Extending SMT Solving with Non-Ground Clause Learning
This research contributes to the theoretical foundations of SMT solving, which is a core component of many automated reasoning and verification tools.
This research contributes to the theoretical foundations of SMT solving, which is a core component of many automated reasoning and verification tools.
Models can represent logical validity internally without reliably expressing it in outputs, affecting how we evaluate AI reasoning capabilities.
这篇论文测了语言模型会不会按给定的模态逻辑规则推理,DeepSeek V4 Flash开推理模式后从4.4%直接飙到88.1%,反差很大,可以看看。
COCOLogic-V2 这个新数据集专测视觉推理,正反例分类特别细,模型在近边界上直接翻车,做可解释 AI 的可以看看。