Evaluation of CART on Longer Continual Task Sequences

Evaluate the Compatible Routing (CART) method on substantially longer multimodal continual instruction-tuning task sequences to assess its long-term knowledge retention.

Background

The paper evaluates CART on the CoIN and UCIT benchmarks, which contain eight and six sequential tasks, respectively. The authors acknowledge that these experiments do not establish how the method behaves over substantially longer continual-learning streams, where expert consolidation, routing, and accumulated adaptation may pose additional challenges. Extending the evaluation is therefore identified as necessary for a more comprehensive assessment of long-term knowledge retention.