DiJia (Andy) Su — AI Research Scientist

Reasoning, learning,
and intelligent agents.

I’m Andy, an AI research scientist at Meta Superintelligence Labs, working on Muse Spark and leading an effort in agentic LLM post-training and reinforcement learning.

My research spans language model reasoning, reinforcement learning, and planning. I received my PhD from Princeton University.

Background

At Meta, I have worked with Ruslan Salakhutdinov and Yuandong Tian.

At Princeton, I was advised by H. Vincent Poor and John Mulvey, and worked closely with Jason Lee and Chi Jin. During my time at Google, I worked with Dale Schuurmans and Craig Boutilier.

Selected research

All publications ↗

ICML2025

Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning

DiJia Andy Su, Hanlin Zhu *, Yingchen Xu*, Jiantao Jiao, Yuandong Tian^, Qinqing Zheng^.

COLM2025

Training Large Language Models to Reason in a Continuous Latent Space

Shibo Hao, Sainbayar Sukhbaatar, DiJia Andy Su, Xian Li, Zhiting Hu, Jason Weston, Yuandong Tian

ICLR2025

Dualformer: Controllable Fast and Slow Thinking by Learning with Randomized Reasoning Traces

DiJia Andy Su, Sainbayar Sukhbaatar, Michael Rabbat, Yuandong Tian, Qinqing Zheng

ICLR2026

SPG: Sandwiched Policy Gradient for Masked Diffusion Language Models

Chenyu Wang, Paria Rashidinejad, DiJia Su, Song Jiang, Sid Wang, Siyan Zhao, Cai Zhou, Shannon Zejiang Shen, Feiyu Chen, Tommi Jaakkola, Yuandong Tian, Bo Liu

ICML2026

dTRPO: Trajectory Reduction in Policy Optimization of Diffusion Large Language Models

Wenxuan Zhang, Lemeng Wu, Changsheng Zhao, Ernie Chang, Mingchen Zhuge, Zechun Liu, DiJia Su, Hanxian Huang, Jun Chen, Chong Zhou, Raghuraman Krishnamoorthi, Vikas Chandra, Mohamed Elhoseiny, Wei Wen

Preprint2025

Galore 2: Large Scale Low Rank Memory Efficient LLM Pretraining on 500B tokens

DiJia Andy Su, Andrew Gu, Jane Xu, Yuandong Tian, Jiawei Zhao

COLM2024

Beyond A*: Better Planning with Transformers via Search Dynamics Bootstrapping

Lucas Lehnert, Sainbayar Sukhbaatar, DiJia Andy Su, Qinqing Zheng, Paul Mcvay, Michael Rabbat, Yuandong Tian

More publications & preprints

Preprint2025

The Path Not Taken: RLVR Provably Learns Off the Principals

Hanqing Zhu, Zhenyu Zhang, Hanxian Huang, DiJia Su, Zechun Liu, Jiawei Zhao, Igor Fedorov, Hamed Pirsiavash, Zhizhou Sha, Jinwon Lee, David Z Pan, Zhangyang Wang, Yuandong Tian, Kai Sheng Tai

ICRA2022

Narrowing the Coordinate-Frame Gap in Behavior Prediction Models: Distillation for efficient and accurate scene-centric motion forecasting

DiJia Andy Su, Bertrand Douillard, Rami Al-Rfou, Cheol Park, Benjamin Sapp

ICASSP2022

Competitive Multi-Agent Reinforcement Learning with Self-Supervised Representation

DiJia Andy Su, Jason D Lee, John M Mulvey, H Vincent Poor

Preprint2022

A Deep Reinforcement Learning Approach for Finding Non-Exploitable Strategies in Two-Player Atari Games

DiJia Andy Su*, Zihan Ding*, Qinghua Liu, Chi Jin

Preprint

MUSBO: Model-based uncertainty regularized and sample efficient batch optimization for deployment constrained reinforcement learning

DiJia Andy Su, Jason D Lee, John M Mulvey, H Vincent Poor

IEEE TAI

ICCSE

Autonomous robot navigation with self-learning for collision avoidance with randomly moving obstacles

Yunfei Zhang, Clarence W de Silva, DiJia Andy Su, Youtai Xue

Best paper award

ICML2020

ConQUR: Mitigating Delusional Bias in Deep Q-Learning

DiJia Andy Su, Jayden Ooi, Tyler Lu, Dale Schuurmans, Craig Boutilier

Manuscript

Machine Learning Optimization with Crypto Currency

DiJia Andy Su et al.

Patents

Mitigating delusional bias in deep q-learning for robotic and/or other agents

Tyler Lu, Jayden Ooi, Craig Boutilier, Dale Schuurmans, DiJia Andy Su
US20220101111A1

Training agent trajectory prediction neural networks using distillation

Bertrand Douillard, DiJia Andy Su
US20230082079A1

Reasoning, learning,
and intelligent agents.

Background

Selected research

Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning

Training Large Language Models to Reason in a Continuous Latent Space

Dualformer: Controllable Fast and Slow Thinking by Learning with Randomized Reasoning Traces

SPG: Sandwiched Policy Gradient for Masked Diffusion Language Models

dTRPO: Trajectory Reduction in Policy Optimization of Diffusion Large Language Models

Galore 2: Large Scale Low Rank Memory Efficient LLM Pretraining on 500B tokens

Beyond A*: Better Planning with Transformers via Search Dynamics Bootstrapping

The Path Not Taken: RLVR Provably Learns Off the Principals

Narrowing the Coordinate-Frame Gap in Behavior Prediction Models: Distillation for efficient and accurate scene-centric motion forecasting

Competitive Multi-Agent Reinforcement Learning with Self-Supervised Representation

A Deep Reinforcement Learning Approach for Finding Non-Exploitable Strategies in Two-Player Atari Games

MUSBO: Model-based uncertainty regularized and sample efficient batch optimization for deployment constrained reinforcement learning

Optimizing Multi-Document Summarization by Blending Reinforcement Learning Policies

Improving Portfolio Performance via Natural Language Processing Methods

Secrecy capacity and wireless energy harvesting in amplify-and-forward relay networks

Autonomous robot navigation with self-learning for collision avoidance with randomly moving obstacles

ConQUR: Mitigating Delusional Bias in Deep Q-Learning

Machine Learning Optimization with Crypto Currency

Patents

Mitigating delusional bias in deep q-learning for robotic and/or other agents

Training agent trajectory prediction neural networks using distillation

Reasoning, learning,and intelligent agents.

Background

Selected research

Machine Learning Optimization with Crypto Currency

Patents

Reasoning, learning,
and intelligent agents.