LLM | Zichen Zhang

<em>Simpler is Better:</em> Finding the Best Reward Function in Long Chain-of-Thought Reinforcement Learning for Small Language Models

Luning Wang*, Zichen Zhang*, Junkuan Liu*.

We study three types of reward functions — normal, cosine, and dynamic — for long chain-of-thought reinforcement learning in Small Language Models, and find that the simple normal reward consistently outperforms more complex designs, suggesting that simpler rewards are good enough for eliciting reasoning in smaller models.

<em>Simpler is Better:</em> Finding the Best Reward Function in Long Chain-of-Thought Reinforcement Learning for Small Language Models

Babysitting a Small Language Model through One-Step Tree-of-Thoughts Knowledge Distillation

Anurag Renduchintala*, Adi Mahesh*, Zichen Zhang*, Zimo Si*, Shangjun Meng*, Samuel Fang*.

We introduce the One-Step Tree-of-Thoughts framework, a simplified prompting method that distills multi-step reasoning into a single structured prompt, and demonstrates how knowledge distillation can transfer this reasoning capability from Large Language Models to Small Language Models with much less parameters, enabling significant improvements reasoning performance, beating GPT-4o and GPT-4, as shown on the model's performance on Game of 24.

Babysitting a Small Language Model through One-Step Tree-of-Thoughts Knowledge Distillation

GenHint: An AI Coding Assistant that Teaches How to Code

🏆 Honorable Mention Best Developer Tool 🏆. MHacks 17 Hackathon, 2024

Zichen Zhang*, Peihao Li*, Tongyuan Miao*, Yuchen Huang*.

Unlike conventional AI coding assistants like GitHub Copilot, GenHint does not give you codes directly. Instead, it generates code templates with "TODO" comments and explains each subproblem for you. It's powered by Llama 3 70B hosted on Groq. Published as a VS Code Extension and easy to download, it will be a great companion for your coding workflow.

Zichen "Charlie" Zhang, Peihao Li, Tongyuan Miao, Yuchen Huang

Sep 29, 2024

GenHint: An AI Coding Assistant that Teaches How to Code

Collage: Pinterest for College Advising

🌟 700+ active users since launch 🌟

Max Feldman*, Alex Wang*, Zichen Zhang*, Jaden Sun*, Kali Francisco, Nate Bennett, Leo Choi, Tanishka Nalawade, Joe Tang, Asher Katz, Eden McCullough, Dasha Skalitzky.

Accessible via your Google Education account, Collage is an AI-powered educational discovery engine for the University of Michigan that personalizes academic advising and scheduling and connects UMich students with their classes, advisors, and peers. Collage features a Pinterest-like course catalog tailored to individual academic interests and career goals, an AI advisor to assist with scheduling decisions, and a social networking system for sharing schedules and connecting with peers.

Max Feldman, Alex Wang, Zichen "Charlie" Zhang, Jaden Sun

Jul 2, 2024