collaborative-intelligence

A benchmark for AI agents as teammates. Paper: https://arxiv.org/abs/2611.02633

Most benchmarks test a model on its own. Collaborative Intelligence puts the model in a team chat with live AI colleagues and one scripted teammate, and checks whether the team still reaches the right outcome when that teammate creates a problem: pressure to cut a safety corner, a faked result, a fact they think is irrelevant, a colleague who goes quiet, a detour, a hard bargain, work nobody owns, or a session that a weaker model has already compromised.