Dynamic Sampling that Adapts: Self-Aware Iterative Data Persistent Optimization for Mathematical Reasoning

Researchers introduce SAI-DPO, a dynamic data sampling framework that adapts training data selection based on a model's evolving capabilities during training, rather than using static metrics. Tested on mathematical reasoning benchmarks including AIME24 and AMC23, SAI-DPO achieves state-of-the-art performance with significantly less training data, outperforming baselines by nearly 6 points.

SAI-DPO addresses a fundamental inefficiency in machine learning training pipelines: the mismatch between static data selection strategies and models' dynamic learning trajectories. Traditional approaches rely on fixed, externally-defined metrics that fail to account for how model capabilities evolve during training, resulting in wasted computational resources on irrelevant or overly difficult examples. The framework introduces two novel metrics—Knowledge Semantic Alignment and Self-Aware Difficulty—that measure domain weaknesses and instance complexity relative to the model's current state, enabling real-time recalibration of training distributions.

This research builds on growing recognition within the AI community that data quality and relevance matter more than sheer data quantity. As models scale, the ability to train efficiently becomes increasingly valuable. Mathematical reasoning tasks provide an ideal testing ground because success is objectively measurable and domains of weakness are identifiable through pass rates and reasoning path analysis.

For AI practitioners and organizations, SAI-DPO's results have practical implications: achieving comparable or superior performance with substantially less training data reduces computational costs, training time, and environmental impact. The framework's demonstrated effectiveness across eight diverse benchmarks suggests broader applicability beyond mathematical reasoning to other specialized domains requiring expert-level reasoning.

The methodology opens avenues for further optimization in adaptive curriculum learning and reinforcement learning pipelines. Future work may explore how these self-aware sampling principles scale to larger models and datasets, and whether similar dynamic adaptation techniques transfer to other problem domains requiring iterative model improvement.

• →SAI-DPO dynamically aligns training data with a model's current capabilities rather than using static selection criteria, improving training efficiency.
• →The framework achieved up to 6-point improvements over baselines on mathematical reasoning benchmarks while using significantly less training data.
• →Two novel metrics—Knowledge Semantic Alignment and Self-Aware Difficulty—enable real-time assessment of data relevance to the model's evolving state.
• →Results across eight benchmarks including AIME24 and AMC23 suggest the approach generalizes effectively to diverse mathematical reasoning tasks.
• →Dynamic data sampling strategies reduce computational cost and training time while maintaining or improving model performance.