NAVER LABS Europe Trains LLM Agents to Learn From Retrieved Experience

October 24, 2026, 15:07Source: NAVER LABS Europe

Researchers at NAVER LABS Europe have proposed a way to combine fine-tuning and experience retrieval so that LLM agents generalize better to tasks they have not seen before. The paper, "Retrieval-augmented LLM agents: Learning to learn from experience," is listed for the EMNLP conference in Budapest.

Key points

NAVER LABS Europe proposes combining fine-tuning with experience retrieval for LLM agents.
The authors say fine-tuning often fails to extrapolate to new tasks, while retrieval alone often underperforms supervised baselines.
They present a LoRA-based SFT recipe that they say outperforms several state-of-the-art agent training pipelines.
The paper analyzes storage, querying and trajectory selection, and integrates retrieval into fine-tuning.
It is listed for EMNLP in Budapest, 24 to 29 October 2026.

The problem with current agent approaches

Large language models have helped advance general-purpose agents, but the authors say robust generalization to unseen tasks remains a significant challenge. According to the paper, current methods usually follow one of two routes: fine-tuning, or training-free memory-augmented generation that uses retrieved experience.

Each route has a weakness, the researchers write. Fine-tuning often fails to extrapolate to new tasks, while experience retrieval often underperforms compared to supervised baselines. The new work asks whether the two can be merged, and studies systematically how to train retrieval-augmented LLM agents to use retrieved trajectories in-context.

Three contributions

The authors, Thomas Palmeira Ferraz, Romain Deffayet, Vassilina Nikoulina, Hervé Déjean and Stéphane Clinchant, describe three steps in their study.

A supervised fine-tuning recipe. They first establish a robust supervised fine-tuning (SFT) recipe using LoRA, which they say outperforms several state-of-the-art agent training pipelines.
An analysis of retrieval design. They provide a detailed analysis of key design choices for experience retrieval, identifying optimal strategies for storage, querying and trajectory selection.
A combined pipeline. They propose a pipeline that integrates experience retrieval into the fine-tuning process itself.

The source does not give benchmark names or numerical results in the abstract.

What the authors claim and where it will be presented

According to the authors, their results demonstrate that the combined approach significantly improves generalization to unseen tasks. They describe it as a scalable and effective framework for building agents that learn to learn from experience.

The paper is listed for the Conference on Empirical Methods in Natural Language Processing (EMNLP), taking place in Budapest, Hungary, from 24 to 29 October 2026. It was published on the NAVER LABS Europe site by Hervé Déjean, and an arXiv version is linked from the page.

What it means

For teams building LLM agents, the paper points to a practical direction: rather than choosing between fine-tuning and retrieval of past experience, train the model so that it knows how to use retrieved trajectories. The reported analysis of storage, querying and trajectory selection may also serve as guidance for those designing experience memories. As the findings are the authors' own claims from an abstract, details on how large the gains are will depend on the full paper.