Ablation discussion: query head count barely matters early, sparse routing must be learned

stochasticchasm · x · 2026-08-28

A technical self-reply analyzing two interesting ablations. First, two techniques not usually compared — layer-wise vs sequence-wise indexer sharing — could plausibly be composed, and the author wants a composed ablation. Second, query head count shows almost no impact in the first stage (single vs 4 heads, despite far more FLOPs) but becomes noticeable after stage 2; the author guesses the network must first be trained to route retrieval through sparse blocks, explaining the delayed effect.

Original post →

More from Research

Research channel →