AI Paper Atlas

Research map / Multi-agent systems and games / LLM agents

One Tool Is Enough: Reinforcement Learning of LLM Agents for Repository-Level Code Navigation

Zhaoxi Zhang, Yitong Duan, Yanzhi Zhang, Yiming Xu, Zhixiang Wang, Kun Liang, Weikang Li, Jiahui Liang, Deguo Xia, Jizhou Huang, Jiyan He, Shuxin Zheng and 1 more

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RepoNavigator, an LLM agent with a single execution-aware tool trained with reinforcement learning for repository-level code localization.

Problem

Locating files and functions that need modification in large software repositories is hard; existing LLM methods treat it as retrieval and rely on multiple auxiliary tools that ignore execution logic.

Approach

An LLM agent equipped with a single execution-aware tool, trained via reinforcement learning to navigate repositories, replacing multi-tool retrieval pipelines.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add all three extra tools (GetClass, GetFunc, GetStruc) to jump (ablation)RepoNavigator · SWE-bench Verified function-level (7B)IoU: 24.28 -> 13.71Hurt
add GetClass and GetFunc to jump (ablation)RepoNavigator · SWE-bench Verified function-level (7B)IoU: 24.28 -> 21.44Hurt
add GetStruc to jump (ablation)RepoNavigator · SWE-bench Verified function-level (7B)IoU: 24.28 -> 24.00Hurt
train with GRPORepoNavigator 7B · SWE-bench Verified function-levelSample-F1: 16.19 -> 27.49Improved
train with GRPORepoNavigator 14B · SWE-bench Verified function-levelSample-F1: 25.58 -> 29.23Improved
train with GRPORepoNavigator 32B · SWE-bench Verified function-levelSample-F1: 27.12 -> 34.09Improved
train with GRPORepoNavigator 7B · SWE-bench Pro function-levelSample-F1: 10.67 -> 14.29Improved
add tool-calling success rate to the reward (hybrid reward) (ablation)GRPO RepoNavigator 7B · SWE-bench Verifiedlocalization metricsImproved
use RL (GRPO) instead of SFT (ablation)RepoNavigator 7B · SWE-bench Verifiedlocalization metricsImproved
allow more tool-calling turns (ablation)RepoNavigator · SWE-bench VerifiedDICE and IoU: avg slope 0.51 / turn to 1.08 / turnImproved
replace Agentless localization with RepoNavigatorAgentless repair · SWE-bench Verified (14B)resolved (%): 10.12 -> 14.74Improved
add RL training to RepoNavigator localizationRepoNavigator · SWE-bench Verified (14B)resolved (%): 14.74 -> 15.03Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy