Liang Zhao

Foundation models can accelerate scientific computation, but only if they respect physical laws and safety limits, a challenge this paper tackles by deriving design principles for constrained scientific foundation models.

Forget end-to-end molecular design; this retrieval-augmented model lets you steer analog generation with prompts and external references, mimicking medicinal chemists' intuition.

LLM personalities can be steered with fine-tuning-level precision, compositionality, and context-awareness, all without training, by directly manipulating activation vectors in representation space.

Scaling laws hold for interest modeling: bigger LLMs and more inference-time sampling consistently boost news recommendation quality, and can be distilled into smaller, deployable models.