POLIS Framework Boosts LLM Reasoning via Cumulative Cultural Evolution

A new framework called POLIS (Population Orchestrated Learning and Inference Society) enables large language models to improve through interaction-driven cumulative intelligence, mimicking human cumulative cultural evolution. In POLIS, multiple LLM agents generate solutions, verify each other's outputs, and retain validated artifacts in shared cultural memory, which is then internalized via parameter updates. On mathematical reasoning benchmarks, populations of 1–4 billion parameter models achieved average gains of 8.8–18.9 points over base models, narrowing the performance gap to 70B+ parameter monolithic models. Mechanistic ablations identified peer verification as the key ratchet operator, with internalization sustaining accumulation across rounds. The research provides computational evidence that interaction-driven processes can enhance LLM capabilities without relying solely on static corpora or parameter growth.