
Article
MoE: Accuracy jumps from 59.7% to 71.3% via repeated expert reselection
An independent developer testing a looped Transformer with MoE found that reselecting experts at every iteration boosts multi-step math problem solving.
Guides, model comparisons and notes from NeyroLearn members.

Article
An independent developer testing a looped Transformer with MoE found that reselecting experts at every iteration boosts multi-step math problem solving.

Article
FlashAttention-4 84,14 s 1,01× Base PyTorch SDPA 85,07 s 1,00× We expected more.