Adaptive Meta-Curriculum for Test-Time Self-Improvement

Recursive self-improvement (RSI) in large language models has shown promise through test-time compute scaling, but current approaches allocate computational resources uniformly across problems, ignoring heterogeneous difficulty and the varying effectiveness of different improvement strategies. We introduce Adaptive Meta-Curriculum for Test-Time Self-Improvement (AMC-TSI), a framework that meta-learns both a curriculum scheduler to predict compute-benefit trade-offs and an adaptive improvement operator that selects among revision, search, and reflection strategies based on problem characteristics. Our key contributions are: (1) a theoretical framework proving that adaptive compute allocation under learned curricula achieves superior sample complexity bounds compared to uniform allocation, (2) a meta-learning algorithm that jointly optimizes curriculum difficulty estimation and improvement operator selection, and (3) empirical validation showing 2.3$\times$ improved compute efficiency and 18.7% higher accuracy on mathematical reasoning benchmarks. AMC-TSI addresses a critical gap in RSI systems by enabling efficient, problem-aware self-improvement without manual tuning.