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

Hugging Face Daily Papers Papers

Summary

This paper proposes Agentic ESOpt, a method using evolution strategies to enable scalable full-parameter fine-tuning of long-horizon LLM agents with minimal GPU memory requirements.

Reinforcement Learning (RL) has been promising in single-turn LLM fine-tuning. However, long-horizon agentic reasoning introduces increasingly branching interactions and sparse rewards, exposing several limitations of RL: its heavyweight backpropagation-based training stack makes it impractical to fine-tune larger LLMs, and longer-horizon trajectories make credit assignment in RL substantially harder. This paper argues that evolution strategies (ES) can be a better choice for fine-tuning long-horizon LLM agents. Compared with agentic RL, ES offers three key advantages: 1) Model Scalability: ES enables full-parameter optimization with only minimal, inference-level GPU memory, making it possible to fine-tune large LLMs. 2) Flexibility: its lightweight, black-box feedback interface makes ES fine-tuning easy to compose with prompt-space evolution (e.g., skill optimization & test-time compute); and 3) Long-Horizon Scalability: ES performs trajectory-level parameter attribution without decomposing rewards across horizons, yielding better scalability than Agentic RL as the horizon length grows. Based on this insight, we propose Agentic ESOpt, a full-parameter agentic fine-tuning framework tailored to flexible parameter--context co-evolution. At each step, Agentic ESOpt samples perturbations around the current LLM parameters, evaluates the resulting agents with rewards, and applies an online reward-weighted update. To improve the exploration--adaptation trade-off, Agentic ESOpt further introduces a cosine decay schedule of the perturbation scale σ. On WebArena-Lite, full-parameter optimization of Qwen-3.5-27B improves the No Skill baseline by 6.69%. In test-time automatic heuristic design, Agentic ESOpt performs online prompt--parameter co-evolution, improving its matched baseline in 28 of 36 settings.
Original Article
View Cached Full Text

Cached at: 08/19/26, 03:58 AM

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

Source: https://huggingface.co/papers/2608.17310

Abstract

Agentic ESOpt uses evolution strategies for scalable full-parameter fine-tuning of long-horizon LLM agents via trajectory-level reward-weighted updates and parameter-context co-evolution.

Reinforcement Learning(RL) has been promising in single-turn LLM fine-tuning. However,long-horizon agentic reasoningintroduces increasingly branching interactions and sparse rewards, exposing several limitations of RL: its heavyweight backpropagation-based training stack makes it impractical to fine-tune larger LLMs, and longer-horizon trajectories makecredit assignmentin RL substantially harder. This paper argues thatevolution strategies(ES) can be a better choice for fine-tuning long-horizon LLM agents. Compared with agentic RL, ES offers three key advantages: 1) Model Scalability: ES enablesfull-parameter optimizationwith only minimal, inference-level GPU memory, making it possible to fine-tune large LLMs. 2) Flexibility: its lightweight, black-box feedback interface makes ES fine-tuning easy to compose with prompt-space evolution (e.g., skill optimization & test-time compute); and 3) Long-Horizon Scalability: ES performs trajectory-level parameter attribution without decomposing rewards across horizons, yielding better scalability than Agentic RL as the horizon length grows. Based on this insight, we proposeAgentic ESOpt, a full-parameter agentic fine-tuning framework tailored to flexible parameter--context co-evolution. At each step,Agentic ESOptsamples perturbations around the current LLM parameters, evaluates the resulting agents with rewards, and applies an onlinereward-weighted update. To improve the exploration--adaptation trade-off,Agentic ESOptfurther introduces acosine decay scheduleof the perturbation scale σ. OnWebArena-Lite,full-parameter optimizationof Qwen-3.5-27B improves the No Skill baseline by 6.69%. In test-time automatic heuristic design,Agentic ESOptperforms online prompt--parameter co-evolution, improving its matched baseline in 28 of 36 settings.

View arXiv pageView PDFProject pageGitHub3Add to collection

Get this paper in your agent:

hf papers read 2608\.17310

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2608.17310 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2608.17310 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2608.17310 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Poor Man's Agentic Modeling: Simulating Large LLM-Agent Societies on a Laptop

arXiv cs.AI

This paper proposes a method for simulating large LLM-agent societies on a laptop by fitting low-parameter surrogate models from a few hundred queries, using a statistical-physics-based taxonomy to predict when this approximation holds. The approach is validated on EconAgent and several other simulations using DeepSeek-elicited agent behaviors.

EASy: Towards Efficient LLM-Based Agentic System

arXiv cs.CL

The paper proposes EASy, a trainable agentic framework that uses reinforcement learning to jointly optimize task performance and computational efficiency, introducing a milestone-plan-act workflow, dependency-aware execution graphs, and tree-structured rollout for training.

EvoArena: Tracking Memory Evolution for Robust LLM Agents in Dynamic Environments

Hugging Face Daily Papers

EvoArena introduces a benchmark for evaluating LLM agents in dynamic environments with progressive updates across terminal, software, and social domains, while EvoMem proposes a patch-based memory paradigm that records structured evolution; experiments show current agents achieve only 39.6% accuracy on EvoArena, and EvoMem yields average gains of 1.5% on the benchmark and improvements on GAIA and LoCoMo.