CodeI/O

Condensing reasoning patterns via code input/output prediction (ICML 2025 Oral): transforms real code into a natural-language CoT task — predict inputs/outputs given code and test cases — exposing models to universal reasoning primitives (logic-flow planning, state-space search, decision-tree traversal) while decoupling reasoning from code syntax. Consistently improves symbolic, scientific, logic, math, and commonsense reasoning; adopted as a pretraining/mid-training signal in later reasoning recipes.

Joint HKUST / DeepSeek-AI work with an HKUST student lead from Junxian He's group.