Cosine-similarity routing to semantic anchors in MoE models matches linear routing perplexity on WikiText-103 while providing direct traceability of decisions and a bandpass loss that cuts dead experts from 30-45% to 0-6%.
Gshard: Scaling giant models with conditional computation and automatic sharding
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Cosine-Similarity Routing with Semantic Anchors for Interpretable Mixture-of-Experts Language Models
Cosine-similarity routing to semantic anchors in MoE models matches linear routing perplexity on WikiText-103 while providing direct traceability of decisions and a bandpass loss that cuts dead experts from 30-45% to 0-6%.