Djiré Albérick Euraste

Code LLMs don't just memorize training data – some generalize far better than others, and even "leaky" datasets like CVEFixes show surprisingly low memorization advantage.