Did Codex Reset
GitHub

Question's Gambit changes the first retrieval step for deep research agents

elvis

Question's Gambit runs once before a deep research agent starts searching. It splits the question into clues, turns each clue into complementary searches, pools the results, and reranks them so the agent begins with that ranked set already in context. The paper is Question's Gambit (arXiv:2609.14412).

With the same retriever and the same agent loop, the change moves BrowseComp-Plus accuracy from 83.1% to 90.5% for GPT-5.5, from 68.1% to 79.0% for GPT-5.4-mini, and from 71.4% to 76.9% for DeepSeek-v4-pro. It roughly halves calibration error for GPT-5.5 and adds 2.3 to 5.3 tool calls per question. elvis describes the gain as coming from this opening context.

In an error analysis of the remaining GPT-5.5 mistakes, only 3 of 79 come from the gold document never being retrieved. The other 76 occur later, when the agent previews, opens, or uses the evidence.

#question's-gambit#deep-research#browsecomp-plus#retrieval

Best Taste-Bench model chooses the better direction 59.7% of the time

A paper from Microsoft and colleagues reports that the best model answers 59.7% of Taste-Bench questions correctly when choosing the better open direction in long engineering and research tasks. Forks decided by later evidence are much harder, a larger reasoning budget does not raise accuracy, and distilling a teacher’s outcome judgment improves held-out SWE-bench Pro success.