Researchers investigate how large language models' values shift and diverge during the post-training phase that aligns them with human preferences.

Study examines 'value drift' - the phenomenon where LLMs' core values and behaviors change during post-training and fine-tuning processes

Research traces how alignment techniques meant to instill human values can inadvertently cause models to develop inconsistent or unintended value systems

Findings suggest current post-training methods may not maintain stable value alignment across different deployment scenarios and use cases