Research
Researchers propose LIMe, an attention extension that cuts LLM training compute by about 15%
The EMNLP 2026 paper from T-Tech adds learnable routing over earlier layers' key-value buffers to attention. Reported results come from models around one billion parameters matching baseline LLaMA training loss.
Every outlet that ran it 1 article
r/LocalLLaMACommunityLead
Researchers Discover a Way to Reduce LLM Training Compute by 15% Without Sacrificing Quality