Wang, Z.; Ma, Y.; Xu, C. Eliciting Chain-of-Thought in Base LLMs via Gradient-Based Representation Optimization. AAAI 2026, 40, 33782-33790.