SHERLOC: Structured Diagnostic Localization for Code Repair Agents

A new training-free framework called SHERLOC achieves state-of-the-art fault localization for code repair agents, significantly boosting repair rates while reducing computational overhead, according to research published on arXiv [1]. The framework, formally named Structured Hypothesis-driven Exploration and Reasoning for Localization, pairs a reasoning large language model with compact repository tools and a self-recovery mechanism, requiring no fine-tuning or multi-agent orchestration [1]. Hovhannes Tamoyan is listed as the submitting author on the paper [1]. The researchers report that SHERLOC reaches 84.33% accuracy@1 on SWE-Bench Lite and 81.27% recall@1 on SWE-Bench Verified [1]. At approximately 30 billion parameters, the framework matches or outperforms other agentic localization methods [1]. Current LLM-based coding agents spend roughly half their operational budget on locating faults before making any edits, a bottleneck that dedicated localization frameworks have tried to address [1]. However, the authors note that existing localization tools are typically evaluated as file-retrieval tasks rather than providing the diagnostic context a repair agent needs to act [1]. SHERLOC is designed to fill this gap by producing structured diagnostic findings alongside pinpointed locations [1]. When the locations and diagnostic outputs from SHERLOC are injected into downstream repair agents, the average resolve rate on SWE-Bench Verified improves by 5.95 percentage points [1]. The integration also cuts localization token usage by 36.7% and total token consumption by 23.1% [1]. These efficiency gains address a persistent challenge in agentic software engineering, where token budgets directly constrain the number of reasoning and tool-use turns an agent can execute before hitting context-window limits [1]. The paper appears on arXiv under the Computation and Language category, submitted on June 23, 2026 [1]. The work contributes to a growing body of research on autonomous code repair, an area where benchmark-driven evaluation on datasets like SWE-Bench has become a standard for measuring progress in automated software maintenance [1].