Official Review · ICLR 2026

The paper proposes a reinforcement learning framework for continual image classification. The method, RLAP-CLIP, uses Reinforcement Learning–based Prototype Optimization (RLPO) to actively refine class representations and mitigate prototype (representation of each class or a embedding of class) degradation. It incorporates difficulty-aware cross-modal fusion to route samples by complexity and enhanced dual-modal prompting to balance visual and textual adaptation. Experiments across multiple image classification benchmarks show RLAP-CLIP outperforms prior methods. - The reinforcement learning–based prototype update method empirically outperforms prior work. Enhanced dual-modal prompting and difficulty-aware sample handling further contribute to the gains, as demonstrated in experiments. - The paper provides empirical and theoretical evidence of training stability—covering hyperparameter sensitivity and convergence—which is critical for reinforcement learning frameworks. - The method improves continual image classification by combining established components—prototype updates, reinforcement learning, learnable prompts, and routing. While the gains on benchmarks are clear, the pipeline is heuristic and tailored only for classification, making its generalization to other continual recognition tasks (e.g., retrieval, detection...) uncertain. - Key terminology should be clearly defined (e.g., “prototype,” “center-based exemplar selection”) to improve readability and understanding. Only objectives and architectures are only explained in the paper. Training and inference process are not explained. Please more clarify what are frozen or learnable weights in the framework. - What is center-based example selection in the Figure 1. This part is not explained in the paper.