学术相关
GrepSeek: Training Search Agents for Direct Corpus Interaction 1
GrepSeek, an optimized direct corpus interaction (DCI) search agent
使用Unix-style shell commands (rg, head, grep, find, sed, awk, tail, cat, ls, wc, sort, cut, uniq)
直接通过RL学习搜索行为会导致agent行为不稳定，检索命令效率低效，频繁检索额外大语料段
- 创建冷启动成功DCI数据集
- 由ground-truth倒推检索命令
- 使用answer-blind Planner将倒推路径转换为因果有效的前向路径
- SFT
-
GRPO
- model: Qwen3.5-9B
- benchmark
- NaturalQuestions, TriviaQA, PopQA, HotpotQA, 2WikiMultihopQA, MuSiQue, Bamboogle
- baseline
- vanilla LLM, naive RAG, IRCoT, Search-O1, Search-R1
这类检索命令对于LLM比较熟悉的语境有优势，容易生成精准的检索词，但在语义复杂的环境下可能表现不佳，混合策略可能还是会是个更好的方向
-
Salemi, A., Zeng, C., Nijasure, A., Chung, J.-H., Rahimi, R., Diaz, F., & Zamani, H. (2026). GrepSeek: Training search agents for direct corpus interaction (Version 1). arXiv. https://doi.org/10.48550/ARXIV.2605.29307 ↩