ML research · long-context retrieval

Patrick McConlogue

Truth In 30 Million Token Context Windows: retrieve the needle with a certificate, or abstain when the bank does not contain the answer.

Measured needle retrieval 1.000 versus RAG 0.000–0.003 (n=300, 30M-token corpus). HiPPO bank 7.5 MB at 30M tokens versus 4.42 TB attention KV on Qwen3-8B.