Projects
馃 Research Projects
馃敼 LLM Collaboration With Multi-Agent Reinforcement Learning
Keywords: Dec-POMDP 路 CTDE-inspired training 路 MAGRPO 路 Multi-turn LLM collaboration
Formalized multi-LLM collaboration as a cooperative Dec-POMDP: each agent observes partial, prompt-based information and outputs natural-language actions; the system/reward model provides joint rewards over multi-turn interactions
Introduced MAGRPO, extending GRPO to multi-agent, multi-turn settings by sampling groups of joint rollouts and using group-relative Monte-Carlo advantages to coordinate agents without training a large centralized value model
Evaluated on two collaboration domains:
Writing: 2 Qwen agents learn complementary roles (concise TLDR + detailed summary; background + method/experiments for arXiv-style expansion), improving structure/style coherence vs. prompt-level baselines
Coding: 2 Qwen coder agents generate a helper + main function; rewards include structure/syntax/tests and an explicit cooperation bonus when main correctly uses helper; introduced CoopHumanEval to reduce non-cooperative noise in HumanEval
Empirically, MAGRPO outperforms single-model and prompt-only multi-agent baselines in overall return and cooperation metrics, especially in multi-turn coding with external feedback signals
馃敼 YUSE: Symbolic Testing for Floating-Point Programs
Keywords: Symbolic Execution 路 Floating-Point Errors 路 Software Testing
- Developed techniques for detecting floating-point inconsistencies and corner-case bugs
- Focused on numerical instability, path explosion, and solver scalability
- Contributed to modeling, testing, and evaluation components of the framework
- This project strengthened my interest in robust and reliable systems
馃敼 Parallel Computing on RTEMS (Paor Project)
Keywords: Parallelism 路 Embedded Systems 路 Performance Optimization
- Worked on system-level parallel computation in real-time operating systems
- Designed and evaluated parallel strategies under strict timing constraints
- Gained experience bridging systems research with algorithmic design
You may find more and smaller-scale projects or in-class notes I took at https://github.com/zedyelllion?tab=repositories.