Original title: OSR: Output Space Redistribution for Adaptive Label Removal in Classification Models
Minyi Peng, Darian Gunamardi, Ivan Tjuawinata, Yongsen Zheng +1 · September 2026
Classification systems often need to delete categories when real-world labels change. Usually, fixing the model requires training it again on the original data or altering internal features. These old fixes cost a lot of computing power and storage, give inconsistent results, scale poorly, degrade model utility, and risk privacy. To fix this, researchers created a new method called Output Space RedistributionAdjusting the probability scores the model outputs for each category.. Instead of retraining or modifying the core model, it acts as a modular filter right at the output stage. It uses statistics to shift the model's confidence scores toward the remaining labels, matching a retrained model. Tests across several classification tasks show it competes with full retraining while using less compute and protecting privacy because it avoids original data.
Imagine you have a drawer organizer with sections labeled for different types of socks, but one type of sock is discontinued. Now, you still have old socks of that type, and you need to sort new laundry without reorganizing the entire drawer from scratch. In this paper, the method adjusts how the organizer distributes socks into sections so that it behaves as if the discontinued section never existed, using only what it already knows about sock types and past sorting results.
When categories change in systems like image recognition or spam filters, updating models used to be expensive and risky for privacy. This method lets them adapt quickly without retraining, reducing costs and protecting data.
The abstract does not specify how the method handles cases where prior output confidences are unavailable or if it applies to all types of classification tasks. It approximates retrained performance, so it may not be perfect in all scenarios. The method is only demonstrated in several classification tasks, not all possible domains.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
OSR: Output Space Redistribution for Adaptive Label Removal in Classification Models — Minyi Peng, Darian Gunamardi, Ivan Tjuawinata et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.