Dylan Sam

Predicting the Performance of Black-box Language Models with Follow-up Queries

Safety Pretraining: Toward the Next Generation of Safe AI

Analyzing Similarity Metrics for Data Selection for Language Model Pretraining

Eliciting Black-Box Representations from LLMs through Self-Queries

Bayesian Neural Networks with Domain Knowledge Priors

The Promises and Pitfalls of Language Models for Structured Numerical Data

Finetuning CLIP to Reason about Pairwise Differences

Understanding prompt engineering may not require rethinking generalization

Computing Low-Entropy Couplings for Large-Support Distributions