Question's Gambit：检索前先走一步，GPT-5.5 涨 7 个点
Impressive paper showing how much the first retrieval step matters for deep research agents. It hel...
智能体出发前先搜一轮，GPT-5.5 在 BrowseComp-Plus 涨 7 个点，做深度研究智能体的可以试试这招
这篇论文提出 Question's Gambit，在智能体开始搜索前先把问题拆成线索，将每条线索转成互补的搜索查询，汇总结果后重排序。同样的检索器和同样的智能体循环下，GPT-5.5 在 BrowseComp-Plus 上的准确率从 83.1% 升到 90.5%，校准误差约减半。GPT-5.4-mini 从 68.1% 升到 79.0%，DeepSeek-v4-pro 从 71.4% 升到 76.9%，代价是每个问题多花约 2.3 到 5.3 次工具调用。错误分析显示，GPT-5.5 剩余的 79 个错误中只有 3 个来自金标文档从未被检索到，其余 76 个都发生在智能体预览、打开或使用证据的阶段。
Impressive paper showing how much the first retrieval step matters for deep research agents. It hel...
Impressive paper showing how much the first retrieval step matters for deep research agents. It helps to improve GPT-5.5 from 83.1% to 90.5% on BrowseComp-Plus with the same retriever and the same agent loop. It seems that the gain comes from the opening context. The authors propose Question's Gambit which runs once, before the agent starts searching. It splits the question into clues, turns each clue into complementary searches, pools the results, and reranks them. The agent then starts its loop with that ranked set already in context. The same change lifts GPT-5.4-mini from 68.1% to 79.0% and DeepSeek-v4-pro from 71.4% to 76.9%, and roughly halves calibration error for GPT-5.5. It costs between 2.3 and 5.3 extra tool calls per question. In an error analysis, only 3 of the 79 remaining GPT-5.5 errors come from the gold document never being retrieved. The other 76 happen later, when the agent previews, opens or uses the evidence. Paper: arxiv.org/abs/2609.14412 Chat with Paper: academy.dair.ai/papers/questio… 💬 3 🔄 0 ❤️ 2 👀 980 📊 3 ⚡