Fault-driven tests linked to higher scores for a code model
A preprint reports higher code benchmark scores after faulty-code test synthesis and dense rewards, while an audit found synthetic tests remained invalid.
Developing Light ยท https://developinglight.com/editorial/developing-light