Radar ·
HOLA combines recurrent state with a limited KV cache
In a post published July 6, 2026, HOLA combines recurrent state with a small KV cache for exact memory. The post reports a 16.1% reduction in perplexity and robust retrieval at 32k tokens.
Published July 6, 2026, the post describes HOLA, a method that combines recurrent state with a small KV cache to retain exact memory. Tokens with high prediction error are directed to the cache. The project explores whether this limited buffer can improve long-context retrieval in linear-attention models.
According to the post, the method reduces perplexity by 16.1% and maintains robust retrieval at 32k tokens. To assess these findings, consult the original post and check its method description, evaluation conditions, and metrics; those details are not included in the available summary.
Original source: HOLA combina estado recorrente com um KV cache limitado