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

One Tool Is Enough: Reinforcement Learning of LLM Agents for Repository-Level Code Navigation Zhaoxi Zhang 1 Yitong Duan 2 Yanzhi Zhang 2 Yiming Xu 1 Zhixiang Wang 1 Kun Liang 1 Yang Li 3 Jiahui Liang 3 Deguo Xia 3 Jizhou Huang 3 Jiyan He 2 Yunfang Wu 1 Abstract Locating files and functions requiring modifica- tion in large software repositories is challenging due to their scale and structural complexity. Exist- ing LLM-based methods typically treat this as a repository-level retrieval task and rely on multiple auxiliary tools, which often overlook code exe- cution logic and complicate model control. We propose RepoNavigator, an LLM agent equipped with a single execution-aware tool: jumping to the definition of an invoked symbol. This unified design reflects the actual flow of code execution while simplifying tool manipulation. RepoNavi- gator is trained end-to-end via Reinforcement Learning (RL) directly from a base pretrained model, without relying on closed-source distilla- tion. Experiments demonstrate that RL-trained RepoNavigator achieves state-of-the-art perfor- mance, with the 7B model outperforming 14B baselines, the 14B model surpassing 32B com- petitors, and the 32B model exceeding closed- source models such as GPT-5 on most metrics. These results confirm that integrating a single, structurally grounded tool with RL training provides an efficient and scalable solution for repository-level issue localization. 1. Introduction With the rapid advancement of Large Language Models (LLMs) (Liu et al., 2024a; Team, 2024; Yang et al., 2025a), equipping LLMs with pre-built tools to form LLM agents has become a common paradigm for expanding their capa- bilities (Shen, 2024; Yuan et al., 2024; Lu et al., 2024). In the domain of software engineering (SWE), although LLM agents can effectively handle simple programming tasks 1School of Computer Science, Peking University 2Zhongguancun Academy 3Baidu Inc. Correspondence to: Yitong Duan , Yunfang Wu . Figure 1: Illustration of a LLM navigating through a code repository. The LLM is equipped with a single yet powerful tool: jump, which is realized through a language server. (Hui et al., 2024; Guo et al., 2024a), their ability to operate on large-scale software repositories remains limited. SWE- BENCH (Jimenez et al., 2023) currently serves as the most comprehensive benchmark for evaluating whether LLMs can resolve real-world GitHub issues. All pretrained LLMs cannot process the whole repository directly due to context limits (Yang et al., 2024b). While SWE-AGENT (Jimenez et al., 2023) provides moderate gains, it remains far from enabling robust repository-level reasoning. Most existing agents rely on test-time scaling applied di- rectly to pretrained LLMs (Liu et al., 2023; Chen et al., 2025; Schmidgall et al., 2025). In SWE tasks, tools are essential rather than optional: real-world repositories are far larger than the context window of current LLMs, making it impossible to process an entire codebase in a single forward pass. Agents must therefore iteratively invoke tools to re- trieve partial information from the repository and interleave natural-language reasoning with tool calls. Mainstream LLM agents (Chen et al., 2025; Liu et al., 2025; Xiang et al., 2025; Wang et al., 2023; Chen et al., 2024) are rarely exposed to such agentic interaction patterns during pretraining and typically acquire tool usage only through few-shot prompting which is insufficient for learning com- plex multi-step tool-chaining behaviors. Moreover, because tool definition spaces are effectively unbounded, pretrained models cannot fully internalize their semantics without post- training. To mitigate these issues, post-training paradigms such as Supervised Finetuning (SFT) (Ma et al., 2025) and Reinforcement Learning with Verifiable Rewards (RLVR) (Yu et al., 2025a; Yue et al., 2025) have been applied, with promising results in domains including retrieval agents (Jin 1 arXiv:2512.20957v5 [cs.SE] 24 Jan 2026 One Tool Is Enough: Reinforcement Learning of LLM Agents for Repository-Level Code Navigation et al., 2025a), GUI agents (Hong et al., 2024), and math agents (Yan et al., 2025). Directly training an agent to fix software issues, however, remains difficult. A single bug often admits multiple valid patches, making string-level evaluation unreliable. The only precise evaluation method requires executing candi- date patches inside a dedicated Docker environment for each repository (Luo et al., 2025), which is prohibitively expen- sive CPU resource for supervised training. To make training more tractable, we adopt a simplified yet widely general- izable assignment: issue localization. Prior work shows that a software issue becomes substantially easier to resolve once the relevant functions and files are correctly identified (Chen et al., 2025; Ma et al., 2025; Xia et al., 2024; Jiang et al., 2025). Since modern software repositories contain a signif

…(Full text truncated)…

This content is AI-processed based on ArXiv data.