Open Source Agent Team Work Zone A persistent management layer for long-lived Claude Code agent teams, designed to preserve working state across compaction and restarts. Research How to Loop MoE Foil flattens expert layers and unties attention to improve looped MoE models at fixed parameters and compute. Path-Lock Expert Architecture-level separation of think and no-think modes through deterministic control-token routing. Demystifying Hybrid Thinking An empirical study of whether hybrid-thinking language models truly separate think and no-think behavior. Efficient Protein Language Models A taxonomy and practical survey of data, architecture, training, and inference efficiency in protein language models. Mid-Think Training-free intermediate-budget reasoning through token-level triggers that induce or suppress thinking. Enhancing Fighting-Game Enjoyment with DRL and LLM Agents A two-tier system that combines diverse deep-reinforcement-learning opponents with an LLM hyper-agent that adapts opponent selection to player data and feedback. Earlier Work Simulation for Sensorimotor Control Use Reinforcement Learning method to demonstrate human Center Nervous System's adaptability Sparse Logistic Regression with Majorization Optimization for sparse logistic regression with majorization methods. Alien Shooter-like Game Development Developed a Alien Shooter-like video game using the Unity3D engine. Motion-Controlled Parkour Game Developed a gesture-controlled parkour video game using the Unity3D engine and CNN model.
