Lonneke van der Plas

Training LLMs for reasoning warps their internal representations *away* from the neural signatures of human creativity, while fine-tuning on human behavior boosts alignment across the board.