Stripe's Fraud Model Gets 39% Higher Recall by Training on Explanations in the Right Order

emilygsands · x · 2026-11-20

A Stripe engineer shares how teaching a model the reasoning behind decisions helps—but only with the right training order. Mixing labels and explanations in standard SFT actually hurts, since long explanations dominate the gradient. Their fix: two-stage SFT (labels+explanations first, labels alone second), then on-policy distillation with reverse KL to recover explanation quality. Result: 39% higher recall at human-review precision on fraud review, with explanations that beat the teacher's. Full write-up on the Stripe Dot Dev blog by Hangzhi Guo.

Related event: Stripe Boosts Fraud Review Recall 39% by Learning from Explanations(2 posts)→

Original post →

More from coding & agent

coding & agent channel →