Gaurav Ghosal

I'm a second year Ph.D. student at the Carnegie Mellon University Machine Learning Department advised by Professor Aditi Raghunathan. My interests lie broadly in improving the reliability of foundation and large language models. Some areas I am especially focused on include factuality: ensuring that LLMs learn and correctly apply factual knowledge in downstream tasks, model updating: designing LLMs that can be efficiently and reliably updated with new information without losing their general capabilities, and understanding training dynamics: characterizing how LLM failure modes arise during training and developing principled methods to mitigate them. Prior to CMU, I was an undergraduate student and researcher at UC Berkeley where I worked on human-AI interaction with Anca Dragan.

Email  /  CV  /  Scholar  /  Twitter  /  Github

Research

Most of my current research focuses on improving reliability of large language models. I pay special attention to problems in improving LLM factuality, enabling efficient updating and unlearning, and studying LLM training dynamics to better understand the origin of their failure modes. In my undergraduate, I published papers on learning from human feedback and robustness to spurious correlations.

Large language models are exposed to a large amount of factual knowledge during pretraining. How can we ensure that this knowledge is applied in downstream tasks? We make the surprising discovery that finetuning on facts the model knows well ensures leads to better factuality downstream. We substantiate our findings via a theoretical analysis of fine-tuning dynamics.

In many real-world settings, such as autonomous driving, a feature can be important to pay attention to in one context (i.e. a crosswalk), but a distractor in another (a traffic light). How can we make sure that deep neural networks rely on the right features in the right context (without using distractors)? We propose a two stage training method that leverages a limited set of human annotations to enforce robustness. We validate our proposed method on the real-world OpenWaymo challenge.

Inferring desired behavior from users is critical for safe and reliable AI, however, users often can give noisy or systematically biased feedback. In this work, we study the effect of modeling the user rationality level when learning a reward function from human feedback. We find that modelling rationality level can be especially desirable in active multi-feedback-type reward learning settings: different types of feedback queries can be optimal depending on the rationality level.

Credit to Jon Barron for the website template source code.