Toward collective intelligence: evolutionary pressures for cooperative language models

Many of life's major evolutionary transitions arose when previously independent units learned to coordinate at a larger scale, from the endosymbiotic origins of eukaryotic cells to increasingly complex animal societies. Inspired by this pattern, we ask whether language models can be trained across widening scales of cooperation, from individual interests to coordination with partners and groups. We translate three of Nowak's mechanisms for the evolution of cooperation --- direct reciprocity, indirect reciprocity, and group selection --- into reinforcement-learning pressures at progressively wider scales of the self. Preliminary results show that training changes not simply how much a model cooperates, but with whom: cooperation with reciprocating partners rises while exploitation falls. These effects hold across Donor game conditions, track the mechanisms, and transfer to an unseen Prisoner's Dilemma. Transfer to broader benchmarks is modest, with positive trends. These results suggest that evolutionary pressures are a promising avenue to induce robust cooperation in LLM systems.