convergence-study-tested-16-models-800-problems
IN premise — summaries/2026/08/24/convergence-without-understanding-2026-s6-limitations.md
Created 2026-08-24T17:10:52+00:00
The Usama & Chang (2026) convergence study tested 16 models (1.5B–72B parameters) on 800 reasoning problems to identify three dissociations between representational and reasoning convergence.
Summary
A 2026 study by Usama and Chang ran 800 reasoning tasks across 16 different-sized language models and found that a model's internal knowledge and its actual reasoning ability can stabilize at different rates, meaning that just because a model "knows" something doesn't guarantee it reasons about that thing reliably. This serves as the empirical anchor for any downstream claims about when and how model capabilities converge.