SAEs reveal fine-grained specialization in MoE models
Sparse autoencoders trained on Qwen3-30B-A3B show distinct experts for specific tasks like Java coding and geometry.
1 stories, newest first · one card per story
Sparse autoencoders trained on Qwen3-30B-A3B show distinct experts for specific tasks like Java coding and geometry.
Today 1.1K readers · 1.5K views · 483 reads · all time 184K views