Shruti Joshi

Simple additive and multiplicative feature transforms can unlock surprisingly robust domain generalization and test-time adaptation in both discriminative and generative models.

Sparse autoencoders' failure to generalize compositionally isn't due to amortized inference, but because they learn lousy dictionaries in the first place.

Interpretability studies on LLMs often overreach, but causal inference offers a framework to ensure claims about model behavior are actually valid and generalizable.