According to multiple papers published on arxiv.org in March 2026, researchers are advancing methods for training large language model (LLM) agents that can learn from accumulated experience.
One team proposed combining fine-tuning with experience retrieval to help agents generalize to unseen tasks. According to arxiv.org, their work “establish[ed] a robust supervised fine-tuning (SFT) recipe using LoRA that outperforms several state-of-the-art agent training pipelines” and integrated experience retrieval into the fine-tuning process. The researchers stated that “this combined approach significantly improves generalization to unseen tasks.”
Separately, researchers introduced SLEA-RL (Step-Level Experience-Augmented Reinforcement Learning), which addresses a limitation in existing methods. According to arxiv.org, while previous approaches “retrieve experiences only once based on the initial task description,” SLEA-RL “retrieves relevant experiences at each decision step conditioned on the current observation.” The framework includes step-level observation clustering, a self-evolving experience library, and policy optimization with step-level credit assignment. Experiments on long-horizon benchmarks showed SLEA-RL “achieves superior performance compared to various reinforcement learning baselines,” according to the source.
Another research team tackled the “verbalization” challenge—converting structured user data into effective natural language inputs for LLMs. According to arxiv.org, their data-centric framework uses reinforcement learning to optimize how interaction logs are transformed into text. Experiments on a Netflix streaming dataset demonstrated “up to 93% relative improvement in discovery item recommendation accuracy over template-based baselines.”