ARI: a retriever that updates weights during search using inference-time gradients
CShorten30 · x · 2026-11-07
dianetc et al. introduce the Amortized Reasoning Index (ARI), a new kind of retriever that updates its weights during search using gradients derived at inference time from a reasoning model's ranking preferences.
- Motivation: agents spend lots of reasoning judging which search results matter, then throw that judgment away on compaction; ARI amortizes it so past reasoning improves future searches
- Needs no relevance labels: a 400M cross-encoder starts weak and grows Recall@100 on held-out queries
- Antoine Chaffin: backpropagating production signal directly can unlock what generalization can't, though open questions remain (working on pre-computed representations, cheap global updates)
More from coding & agent
- YC-backed Trepo debuts kitchen hardware that lets AI agents shop for groceries in the background — ycombinator · 2026-11-07
- Cognition ships Devin Review 2.0: 5× cheaper, 1.5× faster PR review — smtabatabaie · 2026-11-07
- Air moves its agentic development experience directly into JetBrains IDEs, free tiers included — therealdanvega · 2026-11-07
- PSA: keep CSS :has() rules tight — agents' broad selectors trigger huge style recalcs — jh3yy · 2026-11-07
- Solo founder credits AI chief-of-staff bots for record-breaking revenue month — joshgonsalves_ · 2026-11-07
- Trail of Bits: Three Lessons From Running Hundreds of AI Agents on Critical OSS — banteg · 2026-11-07