Global Resolution: Optimal Multi-Draft Speculative Sampling via Convex Optimization

ICLROral2026

Authors: Rahul Krishna Thomas, Arka Pal
Affiliation: Columbia University
Venue: ICLR 2026
Track: Oral

TL;DR

We reduce optimal multi-draft speculative sampling to a convex minimization problem, and solve a truncated version to achieve state-of-the-art acceptance with negligible performance degradation.

Opening excerpt from the authors’ abstract. source

Read the paper

OpenReview

Topics

optimization

← All ICLR 2026 Oral papers · Browse the whole archive