Question's Gambit Boosts Agentic Search Accuracy From 83.1% to 90.5% on BrowseComp-Plus
_reachsumit · x · 2026-09-15
A new paper studies the first retrieval move of deep research agents:
- Agents often retrieve evidence-bearing documents but fail to connect them to the gold documents.
- Question's Gambit, a pre-loop module, decomposes the question into clues, reformulates them into complementary searches, consolidates retrieved results, and reranks the candidate pool before the agent's iterative search-and-reasoning loop begins.
- On BrowseComp-Plus with gpt-5.5, answer accuracy improves from 83.1% (Pi-Serini, the strongest reported agentic baseline) to 90.5%; gains also transfer to MultiHop-RAG.
- Takeaway: effective agentic deep search depends not just on in-loop tools but on the opening move.
More from coding & agent
- MiniMax launches Design: an agent-driven canvas that turns one brief into a full video production — Hailuo_AI · 2026-09-15
- Anthropic explains how it contains its own agents with gVisor, system sandboxes and restricted VMs — emmanuelvivier · 2026-09-15
- How to write tests in the AI era: 6 rules to stop mock-riddled brittle tests — dotey · 2026-09-15
- Qdrant engineer to demo a self-healing research agent that learns from its own run history — qdrant_engine · 2026-09-15
- XWiki clarifies its MCP server is optional, not a forced-in AI assistant — LorinaBalan · 2026-09-15
- Dev Finds Screen Casting Trivial to Implement, but Codex and Claude Both Took the Long Way Around — ezshine · 2026-09-15