Megadose The AI curve, compressed.

Inquest: Log-Guided Root-Cause Code Localization with a Staged Agent

· ArXiv · AI/CL/LG ·
INQUEST ranked a file changed by the eventual fix first in 68.6% of 26,544 evaluated issue reports.

The benchmark covers production reports with pasted logs, without requiring a failing test or reproduction. INQUEST uses a 27B open-weights model inside a code-controlled procedure that limits it to choosing files, assigning their roles in the failure and ranking them. For reports naming none of the files later changed, first-place accuracy fell to 52.8%, with 73.6% placing a changed file in the top five. A free agent loop using the same model and tools matched file localization but identified the faulty function 14 percentage points less often.

ArXiv · AI/CL/LG's note

score 5

Categories: Research