RT by @_akhaliq: Agentic ESOpt: fine-tuning long-horizon LLM agents with minimal GPU memory This…

TL;DR - Agentic ESOpt is a framework for fine-tuning long-horizon LLM agents using evolution strategies rather than backpropagation. It enables full-parameter optimization at inference-level GPU memory and reports strong improvements on WebArena-Lite and other unspecified evaluations.

• Replaces gradient-based backpropagation with evolution-strategy optimization.
• Reduces training memory requirements to approximately those of inference.
• Supports full-parameter optimization rather than limiting updates to a small adapter.
• Targets long-horizon agent tasks and reports gains on WebArena-Lite.