Dutch Books for Language Models
The paper tests whether language-model probabilities can be arbitraged for guaranteed profit.
Andrews and Sarkar elicit forecasts from language models on events generated from stock-return data, then use linear programs to calculate the largest Dutch-book profit against those probabilities. They treat that guaranteed profit as a measure of incoherence, without needing the real-world outcomes to have resolved. The authors report substantial incoherence, worse when event relationships are more logically complex. They also find irrelevant contextual details can raise incoherence by an order of magnitude. ArXiv · AI/CL/LG's note
Andrews and Sarkar elicit forecasts from language models on events generated from stock-return data, then use linear programs to calculate the largest Dutch-book profit against those probabilities. They treat that guaranteed profit as a measure of incoherence, without needing the real-world outcomes to have resolved. The authors report substantial incoherence, worse when event relationships are more logically complex. They also find irrelevant contextual details can raise incoherence by an order of magnitude. ArXiv · AI/CL/LG's note
score 5