kandpal-2023-bloom-176b-triviaqa-accuracy-vs-doc-count

IN premise — summaries/2026/08/24/kandpal-2023-long-tail-knowledge-s1-introduction.md

Created 2026-08-25T02:58:05+00:00

BLOOM-176B accuracy on TriviaQA rises from approximately 25% at 10^1 relevant documents to over 55% at 10^4 relevant documents.

Summary

Giving the 176-billion-parameter BLOOM model more relevant background documents roughly doubles its chance of answering a trivia question correctly, jumping from about a quarter of the time to over half. This means the model's factual performance is heavily driven by how much good context it receives rather than by its internal knowledge alone, so a system built on it should prioritize strong retrieval and context delivery.