Research projects

WSDM 2027 · Accepted

Evidence selection in RAG

When a model proposes a set of passages, is it worth measuring that set once before returning it? The study compares that check with choosing the best context already measured.

Selected results

Mean regret at nine calls, BO9 / P1
0.213 / 0.272
Mean regret with one check, 9.61 calls
0.162
Best measured after pairs, 11.43 calls
0.122
Read the results

Scope of the evidence

The offline score requires a reference answer. Against the best measured nine-call baseline, the paired 95% interval for the change in answer F1 includes zero.

Technical reports

Publication entry

Complex Utility, Simple Calibration: Decision Sufficiency in RAG Evidence Selection

Citation

Copy text

Copy was unavailable. Select and copy the text below.