Ask, Don't Equip: Question-Driven Coding Agents for Deep Kernel Optimization
KernelSeek reached 1.13× FlashAttention-3’s speed where a tool-rich autonomous agent plateaued at 0.50×.
The authors report that comparison used the same task, evaluator, hardware and week. KernelSeek guides a general-purpose coding agent with questions at decision points, initially from a human and later from a question bank it applies to itself. Its standalone CUDA attention kernels passed all 182 test scenarios per architecture, averaging 1.16× FlashAttention-3 on Hopper and 1.46× FlashAttention-4 on Blackwell by geometric mean. Across the recorded work, the authors say every large gain followed a question that shifted the agent’s focus or challenged a claimed limit.
HF Daily Papers' note
The authors report that comparison used the same task, evaluator, hardware and week. KernelSeek guides a general-purpose coding agent with questions at decision points, initially from a human and later from a question bank it applies to itself. Its standalone CUDA attention kernels passed all 182 test scenarios per architecture, averaging 1.16× FlashAttention-3 on Hopper and 1.46× FlashAttention-4 on Blackwell by geometric mean. Across the recorded work, the authors say every large gain followed a question that shifted the agent’s focus or challenged a claimed limit.
HF Daily Papers' note
score 5