WSDM 2027

Complex Utility,
Simple Calibration

Decision Sufficiency in RAG Evidence Selection

Follow one archived decision

Start with the empty context and eight single passages. Compare the best measured context with the additive proposal, then see what changes after measuring the proposal.

Loading the archived cases…

Measured utility is centered reference-answer log likelihood per token. Token F1 and exact match come from the archived answers.

Across 1,440 matched settings

The main comparison uses eight candidate passages and a selection budget of three. Lower mean normalized regret is better.

Loading the archived comparison…

For greedy answers, the paired token F1 gain over P1 is 0.038, with a 95% interval of [0.020, 0.056]. The interval for the gain over BO9 includes zero.

Reproduce the paper on CPU

The frozen v1.0.0 archive includes the numerical inputs, replay code and reference outputs. The published verification matched 22 result CSVs and all four manuscript table bodies.