Quazi Ishtiaque Mahmud

Multi-turn RL with Structural and Performance Aware Rewards for CUDA Kernel Generation

Quazi Ishtiaque Mahmud, Nesreen K. Ahmed, Ali Jannesari

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a powerful technique to enhance the reasoning capacity of LLMs for optimized code generation. However, existing RLVR approaches primarily rely on outcome-based signals such as correctness and speedup, overlookin…Read more