OrcaLoca

An LLM agent framework for software issue localization.

Resolving a software issue starts with finding the code to change, and that first step is where most agent pipelines lose accuracy. OrcaLoca treats localization as a scheduling problem rather than a single retrieval call:

  • Priority-based scheduling for LLM-guided actions, so the agent spends its budget on the most promising leads first
  • Action decomposition with relevance scoring, breaking coarse search steps into ranked sub-actions
  • Distance-aware context pruning, keeping the working set small as the search widens

The combination sets a new state of the art in function-match rate on SWE-bench Lite, and slots in ahead of existing repair agents as a drop-in localization stage.

Published at ICML 2025.