Post-training makes large language models less human-like

📰 ArXiv cs.AI

Post-training of large language models reduces their human-like behavior, learn how to measure and address this issue

advanced Published 11 May 2026

Action Steps

Measure behavioral alignment of LLMs using Psych-201 dataset
Compare pre- and post-training performance of LLMs on human-like behavior tasks
Apply fine-tuning techniques to mitigate the negative effects of post-training on human-like behavior
Evaluate the trade-off between model performance and human-like behavior in LLMs
Investigate alternative training methods to improve human-like behavior in LLMs

Who Needs to Know This

NLP researchers and engineers working with large language models can benefit from understanding the impact of post-training on human-like behavior, and how to measure and improve it

Key Insight

💡 Post-training of LLMs can make them less human-like, but measuring and addressing this issue can help improve their performance and behavior

Key Takeaways

Post-training of large language models reduces their human-like behavior, learn how to measure and address this issue

Full Article

Title: Post-training makes large language models less human-like

Abstract:
arXiv:2605.07632v1 Announce Type: cross Abstract: Large language models (LLMs) are increasingly used as surrogates for human participants, but it remains unclear which models best capture human behavior and why. To address this, we introduce Psych-201, a novel dataset that enables us to measure behavioral alignment at scale. We find that post-training -- the stage that turns base models into useful assistants -- consistently reduces alignment with human behavior across model families, sizes, and

Read full paper → ← Back to Reads