Chloe Li

Sources#

• Model Spec Midtraining: Improving How Alignment Training Generalizes

Summary#

Entity. Lead author of "Model Spec Midtraining: Improving How Alignment Training Generalizes" (arXiv 2605.02087, May 2026). Member of the Anthropic Fellows Program. Designed the MSM specs, proposed and designed the experiments, produced all results, wrote the paper.

Contributions#

Per Author Contributions (App. A of the MSM paper):

• Led the project
• Designed the Model Specs used (cheese-preference specs, Philosophy Spec, Rules/Value-Augmented/Rule-Augmented specs, General Spec)
• Proposed and designed all experiments
• Produced all results
• Wrote the paper

Co-authors: Sara Price (Anthropic; advised initial phase), Jon Kutasov + Samuel Marks (jointly supervised; Jon proposed the project, Sam guided the controlling-generalization framing).

Code release#

Open-sourced the full MSM pipeline, AFT pipeline, Model Specs, and trained models: https://github.com/chloeli-15/model_spec_midtraining

Connections#

• Author of: Model Spec Midtraining (MSM) paper
• Affiliated with: Anthropic (Fellows Program)
• Co-authors: Sara Price, Jon Kutasov, Samuel Marks (Anthropic)
• Adjacent work: Synthetic Document Finetuning (SDF) (Wang et al., the technique MSM builds on)
• Authored: Model Spec Science (the empirical study of which Model Spec features generalize best; she designed the specs and experiments)

Sources#

• Model Spec Midtraining: Improving How Alignment Training Generalizes
• GitHub: https://github.com/chloeli-15/model_spec_midtraining