What's this about?
Uses RL to train models to optimize code for performance — making it faster, not just correct.
Why it matters
Most code models target correctness; rewarding measured performance (6 upvotes) points models at the harder, more valuable goal of efficient code.