MAI-Thinking-1: Building a Hill-Climbing Machine
A technical report on training reasoning models from scratch through scalable pre-training and sustained reinforcement learning.
Research & engineering
Richie Feng
Member of Technical Staff Microsoft AI Mountain View, California
I work on improving language models across a broad range of capabilities while maintaining overall quality and reliability. My focus is on post-training, reinforcement learning, and model behavior, including learning from feedback and adapting to people across languages.
Previously, I worked at Google DeepMind on Gemini’s multilingual capabilities, personalization, safety, and Gemini Live. Before that, I spent five years at Twilio building conversational AI and machine learning systems.
Research
interests
Post-training & alignment
Reinforcement learning, reward modeling, and learning from human preferences.
Model behavior & evaluation
Reasoning efficiency, factuality, conversational evaluation, and personalization.
Multilingual AI
Models that communicate reliably across languages, cultures, and contexts.
Selected
publications
CoReflect: A Reflective Co-Evolution Framework for Improving Conversational Evaluation
An adaptive evaluation framework where dialogue simulation and reflective rubric refinement improve together.
Gemini: A Family of Highly Capable Multimodal Models
A family of multimodal models built to understand and reason across text, images, audio, and video.
Experience
Microsoft AI
Mar 2026 — present
Member of Technical Staff
Foundation model post-training, including MAI-Thinking-1. Develop reinforcement learning and reward modeling methods to improve model behavior and factuality, and lead multilingual post-training and evaluation.
Google DeepMind
Nov 2021 — Mar 2026
Software Engineer, Machine Learning
Worked on Gemini’s expansion to 43 languages, multilingual safety, and Gemini Live. Led an early prototype for personalization for learning, and launched natural language generation systems for Assistant and Search.
Twilio
Dec 2016 — Nov 2021
Machine Learning Engineer
Founding team member of Twilio Autopilot, building intent classification and entity extraction. Improved SMS send-time optimization to increase engagement.
Education
University of California, Berkeley
Master’s degree in Statistics
2015 — 2016
University of Illinois Urbana-Champaign
Bachelor’s degree in Mathematics · Computer Science
2011 — 2015