Learning What Matters: Prioritized Concept Learning via Relative Error-driven Sample Selection

TL;DR

Experiments across multiple instruction-tuning datasets of varying scales demonstrate that PROGRESS consistently outperforms state-of-the-art baselines with much less data and supervision, validating PROGRESS as a scalable solution for efficient learning.

Abstract

Instruction tuning has been central to the success of recent vision-language models (VLMs), but it remains expensive-requiring large-scale datasets, high-quality annotations, and large compute budgets. We propose PRioritized cOncept learninG via Relative Error-driven Sample Selection (PROGRESS), a data- and compute-efficient framework that enables VLMs to dynamically select what to learn next based on their evolving needs during training. At each stage, the model tracks its learning progress across skills and selects the most informative samples-those it has not already mastered and that are not too difficult to learn at the current stage of training. This strategy effectively controls skill acquisition and the order in which skills are learned. Specifically, we sample from skills showing the highest learning progress, prioritizing those with the most rapid improvement. Unlike prior methods, PROGRESS requires no upfront answer annotations, queries answers only on a need basis, avoids reliance on additional supervision from auxiliary VLMs, and does not require compute-heavy gradient computations for data selection. Experiments across multiple instruction-tuning datasets of varying scales demonstrate that PROGRESS consistently outperforms state-of-the-art baselines with much less data and supervision. Additionally, we show strong cross-architecture generalization and transferability to larger models, validating PROGRESS as a scalable solution for efficient learning.

Figures (13)

• Figure 1: Comparison with Prior Efficient Learning Methods for VLMs.Green denote desirable properties for efficient learning, while Red indicate limitations. PROGRESS satisfies all key desirable criteria while requiring only 20% data. See Appendix \ref{['app:baselines']} for details of prior approaches.
• Figure 2: Overall Pipeline. Our framework consists of two stages: (1) Multimodal Concept Categorization, which partitions the unlabeled pool $\mathbb{U}$ into distinct skills by assigning each sample $(I, Q)$ to a specific skill cluster, and (2) Prioritized Concept Learning, where the model actively selects the most informative samples—those showing the highest improvement in its objective (e.g., accuracy or loss) relative to its prior state. We query annotations for only these selected samples on a need basis, forming labeled set $(I, Q, A)$, which is used for training.
• Figure 3: Cluster Visualization. Clustering with multimodal DINO-BERT features ensures purer skill clusters with higher intra-cluster and lower inter-cluster similarity compared to uni-modal partitioning. See Word Cloud Visualization Appendix \ref{['app:word_cloud']}.
• Figure 4: Learning Dynamics Across Difficulty Levels. PROGRESS consistently achieves higher accuracy and reduced variance compared to other selection strategies
• Figure 5: Ablation of Softmax Temperature. Both very-low and very-high temperatures lead to a significant performance drop.