Reproduce “Conformal Calibration Transfer” on OpenResearch

About this paper

Achref Doula

Conformal Calibration Transfer enables prediction sets in a target input space when labeled calibration data exist only in a different source space. It transports labeled source examples using unlabeled paired data, then corrects residual mismatch with either a conservative surrogate-based safeguard or density-ratio weighting; across image corruption and cross-modal benchmarks, the safeguard recovered coverage in all 11 identified under-coverage cases, while weighted calibration generally preserved set efficiency.

Key insights

• Transport separates the problem into two stages: moving labeled calibration information into the target space and controlling the remaining mismatch there, rather than attempting direct source-domain calibration.
• TCC-KS prioritizes coverage protection and becomes more conservative as the observable difficulty gap grows, whereas weighted-TCC is more efficient but depends strongly on stable density-ratio estimates and effective sample size.
• Predictive alignment during transport can reduce residual mismatch and improve weighting reliability, but severe modality or corruption shifts may still require larger prediction sets or additional labeled target auditing.

Summary by alphaXiv.