NVIDIA Takes Second in KDD Cup Using Nemotron

The NVIDIA KGMON team achieved its second-place finish in the KDD Cup 2026 Data Agents competition by focusing on harness optimization rather than expanding model size. Powered by the NVIDIA Nemotron LLM, the team's agent had to answer complex natural-language questions across diverse, messy datasets, including SQL databases, CSV and JSON files, PDFs, and even briefing videos. Because the competition mandated a small, fixed language model, the developers focused on restricting the agent's action space to prevent routing errors and preserve reasoning capacity.

To achieve this, the team consolidated all structured data sources into a single SQLite database, giving the agent a unified SQL interface. They introduced a read-only schema-scouting preflight step to brief the agent on tables, join keys, and data-quality issues before the main reasoning loop started. The agent's toolset was strictly limited to four functions: schema(), sql(query), write_answer(df), and prose_helper(). To handle unstructured documents without overwhelming the context window, the prose_helper tool extracted targeted information using a separate zero-temperature LLM call. For tasks involving video, the team preprocessed the files by extracting keyframes and aligning them with transcribed audio.

For AI practitioners, these results demonstrate that agent reliability often stems from engineering a strict environment rather than relying on open-ended model capabilities. By logging execution traces, the team used a specialized inspector agent to categorize failures and systematically refine the system. The KGMON team also utilized repeated attempts and answer-selection ensembling to boost accuracy, though they caution that this approach increases compute costs. Ultimately, the project provides a blueprint for building dependable enterprise data assistants using smaller, cost-effective open models.