Theory of Mind and Self-Attributions of Mentality are Dissociable in LLMs

1. 1Investigate the relationship between Theory of Mind (ToM) and self-attributions of mentality in LLMs
2. 2Conduct safety ablation and mechanistic analyses of representational similarity to understand the impact of suppressing mind-attribution tendencies on ToM
3. 3Analyze the results to determine if ToM and self-attributions of mentality are dissociable in LLMs
4. 4Apply the findings to improve safety fine-tuning and socio-cognitive abilities in LLMs

AI researchers and engineers working on LLMs and safety fine-tuning can benefit from this research to improve model performance and safety, while ML researchers can apply these findings to develop more sophisticated socio-cognitive abilities in AI models

💡 Theory of Mind and self-attributions of mentality can be separated in LLMs, allowing for more targeted safety fine-tuning

💡 New research shows Theory of Mind & self-attributions of mentality are dissociable in LLMs #AI #LLMs