about me
Han Fang is an AI Research Scientist at Meta’s Superintelligence Labs, working at the frontier of RL, agents, and self-improvement. His recent work includes agentic tool-use capabilities in Muse Spark, which achieves state-of-the-art performance on MCP-Atlas. In 2023–2024, Han led production post-training for Llama 2 and Llama 3, helping launch Meta AI and scale it to 1 billion MAU — driving integrated training runs, core capabilities, tool use and orchestration, and the user-data flywheel.
Han holds a PhD in Applied Mathematics and has published in top-tier venues with 10K+ citations. He is a recipient of the President’s Award to Distinguished Doctoral Students, the Woo-Jong Kim Dissertation Award, and the Excellence in Research Award.
Google Scholar / CV / Linkedin / Twitter
blog
The Central Dogma of Artificial Intelligence
February 2026
Every mature science has its central dogma. Biology has DNA → RNA → Protein. What is ours? Intelligence is the compression of experience into generalization.
The RL Environment Field Guide
January 2026
A practical guide to RL environments using Pokemon Red as a case study. Covers the agent-environment loop, observation spaces, reward design, and credit assignment.
Post-training 101: A Hitchhiker's Guide
September 2025
A comprehensive guide to post-training techniques for LLMs, covering supervised fine-tuning, RLHF, reward models, and practical implementation details.
news
Launched Agentic Tool Use in Muse Spark
Excited to introduce Muse Spark 🥑, the first in the Muse family of models from Meta Superintelligence Labs. Natively multimodal reasoning with tool-use, visual chain of thought, and multi-agent orchestration. Drove agentic tool use capabilities to SoTA.
2026
Meta AI reached 1 billion MAU
Improved Meta AI's multilinguality, enabled roll-out to 12 languages and 40+ countries. Blog · News
2025
Launched voice mode and photo editing in Meta AI
Launched updated Llama 3 model for voice mode. Improved Planner for photo editing with multimodal inputs. Blog
Connect 2024
Launched Llama 3 on Meta AI
Developed Meta AI's online RL with Mixture of Judges, improving reasoning, instructions following, safety. Blog
2024
Launched Meta AI with Llama 2
Debut Meta AI into Family of Apps. Created Orchestrator for tool use. My talk at Connect
2023
Meta AI Few-Shot Learner (FSL)
Developed FSL for detecting new forms of harmful content across 100+ languages. Blog
2021
Training AI to detect hate speech
Built RL-based framework to E2E optimize hate speech classifiers. Blog
2020
