Exact risk analysis finds students inherit teacher bias
Original title: Teacher Supervision, Fixed Representations and Calibration: Exact Risk Calculations for Candidate-Scoring Decision Models
On 231,521 CFPB complaints, outcome-trained students beat the teacher's decision cost by 8%.
The paper analyzes candidate-scoring models whose public descriptions cite soft teacher targets, a frozen pretrained backbone, and temperature calibration, including Jev and an open counterpart, JevAny. Exact Gaussian and probit calculations show a student inherits teacher bias that more data and teacher agreement do not remove, while a fixed representation helps only below a sample-size threshold. On 231,521 Consumer Financial Protection Bureau credit-card complaints, an outcome-trained student ends 8% below the logistic teacher's decision cost, and a zero-shot 30B model stays worse after recalibration.
- Temperature scaling removes excess risk only with the correct coefficient direction.
- Under Luce's axiom, coherent scores aggregate by log-sum-exp.
- The 30B model departs from coherence by 17.5 decibans.
- The decision task is whether a complaint ends in monetary relief.
- Organizations
- Consumer Financial Protection Bureau
From the original article
1 min read at arxiv.orgCandidate-scoring decision models return a probability distribution over a presented set of answers. Public documentation of one such system, Jev, and of an open counterpart, JevAny, describes three training ingredients that determine what these probabilities estimate: soft targets from larger models, a frozen pretrained backbone and temperature calibration.
By Agus Sudjianto · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.