Code2Skill: code-derived skill bank lifts agent benchmarks 11.7%
Code2Skill mined 1,006,822 atomic operations from 19,769 GitHub projects into the verified CodeSkillBank. Agents' macro-average score rose from 42.90 to 47.90 (+11.7%), with gains in 57 of 72 protocol-matched evaluations across nine model settings and eight benchmarks.
- CodeSkillBank holds 1,006,822 records from 19,769 GitHub projects
- Agent macro-average score rose from 42.90 to 47.90 (+11.7%)
- Gains in 57 of 72 protocol-matched evaluations across 9 models and 8 benchmarks
- Code2Skill beat the trajectory-bank oracle by 9.5 points (49.5 vs 40.1)
Read next
Business