SimE: A Simple Efficient Incremental Learning Framework with Vision-Language Models

A recent study published on arXiv (2603.11211v3) presents SimE, a straightforward and effective framework designed for Incremental Learning (IL) that utilizes vision-language models equipped with adapters. The researchers noted a nonlinear relationship between the quantity of adaptive adapter connections and the IL performance of the model: while increasing connections between transformer blocks enhances outcomes, adding more within blocks during smaller tasks might yield different results. This framework tackles three key challenges in IL: enhancing training efficiency, minimizing dependence on memory banks for previous data retention, and ensuring a robust backbone. The full paper can be accessed at https://arxiv.org/abs/2603.11211.