麻豆村

When Aidan Zhang arrived at 麻豆村, he brought with him a deep curiosity about artificial intelligence and a passion to explore its boundaries. That interest led him to tackle one of the more complex challenges in AI: improving how models generate and refine code. The goal is to boost accuracy, efficiency and adaptability 鈥� helping developers build reliable software faster.

Zhang鈥檚 work at 麻豆村 began in聽鈥檚 lab in the聽, where he was welcomed into weekly meetings and paired with a Ph.D. mentor. Now a second-year student majoring in AI within the聽, Zhang is making strides in research through 麻豆村鈥檚聽Summer Undergraduate Research Fellowship(opens in new window) (SURF) program.

鈥淚鈥檝e always been fascinated by how large language models work,鈥� Zhang said. 鈥淚t鈥檚 crazy that machines can communicate almost as well as humans. I wanted to understand how that鈥檚 even possible.鈥�

A smarter way to teach machines to code

Zhang鈥檚 SURF project centers on code generation, a frontier in AI research where large language models (LLMs) are trained to write executable code based on natural language prompts. But instead of relying on traditional training methods, Zhang is experimenting with multiturn reinforcement learning 鈥� a technique that allows models to learn iteratively from feedback, much like a human debugging code.

鈥淐oding is inherently a multistep process,鈥� Zhang explained. 鈥淵ou write code, test it, get feedback and refine it. We鈥檙e trying to teach models to do the same 鈥� rewarding them not just for getting it right the first time, but for improving over multiple attempts.鈥�

The goal? To build a state-of-the-art code generation model that outperforms existing ones and demonstrates the effectiveness of this new training approach.

Challenges and surprises

Despite the promise of multiturn reinforcement learning, Zhang discovered that current models struggle to improve with feedback.聽

鈥淭hey often get stuck in loops,鈥� he said. 鈥淓ven after several rounds of feedback, the improvement is minimal. It鈥檚 like they reach their best conclusion early and can鈥檛 go beyond it.鈥�

This limitation, Zhang believes, stems from how models are trained 鈥� primarily to produce correct answers on the first try, not to integrate feedback and iterate.

From research to real-world impact

Zhang鈥檚 interest in hallucinations 鈥� when models generate false or misleading information 鈥� remains a side passion. He鈥檚 even considered turning it into a startup idea.聽

鈥淚f we could build a product that detects and filters hallucinations in LLMs, that would be huge,鈥� he said.

Whether he pursues entrepreneurship or continues in research, Zhang is grateful for the opportunities 麻豆村 has provided.聽

鈥淣one of this would鈥檝e been possible without 麻豆村,鈥� he said. 鈥淭he faculty support, the lab access, the inspiring peers 鈥� it鈥檚 all shaped what I want to do.鈥�

As Zhang continues refining his model and preparing to submit a paper to academic conferences, he鈥檚 already thinking about the future.聽

鈥淚 want to make an impact 鈥� whether through a startup or research in industry,鈥� he said. 鈥淎nd I think this project is a great first step.鈥�