Jayden Teoh

I am currently a student researcher at Google DeepMind working with Vaishnavh Nagarajan. I will be joining MIT CSAIL as a PhD student in Fall 2026.

Previously, I had a fun stint at Microsoft Research New York under John Langford where I worked on Next-Latent Prediction Transformers, a method for learning compact world models.

I gratuated with a Computer Science degree from Singapore Management University.

Email  /  CV  /  LinkedIn  /  Scholar  /  X  /  Github

Mentors I'm Grateful For

• Pradeep Varakantham (Singapore Management University) for taking a chance on me as an undergraduate and exposing me to the world of ML research. I'm forever grateful for his mentorship and support.


• John Langford (Microsoft Research) for providing me the space to grow as a researcher. His belief in my potential has been a driving force behind my decision to pursue a research career and a PhD.

Research

Jayden's research interests lie in the intersection of representation learning and self-improving agents. Jayden has a track record of publishing papers in top-tier conferences including NeurIPS, ICLR, and ICML. ^*

We introduce Next-Latent Prediction (NextLat), which extends standard next-token training with self-supervised predictions in the latent space.

We introduce a decoding algorithm for Masked Diffusion Models (MDMs) that replaces greedy local-certainty heuristics with a principled information-gain objective, yielding more robust generation across math, code, and creative tasks.

We introduce a LLM-guided evolutionary framework for discovering new reward functions to stabilize Adversarial Imitation Learning.

We formalize the concept of generalization in Multi-Objective Reinforcement Learning (MORL) and contribute a novel benchmark to facilitate future studies in this area.

We evaluate the effectiveness of capability elicitation techniques by intentionally training language models with hidden capabilities that are revealed by a password.

By integrating both regret and novelty as complementary objectives for unsupervised environment design, our CENIE framework facilitates effective exploration across the state-action space while progressively increasing curriculum complexity.

GENIE quantifies environment novelty within the Unsupervised Environment Design (UED) paradigm by using Gaussian Mixture Models.

Miscellanea