What Drives LLM Self-Reflection? A Controlled Ablation of Uncertainty Routing in Armed Conflict Forecasting
The episode analyzes a paper testing LLM self-reflection in armed conflict forecasting. Researchers found that simply asking diagnostic questions or using specialized vocabulary does little to improve accuracy. The core finding is that performance gains are driven by 'uncertainty routing'—mapping specific types of uncertainty to different actions—not the act of reflection itself.