Researchers have developed GraphScent, a novel method for localizing relevant code snippets for bug fixing when running large language models on limited hardware. By analyzing directional relevance signals in Gemma 4 models across numerous Python issues, they found that scores on enclosing scopes and upstream callers are more informative than downstream callees or lexical embeddings. GraphScent utilizes these findings to improve code localization accuracy, increasing it from 12.8% to 23.4% for a 31B parameter model within a single-GPU budget. AI
IMPACT Enhances the feasibility of running large language models for code analysis on consumer hardware.
RANK_REASON The cluster describes a research paper published on arXiv detailing a new method for code localization. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →