← Nanshu Wang

Nanshu Wang

LinkedIn GitHub

Staff Software Engineer at Meta (SuperIntelligence Lab) with 7+ years shipping AI products at scale. Deep expertise in LLM post-training: RLHF, curriculum reinforcement learning, SFT, and DPO across LLM capability on multimodal perception, instruction following, agentic systems with tool use and RAG. Led 50+ person teams across AI Glass (Ray-Ban Meta), Llama post-training, and Meta AI, with products reaching hundreds of millions of users.

Experience

Meta — Staff Software Engineer
SuperIntelligence Lab · Wearable AI
Tech lead for LiveAI on Ray-Ban Meta Glasses — real-time multimodal perception over ego-centric video and audio, with VLM post-training via RLHF, curriculum reinforcement learning, and proactive reasoning. Founding tech lead for the Meta AI platform: agents, plugins, tool use, and RAG powering Meta AI's 2023 debut and 2024 Llama 3 upgrade.
Meta — Senior Software Engineer
AI Assistant · Reality Labs
Tech lead for on-device text assist (Smart Compose, Quick Reply) on Quest/Oculus devices using on-device generative LMs, federated learning, and differential privacy.
Meta — Software Engineer
AI Assistant · Messenger
Built Meta's first LM product (Smart Compose) and first on-device compressed model for Messenger. Scaled Smart Compose to billions of MAU on Facebook News Feed.

Education

Carnegie Mellon University
M.S. in Software Engineering
Chinese Academy of Sciences
M.S. in Computer Science
Renmin University of China
B.S. in Human Resources Management

Fun

Been coding since high school — and won First Prize at the National Olympiad in Informatics (NOIP) twice.

Work / Media

Live AI Launch: Ray-Ban Meta Glasses Add Live AI, Live Translation & Shazam Support Meta · 2024
The Llama 4 Herd: The Beginning of a New Era of Natively Multimodal AI Meta · 2025
Meet Your New Assistant: Meta AI, Built With Llama 3 Meta · 2024
Introducing New AI Experiences Across Our Family of Apps and Devices Meta · 2023
AI at Meta — Smart Compose Announcement Meta · 2022
Speechless? Here's How AI Learns to Finish Your Sentences Meta Tech Blog · 2021
How Does AI Learn to Finish Your Sentences? (Smart Compose) Meta Tech Video · 2021
Open-Sourcing PyText for Faster NLP Development Meta Engineering Blog · 2018
M Now Offers Suggestions to Make Your Messenger Experience More Useful Meta · 2017

Open Source

TorchGlow
ML compiler and execution engine for hardware accelerators
Multi-layer LSTM support.
PyText
Deep-learning NLP modeling framework built on PyTorch
ONNX support and contextualized model implementation.
DeepSpeech
End-to-end speech recognition
Data refinement pipeline for loading speech corpus.

Publications & Patents

He, Y., Li, W., Zhang, H., Li, S., …, Wang, N., et al. Rubric-Based Benchmarking and Reinforcement Learning for Advancing LLM Instruction Following. ACL 2026.

Qin, C., Zhou, W., Sankararaman, K.A., Wang, N., et al. Learning Auxiliary Tasks Improves Reference-Free Hallucination Detection in Open-Domain Long-Form Generation. ACL 2025.

Sun, Y., Yin, X., Jiang, J., Sekar, V., Lin, F., Wang, N., et al. CS2P: Improving Video Bitrate Selection and Adaptation with Data-Driven Throughput Prediction. ACM SIGCOMM 2016.

Sun, Y., Jiang, J., Sekar, V., Zhang, H., Lin, F., & Wang, N. Using Video-Based Measurements to Generate a Real-Time Network Traffic Map. ACM HotNets 2014.

Gill, P., Liu, H., Yang, W., Malik, K., Wang, N., & Reiss, D. Methods, mediums, and systems for providing a model for an end-user device. U.S. Patent 11,501,081. 2022. Patent

Gill, P., Liu, H., Yang, W., Malik, K., Wang, N., & Reiss, D. Methods, mediums, and systems for training a model. U.S. Patent 11,455,555. 2022. Patent

Botros, F., Wang, N., Wang, F., et al. Voice-based Auto-Completions and Auto-Responses for Assistant Systems. U.S. Patent Application 17/120,013. 2022. Patent

Gill, P., Liu, H., Yang, W., Malik, K., Wang, N., & Reiss, D. Methods, mediums, and systems for representing a model in a memory of device. U.S. Patent 11,227,122. 2022. Patent