arXiv:2605.27832v1 Announce Type: new 
Abstract: Large Language Models (LLMs) are being applied to increasingly difficult problems and use cases. To navigate their vast solution spaces effectively, LLMs need to be creative. Yet the subjective nature of creativity and the limits of human judgment make training LLMs for creativity especially challenging. As a solution, we train LLMs on Codenames, a word-association game that exercises the two central axes of creativity, divergent and convergent thinking, while yielding objectively verifiable outcomes. This verifiability lets us bypass human judgment and train with Reinforcement Learning with Verifiable Rewards (RLVR). We train Qwen3-1.7B, 4B, and 8B models and evaluate them on ten creativity and four reasoning benchmarks. We find that the precision-diversity trade-off is scale-dependent: the 8B model prioritizes creativity over precision, while the 1.7B and 4B models gain reasoning precision at the cost of creativity. Concretely, the 8B model shows modest but consistent creativity gains (8 of 10 benchmarks) with only minor reasoning degradation, whereas the smaller models achieve substantial gains on reasoning tasks. Our study presents a scalable and effective solution to train LLMs for creativity.

تم تسليط الضوء على التقدم الأخير في تدريب نماذج اللغة الكبيرة (LLMs) من خلال تطبيق التعلم المعزز مع المكافآت القابلة للتحقق (RLVR) في لعبة ارتباط الكلمات المعروفة باسم Codenames. تهدف هذه الطريقة إلى تعزيز الإبداع في نماذج LLMs، وخاصة نماذج Qwen3، من خلال التركيز على التفكير المتباين والمتقارب مع تجاوز الحكم البشري الذاتي.

Los recientes avances en el entrenamiento de Modelos de Lenguaje Grande (LLMs) se han destacado a través de la aplicación del Aprendizaje por Refuerzo con Recompensas Verificables (RLVR) en un juego de asociación de palabras llamado Codenames. Este enfoque busca mejorar la creatividad en los LLMs, específicamente en los modelos Qwen3, al centrarse en el pensamiento divergente y convergente, eludiendo el juicio humano subjetivo.

Des avancées récentes dans la formation des grands modèles de langage (LLMs) ont été mises en évidence grâce à l'application de l'apprentissage par renforcement avec des récompenses vérifiables (RLVR) dans un jeu d'association de mots appelé Codenames. Cette approche vise à améliorer la créativité des LLMs, en particulier les modèles Qwen3, en se concentrant sur la pensée divergente et convergente tout en contournant le jugement humain subjectif.

Recent advancements in training Large Language Models (LLMs) have been highlighted through the application of Reinforcement Learning with Verifiable Rewards (RLVR) in a word-association game called Codenames. This approach aims to enhance creativity in LLMs, specifically the Qwen3 models, by focusing on divergent and convergent thinking while bypassing subjective human judgment.

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

Was this article worth reading? Share it

Ready to build your own newsroom?