FORT-Searcher: Synthesizing Shortcut-Resistant Search Tasks for Training Deep Search Agents
English summary
Researchers propose FORT, a framework for synthesizing training data for deep search agents that resists shortcut learning. It identifies and mitigates four types of shortcut risks: evidence co-coverage, single-clue selectivity, exposed constants, and prior-knowledge binding. The framework uses trajectory signatures to measure and control shortcut risks during data generation. Experiments show that FORT-generated data leads to improved search agent performance on deep search benchmarks. The accompanying tool, FORT-Searcher, outperforms comparable agents on challenging tasks. Code is available on GitHub.
Chinese summary
研究人员提出了FORT框架，用于合成能抵抗捷径学习的深度搜索代理训练数据。该框架识别并缓解了四种捷径风险：证据共覆盖、单线索选择性、暴露常量和先验知识绑定。它利用轨迹签名来测量和控制数据生成中的捷径风险。实验表明，FORT生成的数据使搜索代理在深度搜索基准上性能提升。配套工具FORT-Searcher在挑战性任务上优于同类代理。代码已开源在GitHub。
Key points
FORT framework synthesizes training data that resists shortcut learning for deep search agents.
FORT框架合成能抵抗捷径学习的深度搜索代理训练数据。
Identifies four shortcut risks: evidence co-coverage, single-clue selectivity, exposed constants, and prior-knowledge binding.
识别出四种捷径风险：证据共覆盖、单线索选择性、暴露常量和先验知识绑定。
Uses trajectory signatures to measure and control shortcut risks during data generation.
利用轨迹签名在数据生成过程中测量和控制捷径风险。
FORT-generated data improves search agent performance on deep search benchmarks.
FORT生成的数据在深度搜索基准上提升了搜索代理的性能。
FORT-Searcher tool outperforms comparable agents on challenging search tasks, with code available on GitHub.
FORT-Searcher工具在挑战性任务上优于同类代理，代码已在GitHub开源。