Researchers · Interviewed at ICLR 2026
Columbia University

ICLR 2026 · May 29, 2026 · 8 min
AI agents are powerful - but they could be painfully slow. Why? Because most agent workflows are fundamentally sequential: observe → reason → act → wait → repeat.
In this interview, Naimeng Ye explains Speculative Actions, a new framework that speeds up AI agents by introducing parallelism into agent execution. Inspired by speculative execution in CPUs and speculative decoding in LLMs, the method predicts likely future actions and starts executing them before the main agent finishes reasoning.