Question's Gambit: first-step reranking lifts agentic search accuracy to 90.5%
CShorten30 · x · 2026-11-13
Weaviate Podcast #150 with ZeroEntropy co-founder Ghita Houir Alami (now leading model research at Notion) covers two big search ideas:
- Question's Gambit: the first retrieval move matters most — decompose the query into clues, run BM25 per clue, pool and rerank, take top-5 as the agent's first observation; later turns stay plain BM25. On BrowseComp-Plus with GPT-5.5, accuracy jumps from 83.1% to 90.5% at roughly the same search budget.
- zELO: like chess Elo, sparse pairwise document battles fit absolute scores via a Thurstone model, distilled into the SOTA zerank-2 cross-encoder.
More from coding & agent
- New research tackles the agent experience problem: knowledge trapped in weights — khademinori · 2026-11-13
- YC F26's Acyclic Labs open-sources infra to run 10,000-agent swarms in one session — ycombinator · 2026-11-13
- Floci: free local AWS emulator with 31.6k GitHub stars for faster integration tests — JeremyCMorgan · 2026-11-13
- Open-source REM tool reverse-engineers competitors' ads, SEO and social playbook from one URL — moeinteractive · 2026-11-13
- Codex agent shrinks a 27M-param transformer to 11.46M with equal loss via weight sharing — yoshiyama_akira · 2026-11-13
- Google open-sources Harness²: Agent = Model + Harness, recursive agent harnessing framework — 2C_ornot2C · 2026-11-13