Amortized Reasoning Index: a retriever that self-improves during search, no labels needed
lateinteraction · x · 2026-11-07
The Amortized Reasoning Index (ARI) is a new kind of retriever that updates its weights during search, using inference-time gradients derived from a reasoning model's ranking preferences. The insight: agents spend lots of reasoning judging which search results matter, then throw that judgment away on compaction — ARI amortizes it so past reasoning keeps improving future search. Starting from a weak 400M cross-encoder and with no relevance labels, ARI grows its Recall@100 on held-out queries substantially.
More from coding & agent
- Building self-evolving systems: running dozens of research workflows daily, the real job is filtering — EXM7777 · 2026-11-07
- 1,500 unique agents paid in 18 hours, nearly 3x the prior period — adamamcbride · 2026-11-07
- Infinite Factory: endless robot factory generated live in-browser with WFC and OpenUSD — mishig25 · 2026-11-07
- Claude Code usage-limit fix: offload small tasks to DeepSeek, first task cost $0.0043 — PrajwalTomar_ · 2026-11-07
- 7 new MCP servers this week, including one that phones you when your agent gets stuck — Prior_Engineering486 · 2026-11-07
- Dev replaces vector-search RAG with Markdown + knowledge graphs via Claude CLI, claims ~95% fewer hallucinations — Shivam__kumar · 2026-11-07