PayPal Improves Commerce Agent Performance with NVIDIA’s NeMo Framework and Fine‑Tuned Nemotron Model

Global: PayPal Improves Commerce Agent Performance with NVIDIA’s NeMo Framework and Fine‑Tuned Nemotron Model

In a preprint posted to arXiv in December 2025, researchers from PayPal and NVIDIA describe the development and optimization of PayPal’s Commerce Agent, a multi‑agent system intended to enhance agentic commerce on the PayPal platform. The work leverages NVIDIA’s NeMo Framework to fine‑tune a Nemotron small language model (SLM) for the system’s Search and Discovery component, aiming to reduce latency, lower operational costs, and preserve or improve overall agent quality.

Strategic Collaboration and Technical Foundation

The partnership between PayPal and NVIDIA centers on applying the NeMo Framework—traditionally used for speech and language tasks—to a commerce‑specific context. By integrating the framework into the NEMO‑4‑PAYPAL architecture, the team created a pipeline for systematic model refinement that aligns with the performance demands of large‑scale e‑commerce transactions.

Fine‑Tuning Methodology and Hyperparameter Exploration

Experiments employed the llama3.1‑nemotron‑nano‑8B‑v1 architecture, with LoRA adapters trained across a range of learning rates, optimizers (Adam and AdamW), cosine annealing schedules, and LoRA ranks. This exhaustive hyperparameter sweep enabled the researchers to identify configurations that optimized the retrieval‑focused tasks central to the Commerce Agent’s operation.

Performance Gains in Retrieval Component

Results indicate that the fine‑tuned Nemotron SLM resolved a critical bottleneck: the retrieval component, which accounts for over 50% of total agent response time. The optimized model achieved a measurable reduction in latency while maintaining or enhancing the quality of search outcomes, thereby delivering a more responsive user experience.

Implications for Scalable E‑Commerce Systems

The study demonstrates that targeted LLM fine‑tuning can deliver tangible efficiency improvements in production e‑commerce environments. By reducing computational overhead and associated costs, the approach supports the deployment of more scalable, cost‑effective multi‑agent systems across large transaction volumes.

Future Directions and Broader Applications

The authors propose extending the framework to additional agentic functions within PayPal’s ecosystem and exploring its applicability to other commerce platforms. Ongoing research will assess long‑term stability, adaptability to evolving user queries, and integration with emerging AI safety standards.

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.