Dexter Ju

AI RESEARCH & ENGINEERING

Dexter Ju

AI safety.
RL, reasoning & agents.

I’m a Member of Technical Staff at Microsoft AI, working on safety for foundation models. Previously, I worked on reinforcement learning for reasoning at Meta Superintelligence Labs (MSL) and on conversational AI at FAIR (Facebook AI Research).

You’ll find my publications under Da Ju.

Explore my work Google Scholar
01

Safety & alignment

Foundation-model safety at Microsoft AI.

02

RL & reasoning

Reinforcement learning and reasoning for Llama 4.

03

Agents & dialogue

Travel-planning agents and conversational learning.

SELECTED WORK

Models, agents,
and the work behind them.

My work spans foundation-model safety, reinforcement learning, agentic planning, and conversational AI.

2026 / Microsoft AI · Reasoning model

MAI-Thinking-1

Microsoft AI’s reasoning model, trained from scratch for math, coding, and tool use.

My focus at Microsoft AI · Foundation-model safety

2025 / Meta AI · Model family

Llama 4

Meta’s natively multimodal model family. I worked on reinforcement learning and reasoning for Llama 4.

Model development · RL & reasoning

Model code

2024 / EMNLP · System Demonstrations

To the Globe

I worked on an agentic travel-planning system that turns natural-language requests into plans with guarantees.

Agentic research · Language → plans

Code

2022 / NeurIPS

Staircase Attention

An attention architecture that brings recurrent processing to sequences. We explored looped-transformer ideas in our 2021 preprint.

Oral presentation · Preprint 2021 · Attention → recurrence

Code

2022 / arXiv

BlenderBot 3

I’m a joint first author on BlenderBot 3, a deployed chatbot designed to learn from conversations and engage responsibly.

Joint first author · Conversation → learning

Code & models
Browse all 22 publications & reports

THE RESEARCH LIBRARY

Publications & reports22

On Google Scholar

Peer-reviewed papers, preprints, and team reports. Checked against Google Scholar on September 5, 2026; duplicate versions are combined.

Years refer to the published version where available; earlier preprints are noted. Collective reports retain their team authorship. Code and dataset links point to verified releases. Some repositories are archived; their licenses and setup notes apply.

A LITTLE BACKGROUND

From dialogue systems
to reasoning and safety.

My research spans deployed dialogue systems such as BlenderBot 3, recurrent architectures such as Staircase Attention, and agentic travel planning with To the Globe. More recently, I worked on RL and reasoning for Llama 4; I now focus on foundation-model safety at Microsoft AI.

At FAIR, I worked with teams led by Gabriel Synnaeve, Jason Weston, and Yuandong Tian. Before that, I studied machine learning, human–computer interaction, and data science in Paris.

More on LinkedIn
NOW

Microsoft AI

Member of Technical Staff · AI safety

PREVIOUSLY

Meta Superintelligence Labs (MSL)

Research Scientist · LLaMA reasoning

Reinforcement learning for reasoning models.

EARLIER

Facebook AI Research

Research Engineer · FAIR

2017

Master’s studies in Paris

Télécom ParisTech · Machine Learning & HCI

UPMC / Paris VI · Data Science

2015

Southeast University

Information Science & Engineering

Undergraduate thesis

“High-resolution LED Headlamps Testing System with Web GUI,” completed at L-LAB in Lippstadt, Germany. The thesis remains the intellectual property of HELLA.

Read the thesis (PDF)

A PERSONAL PROJECT · OUTSIDE THE LAB

A little less searching.
A lot more New York.

NYC What To Do brings together events, exhibitions, dining, and interesting things around the city—with filters, a map, and a place to plan your day.

Explore the NYC guide

SAY HELLO

Good ideas start
with a conversation.

Connect on LinkedIn