F-KFS 通过粗粒度空间特征与细粒度时间特征提取筛选表情峰值帧，改进动态表情识别。