Convex Optimization for LLM Preference Elicitation

Convex optimization project • December 2025

This project formulates prompt selection for learning human preferences as an optimal experimental design problem. Using a linear-Gaussian model over answer embeddings, it derives G-optimal and D-optimal formulations, provides convex reformulations (including Schur-complement constraints), and discusses efficient approximate solvers such as Frank–Wolfe, with comparisons against SDP-based approaches.