Specify the target character for powerful language models

Characterize the normative persona or "character" that powerful language models should embody to reduce misalignment risks, including defining desirable traits, values, and behavioral commitments that should be instilled through pretraining and post-training.

Background

The paper motivates alignment pretraining as shaping a model’s initial persona before post-training. However, it remains unsettled what target character attributes powerful LLMs should have for safety and reliability.

Identifying this target would inform data curation, constitutional principles, and evaluation frameworks for both pretraining and post-training.