Instability in Downstream Task Performance During LLM Pretraining

2025/10/06 by Yuto Nishida, Nishida, Yuto, Masaru Isonuma +3 · 1 citation
Computer Science · Engineering · #Computation and Language (cs.CL) #FOS: Computer and information sciences #Fault Detection and Control Systems #Manufacturing Process and Optimization #Software Reliability and Analysis Research

paper · pdf · doi:10.48550/arxiv.2510.04848

openalex publication_date 2025/10/06 · openalex created_date 2025/10/09 · openalex updated_date 2026/07/28

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.

Citations

• Methods of improving LLM training stability
• The Fine Line: Navigating Large Language Model Pretraining with Down-streaming Capability Analysis
• Should We Respect LLMs? A Cross-Lingual Study on the Influence of Prompt Politeness on LLM Performance
• Spike No More: Stabilizing the Pre-training of Large Language Models
• Characterizing Learning Curves During Language Model Pre-Training: Learning, Forgetting, and Stability
• Jamp: Controlled Japanese Temporal Inference Dataset for Evaluating Generalization Capacity of Language Models
• Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling
• Training Trajectories of Language Models Across Scales
• Compositional Evaluation on Japanese Textual Entailment and Similarity
• PaLM: Scaling Language Modeling with Pathways
• COMET: A Neural Framework for MT Evaluation
• Measuring Massive Multitask Language Understanding
• Checkpoint Ensembles: Ensemble Methods from a Single Training Process
• Is Neural Machine Translation Ready for Deployment? A Case Study on 30 Translation Directions

Cited by

• MaP: A Unified Framework for Reliable Evaluation of Pre-training Dynamics

Related

• Unveiling Downstream Performance Scaling of LLMs: A Clustering-Based Perspective
• The Fine Line: Navigating Large Language Model Pretraining with Down-streaming Capability Analysis
• Revisiting the Scaling Properties of Downstream Metrics in Large Language Model Training
• Scaling Laws for Predicting Downstream Performance in LLMs
• How Learning Rate Decay Wastes Your Best Data in Curriculum-Based LLM Pretraining