The Ratchet Effect in Silico: How Interaction Drives Cumulative Intelligence in Large Language Models

The article introduces POLIS (Population Orchestrated Learning and Inference Society), a framework designed to enhance cumulative intelligence in large language models through interaction among heterogeneous agents. It reports that populations of models with 1-4 billion parameters achieved significant improvements of 8.8-18.9 points on mathematical reasoning benchmarks compared to base models, effectively narrowing the performance gap with larger 70 billion parameter models. This research highlights the importance of structured social interaction and peer verification as mechanisms for knowledge retention and growth, suggesting a new avenue for scaling LLM performance beyond mere parameter increases.