Naive Bayes benchmarks favorably against LLMs for text classification

2026-09-25

New research from arXiv benchmarks Complement Naive Bayes against large language models (LLMs) for text classification tasks. The study indicates that while LLMs excel in zero-data scenarios, Naive Bayes performs comparably or better with labeled data, offering significantly higher throughput.

VERA Brief

AI-generated. Grounded in the article and its cited sources.

New research benchmarks Naive Bayes against large language models for text classification. While LLMs excel in zero-data scenarios, Naive Bayes performs comparably or better with labeled data, offering significantly higher throughput.

Key facts

Naive Bayes (NB) performs comparably or better than large language models (LLMs) on text classification tasks when labeled data is available.
LLMs show superiority in zero-data regimes, but this advantage may be prone to contamination.
Naive Bayes offers significantly higher throughput and lower energy consumption per sample compared to LLMs, especially on commodity CPUs.
The choice between Naive Bayes and LLMs for text classification is task-dependent.
Naive Bayes can reach parity with LLMs for topic classification at approximately 10^4 labels.

Source: arXiv · cs.LG

Reported by VERA Newswire.