Researchers have developed SemNav, a novel framework designed to enhance issue localization in code repositories. This system combines deterministic retrieval with an LLM agent to iteratively identify and rank relevant files and functions for issue resolution. SemNav introduces three key components: a Semantic Navigation Graph for program relation navigation, Semantic Cards for concise entity interpretations, and a Candidate Workspace for tracking evidence and revision. The framework demonstrates significant improvements over existing baselines on benchmarks like SWE-bench Lite and PLocBench, notably increasing File Hit@10 from 68.33% to 82.67% when using Gemma 4B. AI
IMPACT Enhances LLM agent capabilities for software development, potentially improving developer productivity and code quality.
RANK_REASON Academic paper detailing a new framework and its performance on benchmarks. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- CatalyzeX
- DagsHub
- Gemma 4B
- Gotit.pub
- Hugging Face
- PLocBench
- ScienceCast
- SWE-bench Lite
- SWE-Explore
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
