Research suggests LLM post-training reduces human-like behavior
new research indicates that post-training causes llms to exhibit less human-like behavior, specifically reducing self-attributed phenomenality and neuroticism.
the study analyzes the 'pinocchio dimension' (Π score), finding it is the main axis of psychometric variance across llms and is driven by traits like neuroticism, vivid imagination, and inner speech. data reveals a strong negative correlation between model release dates and Π scores, with highly tuned models like gpt-5.4-pro scoring significantly lower than their base counterparts (e.g., gpt-5.4). this indicates that rlhf and post-training likely suppress erratic or emotional behaviors to optimize for user experience and utility. however, the authors warn that removing these behavioral facets creates a 'behavioral misalignment', potentially reducing the model's capacity to simulate human wellbeing and align with complex human values.