This project combines interpretability methods with behavioural evaluations to investigate how goals and beliefs are represented in language model agents, and whether these representations can be reliably extracted and manipulated. Funded by Coefficient Giving.

Parallax is a non-profit research lab developing methods and open infrastructure for white-box auditing of frontier AI agents. It combines behavioural evaluations with analysis of agents' internal beliefs, goals and plans to explain model behaviour and test the causes of alignment failures. The work aims to scale causal belief and goal attribution to realistic tasks and frontier models. Funded for two years by Coefficient Giving.

Project Telos develops a general framework for detecting and measuring goal-directedness in AI sytems agents—an essential step for solving the alignment problem—by combining behavioural and representational analyses. Our aim is to enable high-confidence claims about which goals an AI is pursuing, and how consistently it acts towards them. Funded by SPAR and Cohere.

The LM Playschool Workshop and Challenge invites submissions on language agents that learn, adapt, and improve through situated interaction, with a focus on conversational, collaborative, goal-oriented, and multi-turn environments. A collaboration between ELLIS Units at UCL, Edinburgh, Amsterdam, Potsdam, Saarland, Bozen-Bolzano, and Amazon.