Inducing Reasoning Primitives from Agent Traces
The paper introduces a method called Reasoning Primitive Induction, which aims to enhance the performance of ReAct-style LLM agents. By mining successful agent traces and clustering reasoning moves, the method creates a library of pseudo-tools that can be utilized during testing. The results indicate that these induced libraries significantly outperform the original agents across various tasks.
- ▪Reasoning Primitive Induction mines successful ReAct traces to create a library of pseudo-tools.
- ▪The induced libraries outperform the original agents by significant margins on various tasks.
- ▪The method improves performance on subtasks involving narrative deduction, rule application, and constraint-satisfaction planning.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2606.02994 (cs) [Submitted on 2 Jun 2026] Title:Inducing Reasoning Primitives from Agent Traces Authors:Zhihan Lei, Jiarui Yan, Joshua Momo, William W. Cohen View a PDF of the paper titled Inducing Reasoning Primitives from Agent Traces, by Zhihan Lei and 3 other authors View PDF HTML (experimental) Abstract:ReAct-style LLM agents often rediscover the same reasoning routines across problems, yet leave those routines trapped in transient scratchpads. We introduce Reasoning Primitive Induction, a single-pass method that mines successful ReAct traces, clusters recurrent reasoning moves, and converts the most frequent moves into a compact library of typed pseudo-tools.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
