Apple ML Research unveils RSI-Inference Bench to test AI agents on model-serving tasks

Apple ML Research says RSI-Inference Bench measures how coding agents handle model-serving tasks, from adding support for new models to improving speed and fixing slowdowns. Its leaderboard ranks Anthropic's Sonnet-5.5 first overall, with a 26.7% pass@1 on new-model support and a 1.68× speed-up on optimization and regression-fix tasks.