Abstract Neural networks and other architectures progressively reshape the geometry of data space from data’s raw form to one that better suits high-level tasks. This geometric perspective suggests that we can manipulate the shape of these embeddings to improve interpretability or performance on certain tasks, and that we can devise …
Abstract LLM alignment methods typically learn a single reward model (either implicitly or explicitly) from pairwise comparison data. This approach implicitly assumes homogeneous preferences across human labelers — an assumption that is violated in practice. As a result, the learned reward model is generally mis-specified: Prior work shows that it …
Abstract Modern LLMs process information by repeatedly applying a basic primitive of matrix multiplication. Estimates show that about 60-84% of the energy consumed by LLMs goes into memory load/store operations. How can we reduce this power consumption? LLM converts text into a sequence of tokens (which can be thought as …
Abstract: We are interested in the problem of unit-level counterfactual inference in the presence of unobserved confounders owing to the increasing importance of personalized decision-making in many domains: consider a recommender system interacting with a user over time where each user is provided recommendations based on observed demographics, prior engagement …