A Comrade IA to do math: he is seen as a peer when he hesitates and gets a little wrong, not when he approves everything

Asked about an agent with a personality inspired by the Big Five, experts first judge him on his social behavior. Too much acquiescence or instructions, and he becomes a guardian again.

Chatbots are increasingly being tested as learning partners in math, but what makes them seen as peers rather than professors? One team designed an LLM-based peer with a Big Five personality and an evolving motivational state, then submitted it to experts in the field before analyzing their interviews.

The first lesson was that the participants judged him on his social behaviour much more than on his mathematical skills. The agent was seen as a peer when he created a space without judgment, reasoned hesitantly, showed well-balanced imperfections and respected the rhythm of each person. Conversely, systematic approval and excessive instructions undermined confidence and made him fall into the role of tutor.

The authors draw from this human-centred design ideas — emotional safety, shared control, collaborative reasoning — for future peer-to-peer learning systems.