Short papers
4–5 pages in NeurIPS style; page limits exclude references and appendices.
Machine Learning/CORE Unranked
Pre-to-Post NeurIPS 2026
NeurIPS 2026 Workshop: Transitioning from Pre-Training to Post-Training
The Pre-to-Post NeurIPS 2026 workshop explores the scientific understanding of the transition between pre-training and post-training stages in large language model development. It aims to move beyond folklore by examining how pre-training decisions influence post-training success, the mechanics of methods like fine-tuning and reinforcement learning, and how to predict or prevent failure modes in the training pipeline.
Paper fit
Contribution paths
A strong submission should clearly identify its contribution and evaluate it appropriately.
Long papers
Follow the main-conference page limit in NeurIPS style; page limits exclude references and appendices.
Research areas in scope
01
Foundations laid during pre-training
How data mixtures, curricula, continued or mid-training, learning-rate decay, and other late-stage pre-training decisions shape downstream capabilities.
02
The mechanics of post-training
Comparisons across supervised fine-tuning, reinforcement learning from human or AI feedback, reinforcement learning with verifiable rewards, and distillation; how these methods sharpen, broaden, suppress, or reorganize capabilities.
03
The development of model behaviors across training
Identifying when alignment, reasoning, instruction following, refusal, persona, and other behaviors emerge during specific stages of post-training, and distinguishing changes created by post-training from capabilities already present after pre-training.
04
Interactions between pre-training and post-training data
How particular pre-training data mixtures, domains, curricula, or objectives make subsequent post-training more or less effective; whether post-training outcomes depend on related knowledge, behaviors, or representations being established during pre-training.
05
Failure modes and fundamental limits
Mode or entropy collapse, reward hacking, capability forgetting, alignment taxes, and theoretical or empirical limits on what post-training can recover or change.
06
Data and optimization across the training transition
Synthetic data, scaling laws for supervised, preference, and reinforcement-learning data, optimizer-state inheritance, learning-rate schedules, regularization, and curriculum design across training stages.
07
Predicting post-training outcomes from pre-training
Developing metrics, representations, or behavioral signals during pre-training that forecast later trainability, alignment, robustness, and capability gains.
08
Reimagining the training pipeline
Folding traditionally post-training data and objectives into pre-training, jointly designing training stages, and allocating data and compute across the full pipeline.
09
Evaluation and open science
Evaluating post-training beyond benchmark improvements through causal experiments, standardized protocols, intermediate checkpoints, and openly reproducible training studies.
Policies worth checking twice
- Submissions are non-archival.
- Work already published at NeurIPS or other major ML conferences is not eligible.
- Each submission must nominate a reciprocal reviewer, who may be contacted to review if additional reviewers are needed.
- Page limits exclude references and appendices for both short and long papers.
- Submissions must be formatted in NeurIPS paper style.
Official sources
Compiled from the official call for papers. The organizers’ pages remain authoritative.
Last verified September 9, 2026