Oz T. Jang

He is a researcher in artificial intelligence and an amateur researcher in mathematics. he is working on Reasoning, AI safety and Multimodal. He is also a incoming visiting faculty, National University of Singapore. He was a visiting researcher in Carnegie Mellon University.

He always supports Slow Science

Github  /  Scholar  / 

If you would like to join his group in any other capacity, please fill this form and then please send him a short email note without any documents.
Selected Publications
AutoRL-LRM: Automated RL Algorithm Optimization for Large Reasoning Models

Oz T. Jang

[paper] (Under reviewing) [code]

An autonomous research agent that discovers, implements, and evaluates RL algorithms to improve mathematical reasoning, guided by metrics.

2025

VRHF: RLVR from Human Feedback

Oz T. Jang

[paper] (Writing)

Propose combining RLHF with RLVR for improving Large Reasoning Model

2025

Double Zero: Self-Evolving Reasoning with Zero Data, Zero RL

Oz T. Jang

[paper] (Writing)

Propose Double Zero Self-Evolving Large Reasoning Models without data, without SFT

2025

SFT-hybrid-RFT Large Reasoning Model

Oz T. Jang

[paper] (Writing)

Propose SFT-hybrid-RFT Large Reasoning Model

2025

LookaheadBio: Bilevel Optimization with Lookahead Moving Averages for LLM Data Reweighting

Ruijie Xie, Chaoyue Zhao, Oz T. Jang

[paper] (Under reviewing)

the first bilevel optimization algorithm that incorporates Lookahead's slow-fast weight mechanism for both model parameters and data weights, achieving comprehensive variance reduction.

2025

Lookahead-LSTM Optimizer: A Meta-Learning K-steps Method

Oz T. Jang, Teng Yang, Xiaozhu Hu, Zi Yang, Chifong Wong

[paper] [code]

Propose a Meta-learning optimization method named Lookahead-LSTM for improving generalization and data transferability.

2020 Spring

Selected Projects
multi Multi-view Learning for Vision-and-Language Planning

Oz T. Jang, Wei Yuan

Generating multi-view videos as a world state representation,combining VLMs with text-to-video models to obtain the ability of long-horizon decision making

2023 Winter

multi Multimodal Perception Fusion for Robotic Manipulation

Oz T. Jang, Zhigang Li, Bowen Fu

Merge object 6-DoF information and improve the performance of object localization for robotic manipulation

2019 Fall

Pro bono office hour

Starting January 2025, He has decided to commit 1~2 hours every week to provide guidance, suggestions, and/or mentorships for students from underrepresented groups or whoever is in need.

Please fill in this form if you are interested.


This template is a modification to Jon Barron's website template