REDaCT: Collapse-Free Unlearning in Large Reasoning Models
Paper • 2610.12769 • Published
REDaCT Leak Detector
This repository contains the binary leak-detector classifier used in REDaCT: Collapse-Free Unlearning in Large Reasoning Models. It is a RoBERTa-based model with a single binary classification head that identifies sentences in a chain-of-thought reasoning trace that leak forget-set information.
The classifier is used by the REDaCT framework to detect the first offending sentence during inference, after which a corrective embedding is injected to regenerate the reasoning in a forget-safe manner.
For the full reproduction pipeline, training scripts, and usage examples, see the GitHub repository.