Supervised fine-tuning (SFT) on benign data can paradoxically erode a language model’s safety alignment, a phenomenon known as catastrophic forgetting of safety behaviors. Although prior work shows that randomly adding safety examples can reduce harmful output, the principles that make certain examples more effective than others remain poorly understood. This paper investigates the hypothesis that the effectiveness of a safety example is governed by two key factors: its instruction-response behavior (e.g., refusal vs. explanation) and its semantic diversity across harm categories. We systematically evaluate sampling strategies based on these axes and find that structured, diversity-aware sampling significantly improves model safety. Our method reduces harmfulness by up to 41% while adding only 0.05% more data to the fine-tuning set.

Hand-crafting high quality prompts to optimize the performance of language models is a complicated and labor-intensive process. Furthermore, when migrating to newer, smaller, or weaker models (possibly due to latency or cost gains), prompts need to be updated to re-optimize the task performance. We propose Concept Distillation (CD), an automatic prompt optimization technique for enhancing weaker models on complex tasks. CD involves: (1) collecting mistakes made by weak models with a base prompt (initialization), (2) using a strong model to generate reasons for these mistakes and create rules/concepts for weak models (induction), and (3) filtering these rules based on validation set performance and integrating them into the base prompt (deduction/verification). We evaluated CD on NL2Code and mathematical reasoning tasks, observing significant performance boosts for small and weaker language models. Notably, Mistral-7B’s accuracy on Multi-Arith increased by 20%, and Phi-3-mini-3.8B’s accuracy on HumanEval rose by 34%. Compared to other automated methods, CD offers an effective, cost-efficient strategy for improving weak models’ performance on complex tasks and enables seamless workload migration across different language models without compromising performance.

Fine-tuning all the layers of a pre-trained neural language encoder (either using all the parameters or using parameter-efficient methods) is often the de-facto way of adapting it to a new task. We show evidence that for different downstream language tasks, fine-tuning only a subset of layers is sufficient to obtain performance that is close to and often better than fine-tuning all the layers in the language encoder. We propose an efficient metric based on the diagonal of the Fisher information matrix (FIM score), to select the candidate layers for selective fine-tuning. We show, empirically on GLUE and SuperGLUE tasks and across distinct language encoders, that this metric can effectively select layers leading to a strong downstream performance. Our work highlights that task-specific information corresponding to a given downstream task is often localized within a few layers, and tuning only those is sufficient for strong performance. Additionally, we demonstrate the robustness of the FIM score to rank layers in a manner that remains constant during the optimization process.

We present SLATE, a sequence labeling approach for extracting tasks from free-form content such as digitally handwritten (or “inked”) notes on a virtual whiteboard. Our approach allows us to create a single, low-latency model to simultaneously perform sentence segmentation and classification of these sentences into task/non-task sentences. SLATE greatly outperforms a baseline two-model (sentence segmentation followed by classification model) approach, achieving a task F1 score of 84.4%, a sentence segmentation (boundary similarity) score of 88.4% and three times lower latency compared to the baseline. Furthermore, we provide insights into tackling challenges of performing NLP on the inking domain. We release both our code and dataset for this novel task.