ABSTRACT
This paper introduces AdaSched, a framework that decouples textual gradient computation from prompt updates in large language models, enhancing convergence speed and final performance.
PAPER · PDF
Loading PDF...
Key findings
AdaSched achieves up to 23% improvement in convergence speed and 15% better final performance compared to baseline textual gradient methods.
AdaSched reduces API call variance by 34%.
The framework introduces a dual-phase optimization process with gradient accumulation and adaptive scheduling mechanisms.
Limitations & open questions
The study focuses on specific benchmarks and may require further validation across a broader range of tasks.
The theoretical analysis of convergence properties may need to be extended to more complex scenarios.