Zeroth-Order Methods Advance Stochastic Nonconvex Optimization

Two zeroth‑order algorithms were introduced to solve stochastic nonconvex nonsmooth composite optimization without smoothness assumptions, providing finite convergence guarantees (6 Oct 2025). getnews.me/zeroth-order-methods-adv... #zerothorder #nonconvex

Zeroth-Order Optimizer Improves LLM Fine‑Tuning Efficiency

Researchers unveiled ZO Fine‑tuner, a zeroth‑order optimizer that slashes GPU memory for LLM fine‑tuning. Tested on four models across seven benchmarks, it outperformed prior baselines in 82.1% of cases. getnews.me/zeroth-order-optimizer-i... #zerothorder #llm

New Distributed Zeroth-Order Optimization Method Reduces Communication Load

Com‑DSZO, a stochastic distributed zeroth‑order algorithm, uses two function evaluations per iteration and attains O(1/√t) convergence while compression cuts communication. getnews.me/new-distributed-zeroth-o... #distributed #zerothorder

Hybrid AI Training with Low‑Rank Surrogates and Zero‑Order Optimization

A framework trains neural nets with photonic or neuromorphic layers via zeroth‑order optimization and a low‑rank surrogate, achieving digital accuracy with hardware queries. Read more: getnews.me/hybrid-ai-training-with-... #zerothorder #lowrank