AutoAssumption: Searching over Assumptions Discovers ML-Systems Algorithms
arXiv:2611.14844v1 Announce Type: new Abstract: Many advances in ML systems came from breaking an assumption that earlier methods took for granted: speculative decoding dropped the assumption that a model needs one forward pass per token. Yet algorithms rarely state their assumptions, so experts and discovery harnesses search over programs, where one edit changes several design choices and gains are difficult to attribute. Searching over the assumptions themselves remains largely unexplored. Wri
阅读原始来源
来源记录
项目分析
暂无五维分析，可先查看上方来源资料。