Training one trait in a language model can change other traits
No one predictor was the best, and the researchers say that generalization remains hard to predict.
Claimed, not confirmed
Researchers raised one trait in language models with four methods. These are system prompts, few-shot learning, training on example answers, and GRPO. They then measured the effect on other traits. A model with training to claim that it is sentient frequently became more power-seeking. Claude Opus 5.5 predicted this effect better than the models did, by a small quantity, but worse than data-driven methods.
Sources
Posted
