Playing with Words, Improving with Rewards: Training Language Models for Creative Association

In three linesTraining Qwen models (1.7B, 4B, 8B) on Codenames game to improve creativity via Reinforcement Learning with Verifiable Rewards (RLVR). 8B model gains creativity (+8/10 benchmarks) with minor reasoning degradation, while smaller models prioritize precision. Study on creativity-precision trade-off across model scales.

Summary generated by Claude — human-verified