blog
Project pages for our recent work.
-
Beyond the paper: mine knowledge, evaluate it, run it on iCode
How the pieces fit: Lingxi Advisor mines procedural knowledge three ways (issue–patch history, agent trajectories, test outcomes), an evaluation kit measures it, and iCode runs it.
-
Lingxi: procedural knowledge for repository-level issue resolution
ISSTA 2026 · ACM SIGSOFT Distinguished Paper. Recovering the procedural knowledge behind a repository's historical issues, and using it to guide a coding agent on new issues.
-
STAIR: Reusing Past Repairs Through Hierarchical Trajectory Abstraction for Coding Agents
Under review. Turning past repair trajectories into repair recipes at several granularities, adapted into stage-by-stage plans for a new issue. 81.2% on SWE-bench Verified.
-
TestGRAD: test-based patch selection for SWE-agent ensembles
Under review. Mining failure patterns from test attempts to evolve the test suite until it tells candidate patches apart. 84.2% on SWE-bench Verified with a 4-agent ensemble.