Transformer-based models have achieved remarkable performance. However, their structural characteristics—multiple layers and attention heads—introduce efficiency challenges in inference and deployment. To address these challenges, various pruning methods have recently been proposed. Notably, gradient-based methods using Head Importance Scores (HIS) have gained traction for interpretability, efficiency, and ability to identify redundant heads. However, HIS alone has limitations as it captures only the gradient-driven contribution, overlooking the diversity of attention patterns. To overcome these limitations, we introduce a novel pruning criterion, HIES (Head Importance-Entropy Score), which integrates head importance scores with attention entropy, providing complementary evidence on per-head contribution. Empirically, HIES-based pruning yields a 1.13% improvement in model quality and 3.52% improvement in stability over the HIS-only method on average, substantially reducing the accuracy and stability cost of aggressive compression. Code will be released upon publication.

Fine-tuning a pretrained LLM into a vision-language model (VLM) can erode the backbone’s text capability, with the damage concentrated on tasks that require following exact output rules, such as instruction following, chain-of-thought reasoning graded on a strictly parsed final answer, and similar evaluations with strict graders. We trace this gap to *attention-sink corruption*: VL fine-tuning perturbs the early sink position that anchors a large fraction of attention probability, and how well the base LLM preserves its sink tracks how much of the affected capability survives adaptation. Building on this view, we introduce *Sink Strength*, a single scalar computed on the base LLM in a few seconds on a single GPU that predicts post-VL degradation without any VL training. It consistently tracks relative degradation across the six VLM–LLM pairs and multiple format-sensitive tasks. Complementing this diagnostic, we find that post-pretraining QK-norm injection fails to reproduce the protection of native QK-norm, while several off-the-shelf weight-merging settings fail to recover the lost capability after VL training. These negative results underscore the value of screening backbones with Sink Strength before VL training and narrow the intervention space toward head-selective training-time protection.

Transformer-based large-scale pre-trained models achieve great success. Fine-tuning is the standard practice for leveraging these models in downstream tasks. Among the fine-tuning methods, adapter-tuning provides a parameter-efficient fine-tuning by introducing lightweight trainable modules while keeping most pre-trained parameters frozen. However, existing adapter-tuning methods still impose substantial resource usage. Through our investigation, we show that each adapter unequally contributes to both task performance and resource usage. Motivated by this insight, we propose Selective Adapter FrEezing (SAFE), which gradually freezes less important adapters early to reduce unnecessary resource usage while maintaining performance. In our experiments, SAFE reduces memory usage, computation amount, and training time by 42.85%, 34.59%, and 11.82%, respectively, while achieving comparable or better task performance compared to the baseline. We also demonstrate that SAFE induces regularization effect, thereby smoothing the loss landscape, which enables the model to generalize better by avoiding sharp minima.