Lili Chen

I'm a Ph.D. student in the Machine Learning Department at Carnegie Mellon University, advised by Deepak Pathak. My research is graciously supported by the Department of Defense NDSEG Fellowship.

Previously, I received my B.A. in Computer Science from UC Berkeley. I was an undergraduate researcher at the Robot Learning Lab, where I was advised by Pieter Abbeel and Kimin Lee.

Email  /  Google Scholar  /  GitHub  /  LinkedIn  /  Twitter

I'm broadly interested in reinforcement learning, self-supervised learning, and LLM reasoning.

We explore the possibility that LLMs could improve without external data, via asymmetric self-play: by generating their own questions as a proposer and answering them as a solver.

We present an unsupervised reinforcement learning method that improves LLM reasoning performance by using the model's own confidence as a reward.

We present a language-conditioned diffusion model which can learn visuomotor policies from language-annotated play data.

We train a visual affordance model on human videos to estimate how a human is likely to interact with objects and deploy this model in robotic control tasks.

We propose to replace traditional offline RL algorithms with a simple transformer model trained on sequences of returns, states, and actions with an autoregressive prediction loss.

We tackle exploration for high-dimensional observation spaces using a k-NN state entropy estimator in the low-dimensional representation space of a randomly intialized CNN.

We present a compute- and memory-efficient modification of off-policy visual RL methods by freezing lower layers of CNN encoders and storing low-dimensional embeddings.

We demonstrate usage of RBMs to solve NP-Hard problems efficiently by mapping the RBM onto a reconfigurable FPGA.

I hope to improve the accessibility of computer science education, at all levels.

Website template from here.