Evaluates how well frontier models interpret what people feel and want in conversation using participant-labeled conversations.
AttuneBench evaluates emotional intelligence in language models using 200 multi-turn conversations between people and AI models. Participants annotated their own emotions, the model’s conversational behavior, and which responses they preferred. The benchmark measures several dimensions, including emotion tracking, response preferences, and conversation-wide understanding, rather than treating emotional intelligence as a single score. Its repository provides code to run and score evaluations, while the public site presents model results and analysis. The benchmark was developed through collaboration between Pareto AI and Thoughtful Lab.
Be the first Cosign
Add it to your Products I use list. You can also share why you use it.
