Instability in Downstream Task Performance During LLM Pretraining

2025-10-06

AI Summary

Downstream task performance during large language model (LLM) pretraining exhibits significant instability across training checkpoints, hindering reliable identification of the optimal checkpoint. This work systematically characterizes this instability phenomenon and proposes two lightweight, plug-and-play post-training methods: sliding-window checkpoint averaging and adjacent-checkpoint ensembling. Both methods require no modification to the training pipeline or additional computational overhead—only leveraging already-saved intermediate checkpoints. Theoretical analysis demonstrates that these approaches reduce prediction variance by smoothing stochastic fluctuations in parameter space. Extensive experiments across diverse LLM scales (e.g., 1B–7B) and downstream tasks (e.g., GLUE, MMLU, ARC) confirm that both methods substantially improve stability—reducing performance standard deviation by an average of 35%—and yield more robust and reliable final model selection. This work establishes a simple, efficient, and theoretically grounded paradigm for enhancing stability in LLM pretraining.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

When training large language models (LLMs), it is common practice to track downstream task performance throughout the training process and select the checkpoint with the highest validation score. However, downstream metrics often exhibit substantial fluctuations, making it difficult to identify the checkpoint that truly represents the best-performing model. In this study, we empirically analyze the stability of downstream task performance in an LLM trained on diverse web-scale corpora. We find that task scores frequently fluctuate throughout training, both at the aggregate and example levels. To address this instability, we investigate two post-hoc checkpoint integration methods: checkpoint averaging and ensemble, motivated by the hypothesis that aggregating neighboring checkpoints can reduce performance volatility. We demonstrate both empirically and theoretically that these methods improve downstream performance stability without requiring any changes to the training procedure.

Problem

• LLM pretraining causes unstable downstream task performance
• Checkpoint selection is difficult due to performance fluctuations
• Methods are needed to stabilize performance without changing training

Innovation

• Checkpoint averaging reduces performance volatility
• Ensemble methods stabilize downstream task metrics
• Post-hoc integration without training procedure changes