Amortized Reasoning Index: a 0.4B retriever continuously RL-trained by a big model matches 8B

lateinteraction · x · 2026-11-07

Diane et al. introduce the Amortized Reasoning Index (ARI), a retriever that updates its own weights during search, using gradients derived at inference time from a reasoning model's ranking preferences — no relevance labels needed.
Motivation: today's agents spend lots of reasoning judging which search results matter, then throw that judgment away at compaction. ARI amortizes it, so reasoning spent on past queries keeps improving future search.
A 400M cross-encoder starts weak but steadily grows Recall@100 on held-out queries; the chart shows the 0.4B ARI (Ettin) matching or beating Qwen3-8B on Twitter-Conflict, Writing, and BRIGHT AoPS, approaching Gemini Embedding 2.
dbreunig highlights this pattern for agent memories, addressing "memory rot" and evolving personalized search.

More from coding & agent