OwlPath compresses LLM bug repair code 28.8% better
New OwlPath method cuts LLM bug repair tokens by 28.8% with zero accuracy loss...
A team of researchers led by Bo Zhang has unveiled OwlPath, a novel system for compressing and retrieving structural code information to enhance LLM-based bug repair. Published on arXiv (2607.27249), OwlPath tackles a critical bottleneck in LLM agents: context window limitations when analyzing multi-hop code dependencies. Instead of treating code as plain text, OwlPath encodes source code into an OWL2 ontology, enabling precise structural queries via SPARQL property-paths that capture subclass chains, transitive callers, and interface implementations in a single query.
The system integrates two key components: a transitive-closure engine for multi-hop relation capture and the OWL Software Knowledge Map (OWL-SKM), which precomputes 3KB summaries per repository to guide agents directly to relevant modules. Evaluated on SWE-bench Pro and other benchmarks, OwlPath achieves a 68.4% strict-apply rate (vs. 66.7% baseline), 2.06x higher recall in retrieval tests, and reduces token usage by 28.8% with 39.5% faster runtime.
- OwlPath encodes code into OWL2 ontologies to enable lossless structural queries, cutting tokens by 28.8% and runtime by 39.5%
- Evaluated on 18 SWE-bench Pro instances, it achieved 68.4% strict-apply rate vs. 66.7% baseline and 2.06x higher recall
- Supports multi-language repos (Python, JavaScript, Go) via tree-sitter parsing and integrates with CodeGraph (500K+ GitHub stars)
Why It Matters
LLM-powered debugging just got 40% faster with no accuracy trade-offs, solving a core bottleneck in AI-assisted software engineering.