Research
Bridge Mines 381,661 Java API Update Instances, and the Best LLM Recommends the Right Replacement Only 37.1% of the Time
Bridge automatically mines ecosystem-scale API update mappings and concrete client update instances, producing 381,661 Java instances covering 18,900 mappings across 2,557 libraries and 277,259 Python instances covering 4,456 mappings across 999 libraries. Validation reports 91.6% precision and 88.7% recall for Java, and 90.1% precision with 64.0% recall for Python. Evaluated on replacement-API recommendation against this data, the best LLM reaches only 37.1% accuracy for Java and 44.4% for Python, and models do markedly better on frequently observed mappings than uncommon ones, which is the memorization signature you would expect and a direct warning for agent-driven library migrations.
↳ Follow the thread