Learning to Stay Safe: Adaptive Regularization Against Safety Degradation during Fine-Tuning

Instruction-following language models are trained to be helpful and safe, yet their safety behavior can deteriorate under benign fine-tuning and worsen under adversarial updates. Existing defenses often offer limited protection or force a trade-off between safety and utility. We introduce a training framework that adapts regularization in response to safety risk, enabling models to remain aligned throughout fine-tuning. To estimate safety risk at training time, we explore two distinct approaches: a judge-based Safety Critic that assigns high-level harm scores to training batches, and an activation-based risk predictor built with a lightweight classifier trained on intermediate model activations to estimate harmful intent. Each approach provides a risk signal that is used to constrain updates deemed higher risk to remain close to a safe reference policy, while lower-risk updates proceed with standard training. We empirically verify that harmful intent signals are predictable from pre-generation activations and that judge scores provide effective high-recall safety guidance. Across multiple model families and attack scenarios, adaptive regularization with either risk estimation approach consistently lowers attack success rate compared to standard fine-tuning, preserves downstream performance, and adds no inference-time cost. This work demonstrates a principled mechanism for maintaining safety without sacrificing utility.

CVOCSemRPL: Class-Variance Optimized Clustering, Semantic Information Injection and Restricted Pseudo Labeling based Improved Semi-Supervised Few-Shot Learning

Few-shot learning has been extensively explored to address problems where the amount of labeled samples is very limited for some classes. In the semi-supervised few-shot learning setting, substantial quantities of unlabeled samples are available. Such unlabeled samples are generally cheaper to obtain and can be used to improve the few-shot learning performance of the model. Some of the recent methods for this setting rely on clustering to generate pseudo-labels for the unlabeled samples. Since the quality of the representation learned by the model heavily influences the effectiveness of clustering, this might also lead to incorrect labeling of the unlabeled samples and consequently lead to a drop in the few-shot learning performance. We propose an approach for semi-supervised few-shot learning that performs a class-variance optimized clustering in order to improve the effectiveness of clustering the labeled and unlabeled samples in this setting. It also optimizes the clustering-based pseudo-labeling process using a restricted pseudo-labeling approach and performs semantic information injection in order to improve the semi-supervised few-shot learning performance of the model. We experimentally demonstrate that our proposed approach significantly outperforms recent state-of-the-art methods on the benchmark datasets.