llama2-primacy-bias-threshold-13b
IN premise — summaries/2026/08/24/liu-2023-lost-in-middle-s20-t-otal-retrieved-documents.md
Created 2026-08-25T02:58:08+00:00
Llama-2 7B shows only recency bias (no primacy bias), while Llama-2 13B and 70B show both primacy and recency bias, indicating primacy bias requires approximately ≥13B parameters.
Summary
Whether a model naturally pays extra attention to information at the very beginning of a sequence depends on its size: the 7B model simply doesn't do this, while the 13B and 70B models do. This means that if a downstream system relies on a small model to weight early context or initial instructions, it will silently drop that signal, and you need at least a 13B-class model to get that kind of positional sensitivity.