HLS-Seek: QoR-Aware Code Generation for High-Level Synthesis via Proxy Comparative Reward Reinforcement Learning

HLS-Seek proposes a QoR (Quality-Latency Tradeoff)-aware code generation framework based on agent-based reward reinforcement learning. This framework uses a model with only 7B parameters, achieving a syntax accuracy of 84.7% (pass@1) and a function accuracy of 81.4% (pass@5) on the HLS-Eval test set, with training speed 8.5 times faster than traditional real-world reward reinforcement learning. To improve efficiency and prevent reward hacking, the framework employs an agent-based reward model to avoid fully synthetic loops, achieving a Pareto advantage accuracy of 99.53%, and introduces a uncertainty-aware Monte Carlo Dropout switching mechanism. In QoR evaluations, HLS-Seek achieved the lowest latency on 19 out of 30 kernels and achieved Pareto dominance over the HLS-specific baseline on 9 kernels, with its function accuracy surpassing GPT-5.1.