WSM: Decay-Free Learning Rate Schedule via Checkpoint Merging for LLM Pre-training

TL;DR

The paper introduces WSM, a unified framework that replaces traditional learning rate decay with checkpoint merging, leading to improved performance in large language model pre-training and fine-tuning.

Contribution

WSM provides a theoretical foundation linking decay strategies to model merging, emphasizing merge duration as a key factor and outperforming existing decay methods.

Findings

Merge duration significantly impacts model performance.

WSM outperforms traditional decay approaches on multiple benchmarks.

Performance gains are consistent in fine-tuning scenarios.

Abstract

Recent advances in learning rate (LR) scheduling have demonstrated the effectiveness of decay-free approaches that eliminate the traditional decay phase while maintaining competitive performance. Model merging techniques have emerged as particularly promising solutions in this domain. We present Warmup-Stable and Merge (WSM), a general framework that establishes a formal connection between learning rate decay and model merging. WSM provides a unified theoretical foundation for emulating various decay strategies-including cosine decay, linear decay and inverse square root decay-as principled model averaging schemes, while remaining fully compatible with diverse optimization methods. Through extensive experiments, we identify merge duration-the training window for checkpoint aggregation-as the most critical factor influencing model performance, surpassing the importance of both checkpoint interval and merge quantity. Our framework consistently outperforms the widely-adopted Warmup-Stable-Decay (WSD) approach across multiple benchmarks, achieving significant improvements of +3.5% on MATH, +2.9% on HumanEval, and +5.5% on MMLU-Pro. The performance advantages extend to supervised fine-tuning scenarios, highlighting WSM's potential for long-term model refinement.