EMNLP Findings2025
Instability in Downstream Task Performance During LLM Pretraining
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.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
1 top-tier paper has cited it since. Know what they found before your reviewers do.
Lune can draw on
the full paper · the top-tier papers that cite it · related work
Free to start. No credit card required.
Lineage
- Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling cites Training Trajectories of Language Models Across Scales
- Training Trajectories of Language Models Across Scales cites Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling
Cited by top-tier papers1
Builds on6
- Measuring Massive Multitask Language UnderstandingDan Hendrycks, Collin Burns, Steven Basart, Andy Zou et al.ICLR 2021 · 7,905 citations
- Pythia: A Suite for Analyzing Large Language Models Across Training and ScalingStella Biderman, Hailey Schoelkopf, Quentin Gregory Anthony, Herbie Bradley et al.ICML 2023 · 1,822 citations
- Training Trajectories of Language Models Across ScalesMengzhou Xia, Mikel Artetxe, Chunting Zhou, Xi Victoria Lin et al.ACL 2023 · 12 citations
- A Method for Building a Commonsense Inference Dataset based on Basic EventsKazumasa Omura, Daisuke Kawahara, Sadao KurohashiEMNLP 2020 · 7 citations
- COMET: A Neural Framework for MT EvaluationRicardo Rei, Craig Stewart, Ana C. Farinha, Alon LavieEMNLP 2020 · 6 citations
Related papers
- Post-Hoc Reversal: Are We Selecting Models Prematurely?Rishabh Ranjan, Saurabh Garg, Mrigank Raman, Carlos Guestrin et al.NeurIPS 2024 · 8 citations
- Safeguard Fine-Tuned LLMs Through Pre- and Post-Tuning Model MergingHua Farn, Hsuan Su, Shachi H. Kumar, Saurav Sahay et al.EMNLP Findings 2025 · 3 citations
- Data Curation Alone Can Stabilize In-context LearningTing-Yun Chang, Robin JiaACL 2023 · 11 citations
- Free Lunch: Robust Cross-Lingual Transfer via Model Checkpoint AveragingFabian David Schmidt, Ivan Vulic, Goran GlavasACL 2023 · 3 citations
- Test-Time Self-Adaptive Small Language Models for Question AnsweringSoyeong Jeong, Jinheon Baek, Sukmin Cho, Sung Ju Hwang et al.EMNLP Findings 2023 · 5 citations
