Developer proposes mixture-of-attention architecture with centroid clustering for token routing

InfamousTrouble7993 · reddit · 2026-08-18

A developer released LogPose, a mixture-of-attention LM architecture addressing attention dilution and dead experts. It uses a differentiable soft K-means clustering router to route tokens to specialized attention heads, plus GRU-based recurrent routing and linear routing. The implementation includes a Llama-style decoder baseline (RMSNorm, RoPE, SwiGLU, GQA), built-in GSM8K and MBPP benchmarks, KV and routing-state caching, and MLflow integration. The author asks if this is novel, fearing prior work.

Original post →

More from Research

Research channel →