Agentic ESOpt: Fine-Tuning Long-Horizon LLM Agents with Minimal GPU Memory

_akhaliq · x · 2026-08-20

Agentic ESOpt is a new framework designed to reduce the high GPU memory costs of training long-horizon LLM agents. It replaces traditional backpropagation with evolution strategies, enabling full-parameter optimization with just inference-level memory requirements. The method demonstrates strong performance gains on benchmarks like WebArena-Lite.

Related event: Agentic ESOpt Fine-Tunes Long-Horizon LLM Agents with Minimal GPU Memory(2 posts)→

Original post →

More from coding & agent

coding & agent channel →