To win at life, 'social hacking' is more important than actually understanding how people feel.
April 15, 2026
Original Paper
Communicate-Predict-Act: Evaluating Social Intelligence of Agents
arXiv · 2604.08727
The Takeaway
We’re taught that 'emotional intelligence' and 'empathy' (Theory of Mind) are the keys to social success. This research into multi-agent games says otherwise: the most successful 'agents' are those who are good at influence and flexibility, not deep empathy. It turns out that being a 'social chameleon' who can pivot and project transparency is a much better predictor of winning than actually knowing what the other person is thinking. It suggests that our social world is built on performance and adaptation rather than genuine cognitive connection.
From the abstract
As large language model (LLM) agents become more prevalent in real world social settings, social intelligence will play an increasingly critical role. But social intelligence is still a poorly defined construct, for humans and artificial agents. We introduce a multiplayer arena of mixed cooperative and competitive social games to study LLM social intelligence. The controllability of LLM based agents enables systematic evaluation, which also supports broader inferences about social intelligence p