Sparse Attention as MIPS (Maximum Inner Product Search)

Source: https://x.com/teortaxestex Author: teortaxestex Date: 2025-2026

Summary

Insight: sparse attention is equivalent to approximate maximum inner product search (MIPS). Leading techniques like Hierarchical Navigable Small World (HNSW) graphs are already well-studied for MIPS. Suggests importing MIPS research into sparse attention design.

Key Claims

  • Sparse attention = approximate MIPS (Maximum Inner Product Search)
  • MIPS is a well-researched field with mature techniques (HNSW, LSH, etc.)
  • Existing sparse attention research reinvents what MIPS already solved
  • Suggested approach: apply HNSW and related techniques to attention directly

Concepts