15M-Parameter Experiment: Linear RNN Matches Transformer Accuracy at Higher Speed

A controlled experiment training two 15M-parameter decision models with matched budgets found that a linear recurrent backbone matched a Transformer in accuracy while training noticeably faster, with reported speedups ranging from 2.4x to 6x.

2026-10-15 ~ 2026-10-15 · 2 related posts

• 15M-param face-off: linear RNN ties Transformer on accuracy, trains ~6-7x faster — AlarmingTrouble5261 · 2026-10-15

1 near-duplicate retellings: AlarmingTrouble5261