Inference-time Unlearning Using Conformal Prediction

Somnath Basu Roy Chowdhury, Rahul Kidambi, Avinava Dubey, David Wang, Gokhan Mergen, Amr Ahmed, Aranyak Mehta · 2026

Read Inference-time Unlearning Using Conformal Prediction by Somnath Basu Roy Chowdhury, Rahul Kidambi, Avinava Dubey, David Wang, Gokhan Mergen, Amr Ahmed, Aranyak Mehta (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-03787.txt.

Description

Removing information from a trained model usually means retraining some of its weights, which can dull its general abilities. Somnath Basu Roy Chowdhury and 6 co-authors move the unlearning to inference time in a 2026 paper. The frozen generative model is paired with a verifier that judges whether each response meets the unlearning requirements.

Responses are refined in rounds from the verifier's feedback, with no parameter updates at all. Conformal prediction enters to cut the computing overhead and to supply distribution-free guarantees on the result. Across hard unlearning benchmarks, the approach reduces unlearning error by up to ninety-three percent against state-of-the-art methods. The approach suits generative models above all, where parameter-editing assumptions often break down.

Catalog record

CREATORSomnath Basu Roy Chowdhury, Rahul Kidambi, Avinava Dubey, David Wang, Gokhan Mergen, Amr Ahmed, Aranyak Mehta
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE45 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-03787

Browse the full library · About the Tower of Babel