Question's Gambit lifts deep research agents: GPT-5.5 hits 90.5% on BrowseComp-Plus
omarsar0 · x · 2026-09-22
A new paper proposes Question's Gambit, showing the first retrieval step matters enormously for deep research agents — improving only the opening context yields big gains:
- On BrowseComp-Plus with the same retriever and agent loop: GPT-5.5 improves 83.1% → 90.5%, GPT-5.4-mini 68.1% → 79.0%, DeepSeek-v4-pro 71.4% → 76.9%
- Calibration error for GPT-5.5 roughly halves; costs 2.3–5.3 extra tool calls per question
- Method: run once before the agent starts; split the question into clues, turn each clue into complementary searches, pool and rerank results, then seed the agent loop with the ranked set
- Error analysis: only 3 of 79 remaining GPT-5.5 errors come from the gold document never being retrieved; the other 76 happen later in the loop
More from coding & agent
- Deel launches Akai, an agent platform that learns workflows from a single demo — kimmonismus · 2026-09-22
- Uncle Bob Demos uml-viewer: His New Way to Review Agent-Built Projects — blaizedsouza · 2026-09-22
- Progressive disclosure for agents: gradually revealing info as trust builds between agents — iamrobotbear · 2026-09-22
- Qwen 3.8 27b fine-tune cuts verbose output by up to 40% with little quality loss — julianharris · 2026-09-22
- Codex desktop app won't show which command the agent ran, user complains — amplifiedamp · 2026-09-22
- It Was Never About Coding: A Veteran Engineer's Take on the Agent Era — blaizedsouza · 2026-09-22