PayPal and NVIDIA Release Optimized Commerce Agent in New arXiv Study

Global: PayPal Unveils Optimized Commerce Agent Using NVIDIA’s NeMo Framework

A new study posted on arXiv in December 2025 details how PayPal, together with NVIDIA, has engineered an enhanced commerce‑agent system called the Commerce Agent, powered by a fine‑tuned Nemotron small language model. The research outlines the collaboration’s goal of improving response latency and operational cost for agentic commerce on the PayPal platform.

System Architecture

The Commerce Agent is built as a multi‑agent framework named NEMO‑4‑PAYPAL. Its design targets the automation of search and discovery tasks within PayPal’s marketplace, integrating large‑language‑model capabilities to handle user queries and product retrieval.

Fine‑Tuning Methodology

Researchers employed NVIDIA’s NeMo Framework to replace the original base model with a Nemotron small language model (SLM). Training leveraged the llama3.1‑nemotron‑nano‑8B‑v1 architecture and applied LoRA adapters. Systematic hyperparameter sweeps explored learning rates, two optimizers (Adam and AdamW), cosine‑annealing schedules, and varying LoRA ranks.

Experimental Findings

Comprehensive experiments demonstrated notable reductions in latency and cost while preserving, and in some cases enhancing, overall agent quality. The fine‑tuned Nemotron SLM addressed a key performance bottleneck in the retrieval component, which accounts for over 50% of total agent response time, leading to measurable speed improvements.

Broader Implications

The authors claim this work represents the first application of NVIDIA’s NeMo Framework to commerce‑specific agent optimization and offers a scalable approach for multi‑agent system refinement in production e‑commerce environments. If adopted widely, the methodology could influence how online marketplaces integrate AI‑driven agents for faster, more cost‑effective user interactions.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.