Call for Papers

Topics of interest

We solicit theoretical, empirical, and methodological work across the following areas. This list is not exhaustive.

• 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.

• 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.

• 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.

• 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.

• 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.

• 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.

• 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.

• 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.

• Evaluation and open science

  Evaluating post-training beyond benchmark improvements through causal experiments, standardized protocols, intermediate checkpoints, and openly reproducible training studies.