AI & ML interests
None defined yet.
Recent Activity
Papers
View all Papers Organization Card
Algorithms and Optimization Group @ Institute of Theoretical Computer Science, ETH Zurich. Led by Rasmus Kyng and Maximilian Probst Gutenberg. More at kyng.inf.ethz.ch.
Latest release
AuXSpec: Learning to Autoformalize Program Specifications via Executable Postconditions
AuXSpec pairs Lean specifications with executable counterparts, so test cases can check generated specifications by execution, which unlocks an execution-based reward for RL on formal specification tasks. Trained with this reward, AuXSpec-27B-RL improves SpecGen pass@4 on AuXSpec-Test 14.3% → 84.3%, outperforms all frontier models and surpasses the open-source state of the art on Verina.
Paper · Website · AuXSpec-27B-RL · AuXSpec-9B-RL · Dataset
AuXSpec: Learning to Autoformalize Program Specifications via Executable Postconditions