Fine-Grained LLM Evaluations DILL Lab research on capability structures for understanding reasoning failures in language models. PoisonPerf Data poisoning attacks in performative prediction. Subliminal Learning Experiments Apart Research hackathon work on toy models of hidden bias transfer. Weight Decomposition for Interpretability MARS work-trial experiments on decomposing learned weights into human-interpretable components. Superposition and Steering Work Trial MARS work-trial analysis of sparse feature encoding, interference, and activation geometry. AI for Discovery in the Sciences ML curriculum and tutorial notebooks for PhD students and industry professionals. Tetsuwan Scientific ML engineering for LLM- and vision-based laboratory automation systems. Galactic Dynamics REU Computational astrophysics research on Antlia 2 and Milky Way outer-disc perturbations. Philosophy of Physics Earlier research on relativity, quantum mechanics, and conceptual foundations.