arxivcs.LGcs.CL2026-07-16
Looped Latent Attention: Cross-Loop KV Compression for Looped Transformers
Looped, weight-tied Transformers reduce parameters by reusing a block, but decoding still stores a separate K/V cache for every recurrence step. We show that this loop-indexed cache is highly structured. For a fixed token, layer and head, K/V vectors trace a short low-rank trajec…