[1/n] Can LLMs instantiate computations of mental disorders—not just talk like they’re “depressed” or “anxious”? In our new paper, “Emergence of psychopathological computations in large language models”, we show that: • Structure: LLMs encode computational structures of psychopathology that can be mechanistically measured and intervened on • Function: Executing the computations leads LLMs to engage in problematic behaviors • Emergence: As LLMs scale, these dysfunctional computations become more structured and more effective This suggests that some “psychopathology-like” behaviors in LLMs may reflect internal computations, not mere superficial mimicry—raising both new AI safety risks concerns and promise for powerful, new tools for psychiatric research. 🔗 Full paper: arxiv.org/abs/2504.08016v2 #AI #LLM #AISafety #AIEthics #ComputationalPsychiatry #MechInterp #SAE #Psychology #MentalHealth #mentaldisorder #Emergence

🚀Call for Papers: PAKDD 2025 Workshop on Graph Learning with Foundation Models - Venue: Sydney, Australia - Deadline: February 22, 202 - Website: pakdd-glfm-2025.github.io We welcome submissions on graph learning with foundation models, including both full and short papers!

🚀Call for Papers: PAKDD 2025 Workshop on Graph Learning with Foundation Models - Venue: Sydney, Australia - Deadline: February 22, 202 - Website: pakdd-glfm-2025.github.io We welcome submissions on graph learning with foundation models, including both full and short papers!

🚀Call for Papers: PAKDD 2025 Workshop on Graph Learning with Foundation Models - Venue: Sydney, Australia - Deadline: February 22, 202 - Website: pakdd-glfm-2025.github.io We welcome submissions on graph learning with foundation models, including both full and short papers!