Combining MoE, MLA, and RoPE in small transformers improves perplexity on TinyStories while cutting KV cache size, but several headline numbers conflict internally.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2025 1verdicts
REJECT 1representative citing papers
citing papers explorer
-
Unifying Mixture of Experts and Multi-Head Latent Attention for Efficient Language Models
Combining MoE, MLA, and RoPE in small transformers improves perplexity on TinyStories while cutting KV cache size, but several headline numbers conflict internally.