Evidence-backed signal

Quantitative reasoning gains increasingly came from composing model scale with specialized training data, chain-of-thought-style prompting and inference-time sampling rather than from raw parameter count alone.

supported inference 96% confidence increasing