Two-Way Translation

Large language models routinely act on underspecified human instructions, filling gaps with their own priors while presenting the result with unwarranted confidence. The human cannot see the model's internal interpretation until after tokens — and money — are spent. This paper documents two protocols developed through