Question's Gambit:优化检索首步提升深度研究Agent性能
Impressive paper showing how much the first retrieval step matters for deep rese…
Agent 开发者必读,这篇论文用详实数据证明了“开局检索质量”对深度研究能力的决定性影响,提供了可复现的线索拆分与重排序思路。
Impressive paper showing how much the first retrieval step matters for deep research agents.
一篇令人印象深刻的论文,展示了第一步检索对深度研究代理的重要性。
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.
在使用相同的检索器和相同的代理循环的情况下,它帮助 GPT-5.5 在 BrowseComp-Plus 上的表现从 83.1% 提升到了 90.5%。
It seems that the gain comes from the opening context.
这种增益似乎来自于初始上下文。
The authors propose Question's Gambit which runs once, before the agent starts searching.
作者提出了“Question's Gambit”(开局策略),它在代理开始搜索之前只运行一次。
It splits the question into clues, turns each clue into complementary searches, pools the results, and reranks them.
它将问题拆分为线索,将每条线索转化为互补的搜索查询,汇总结果并进行重新排序。
The agent then starts its loop with that ranked set already in context.
随后,代理便带着这个已排序好的集合进入其循环,并将其作为上下文的一部分。
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.
同样的改动使 GPT-5.4-mini 从 68.1% 提升至 79.0%,DeepSeek-v4-pro 从 71.4% 提升至 76.9%,并将 GPT-5.5 的校准误差大致减半。每个问题需要额外花费 2.3 到 5.3 次工具调用。
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.
在错误分析中,GPT-5.5 剩余的 79 个错误中,只有 3 个是因为从未检索到黄金文档所致。其余 76 个错误发生在后续阶段,即代理预览、打开或使用证据时。
Paper: https://arxiv.org/abs/2609.14412
论文:https://arxiv.org/abs/2609.14412
Chat with Paper: https://academy.dair.ai/papers/questions-gambit-the-first-move-matters-in-agentic-deep-search-2609.14412
与论文对话:https://academy.dair.ai/papers/questions-gambit-the-first-move-matters-in-agentic-deep-search-2609.14412
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力