AIAny
Icon for item

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

Fine-tunes long-horizon LLM agents with evolution strategies so full-model updates run at inference-level GPU memory. Emphasizes trajectory-level credit via black-box rewards, online prompt–parameter co-evolution, and a cosine decay for perturbation scale to balance exploration and adaptation; suited for limited-GPU settings.

Introduction

Most LLM fine-tuning uses gradient-based RL, but long-horizon agentic tasks stretch credit assignment and require heavy backprop memory—blocking full-parameter tuning on modest hardware. The core insight here is pragmatic: use evolution strategies (ES) to treat agent evaluation as a black-box, enabling full-parameter updates with only inference-time memory while keeping trajectory-level rewards intact.

Key Findings
  • ES-based full-parameter fine-tuning scales to large LLM agents with much lower GPU memory requirements, enabling experiments that would be infeasible with RL's backprop stack.
  • On WebArena-Lite, full-parameter optimization of Qwen-3.5-27B with Agentic ESOpt improves the No Skill baseline by 6.69%.
  • In online test-time heuristic/prompt co-evolution, Agentic ESOpt outperforms its matched baseline in 28 of 36 evaluated settings, showing practical gains from joint prompt–parameter adaptation.
  • Method design choices include an online reward-weighted update for sampled perturbations and a cosine decay schedule for the perturbation scale σ to trade off exploration vs. rapid adaptation.
Who it's for and trade-offs

Great fit if you need to fine-tune large LLM agents on long-horizon, sparse-reward tasks but only have one or a few GPUs; the approach preserves trajectory-level credit without needing a learned critic or large GPU memory. Look elsewhere if you require lowest possible wall-clock GPU-hours for short single-turn tasks where gradient-based RL or LoRA-style methods may converge faster; ES can demand many forward-rollouts and careful population/evaluation budgeting. The paper highlights practical gains on agentic benchmarks and demonstrates scenarios (prompt–parameter co-evolution) where black-box perturbation is an advantage, but practitioners should plan evaluation budgets and perturbation schedules to control wall-clock cost.

Information

  • Websitearxiv.org
  • AuthorsZhi Zheng, Rongsheng Chen, Yunpeng Ba, Zhenkun Wang, Yee Whye Teh, Wee Sun Lee
  • Published date2026/08/18

More Items

Encodes videos into a Film Knowledge Graph and reconstructs them to learn agent-native, editable video representations for agentic reasoning and manipulation. Uses agentic auto-encoding with dual-loop textual-gradient optimization, reports large reconstruction gains, and releases a benchmark and dataset.

Turns embodied navigation into 2D visual prompting where a vision-language model selects image pixels that are projected to 3D actions; adds selective chain-of-thought, compressed anchor-trajectory memory, and a two-level alignment objective to improve sample and runtime efficiency.

Enables interactive serving of large Mixture-of-Experts (MoE) models on personal machines by adapting offload and execution to measured device bandwidth and agentic workload patterns. Key features include bandwidth-adaptive execution, semantic-aware caching of recurrent state, and an elastic GPU expert cache; supports 20+ MoE models and runs models from ~35B to 753B on consumer/workstation GPUs.