KACE: Knowledge-Adaptive Context Engineering for Mathematical Reasoning

In three linesKACE decouples storage from usage in context for mathematical reasoning. An epistemic tree stratified by difficulty and domain is built offline via self-reflective loop. At evaluation, tiered self-consistency dynamically classifies problems and selectively retrieves matching cards. On AIME 2025: 62.2% accuracy (+10.4 points vs Best-of-5).

Summary generated by Claude — human-verified