@mircomusolesi @ai_ucl We also point to open research questions in this area, including ones relating to defining rationality for generative AI, evaluation metrics that incorporate multiple dimensions, human-AI interaction in the presence of irrationality, and rationality in multi-agent systems.

@mircomusolesi @ai_ucl The "rationality entanglement” in generative AI highlights a conflict between the training objectives and the behaviour we expect from existing models. We can view this as a tension between predictive/generative and epistemic rationality.

@mircomusolesi @ai_ucl We discuss how ‘irrational’ behaviour can at times be optimal under certain constraints, particularly over longer-term horizons. We also survey existing methods that can be applied to identify and interact with potentially irrational agents.

@mircomusolesi @ai_ucl Fields like economics, philosophy, psychology have shaped the conception of rationality in AI. Each has a distinct understanding of the concept, leading to the lack of a unified definition in AI.

🚀 Just published! Our paper “(Ir)rationality in AI: state of the art, research challenges & open questions” is out in Artificial Intelligence Review. We dive into what rationality means in AI and how to handle irrational agents 🧵 https://t.co/Yj3yX1W5ZO @mircomusolesi @ai_ucl

Proud to see our work featured in print! Our article on agent-based modelling of civil conflict is highlighted in an analysis of game theory in conflict modelling. @mircomusolesi @AkinUnver @Epsiloon_mag article: https://t.co/ySD8aGccGS Full paper: https://t.co/IG9WQCMabV

Two papers out on the same day! Using tasks from cognitive psychology we evaluate 7 LLMs and show that they display irrationality, but not in the same way humans do. They are often incorrect in ways that differ from human-like biases. @mircomusolesi https://t.co/4bEfDKNcgy

Very excited to see this paper out! We use an agent-based model (ABM) rooted in game theory to study the microdynamics of conflict in northeastern Syria. We focus on retaliation and reciprocity between ISIS and Kurdish forces. @AkinUnver @mircomusolesi https://t.co/QOcQpXEnBN

The event data was so dense, that it allowed a range of game-theoretical and ABM applications to better understand iterated reciprocity and retaliation dynamics between non-state armed groups. So @olivia_macs, @mircomusolesi, and I did just that.

Finally happy to see our new paper out. Remember the 'Siege of Kobani' nearly a decade ago? It turns out it is still the best-documented dyadic extended micro-level conflict event we have in any major conflict dataset. https://t.co/GSvetBOZVS

Had the chance to present a working paper on the spread of protest onset at the UCL Conflict & Change workshop. Such a great chance to get feedback, and an incredibly collaborative environment. Thank you @uclconflict for a great workshop!

Machine Psychology - models are biased like us, but are they as irrational🤪? @olivia_macs @mircomusolesi https://t.co/q2fz1SPUCu

Do LLMs reason rationally? With @mircomusolesi we explore this question in our new preprint, using methods from cognitive psychology to assess the reasoning of LLMs. Check it out at: https://t.co/LqYVXiQh43

What better way to start 2024 than by reliving one of the highlights of 2023: the first ever Cooperative AI Summer School! For those who couldn't attend in person, recordings of all lectures are *finally* available online: https://t.co/0ZJ3O59g5V. More details in the thread...

🚨New paper 🚨 What does it mean to create rational machines? We survey the role of rationality and irrationality within artificial intelligence, and discuss research challenges and open questions in this area. Co-author: @mircomusolesi https://t.co/R7mvKAAMJB