Eliciting Chain-of-Thought in Base LLMs via Gradient-Based Representation Optimization

Open in new window