AdaExplore: Search for Efficient Kernel Generation
Article URL: https://stiglidu.github.io/AdaExplore/ Comments URL: https://news.ycombinator.com/item?id=47943943 Points: 1 # Comments: 0
Article URL: https://stiglidu.github.io/AdaExplore/
Comments URL: https://news.ycombinator.com/item?id=47943943
Failure-Driven Adaptation and Diversity-Preserving Search for Efficient Kernel Generation
Weihua Du · Jingming Zhuo · Yixin Dong · Andre He · Weiwei Sun · Zeyu Zheng Manupa Karunaratne · Ivan Fox · Tim Dettmers · Tianqi Chen · Yiming Yang · Sean Welleck
Carnegie Mellon University University of Washington Arm Ltd.
Kernel generation is hard for LLMs for two reasons. First, correctness has a sharp feasibility boundary : even small mistakes in syntax, memory access, or parallelization often cause compilation failures or runtime errors. Kernel generation is also underrepresented in LLM pretraining data, so models tend to repeat the same mistakes across tasks instead of learning from them. Second, performance lives on a rugged optimization landscape : meaningful speedups usually require coordinated structural changes, such as different tiling, parallelization, or fusion strategies, rather than small local edits. A search that only patches the current best often gets trapped in a weak strategy, while a search that only regenerates from scratch never fully refines promising candidates.
AdaExplore addresses both challenges directly. The Adapt stage builds a cross-task memory of recurring failures, helping the agent avoid mistakes it has already seen. The Explore stage runs a tree search that alternates between small steps for local refinement and large steps for structural regeneration, while keeping multiple promising branches alive in parallel. Together, these two stages handle both the sharp feasibility boundary and the highly non-linear performance landscape.