PhD researcher · NUS Singapore

Abhijeet Sinha

I study how to make generative models explore. My work trains policies that find many good answers instead of collapsing onto one.

Reinforcement learning
Generative modeling
LLM post-training & alignment
Diversity & creativity
Interpretable AI

About

I am a PhD researcher in Machine Learning at the National University of Singapore, in the Artificial Scientific Intelligence Lab, supervised by Dr. Dianbo Liu. My research covers reinforcement learning, generative modeling, LLM post-training and alignment, diversity, creativity, and interpretable AI.

My current work studies outcome-level mode collapse in reinforcement learning and reward-proportional sampling. I develop methods that encourage diverse, high-reward generation while preserving alignment and task performance. I also work on interpretable generative systems, including reinforcement-learning agents that edit discrete latent representations.

Before NUS, I completed a dual Bachelor’s and Master’s degree in Biotechnology at IIT Madras, and worked on brain-inspired visual attention, neurological healthcare applications, document understanding, and enterprise AI systems.

Current focus

Fig. 1

160 samples from a policy trained for reward-proportional sampling. Every good outcome is found, in proportion to its reward.

Fig. 2

The same idea over a two-dimensional outcome space. Shading and contours show reward; each dot is a sampled outcome. Reward-proportional sampling covers every peak, in proportion to its reward.