ARI: a retriever that trains itself at search time using a reasoning model's gradients

CShorten30 · x · 2026-11-07

dianetc introduces the Amortized Reasoning Index (ARI), with lateinteraction (Charles Packer) calling it a fundamental change in how we build retrievers and a new kind of agent memory.

Core idea: agents spend lots of reasoning judging which search results matter, then throw that judgment away on compaction. ARI "amortizes" it — the retriever updates its own weights during search using gradients derived from a reasoning model's ranking preferences, effectively test-time training for retrievers.
Results: with no relevance labels, a 400M cross-encoder starts weak and steadily improves Recall@100 on held-out queries.
Open problems: making oblique search queries as manageable as non-oblique ones, likely requiring sub-linear late-interaction-style search.

Original post →

ARI: a retriever that trains itself at search time using a reasoning model's gradients

More from coding & agent