Reinforcing Agentic Creativity in Scientific Ideation with Night Science
The paper claims LLM creativity can be trained, not just sampled hotter.
AI Night-Scientist uses reinforcement learning to push models beyond predictable scientific-idea generation. The authors frame creativity across action, process, and outcome, then train with GRPO to vary how and when the model explores. They report broader proposal diversity, including 27.8% more research directions and 14.9% more contribution types than the base model. The abstract says temperature alone did not reproduce the gains; semantic guidance about the kind of creativity mattered. ArXiv · AI/CL/LG's note
AI Night-Scientist uses reinforcement learning to push models beyond predictable scientific-idea generation. The authors frame creativity across action, process, and outcome, then train with GRPO to vary how and when the model explores. They report broader proposal diversity, including 27.8% more research directions and 14.9% more contribution types than the base model. The abstract says temperature alone did not reproduce the gains; semantic guidance about the kind of creativity mattered. ArXiv · AI/CL/LG's note
score 5