Recent years have witnessed extensive efforts to enhance Large Language Models (LLMs) across various domains, alongside growing attention to their ethical implications. However, a critical challenge remains largely overlooked: LLMs must balance between rejecting harmful requests for safety and accommodating legitimate ones for utility. This paper presents a Direct Preference Optimization (DPO) based alignment framework that achieves better overall performance by addressing this ethical-utility trade-off, using chemical domain applications as a proof-of-concept. Our alignment pipeline starts with a GPT-assisted three-phase data generation scheme, in which we create LibraChemQA, a chemical question-answering dataset comprising 31.6k triplet instances. By incorporating an innovative balanced seed in the data generation process, our framework systematically considers both legitimate and illegitimate requests. The framework also introduces a rephrasing mechanism for efficient data augmentation that enhances the model's chemical comprehension. We further develop a novel hybrid evaluation scheme with LLM judges for precise assessment of both safety and utility. Experimental results demonstrate our model's substantial improvements in overall performance where both safety and utility are considered - our resulting model, LibraChem, outperforms leading LLMs including Claude-3, GPT-4o, and LLaMA-3 by margins of 13.44%, 7.16%, and 7.10% respectively on our released benchmark.

Spatio-temporal forecasting is crucial in real-world dynamic systems, predicting future changes using historical data from diverse locations. Existing methods often prioritize the development of intricate neural networks to capture the complex dependencies of the data, yet their accuracy fails to show sustained improvement. Besides, these methods also overlook node heterogeneity, hindering customized prediction modules from handling diverse regional nodes effectively. In this paper, our goal is not to propose a new model but to present a novel low-rank adaptation framework as an off-the-shelf plugin for existing spatial-temporal prediction models, termed ST-LoRA, which alleviates the aforementioned problems through node-level adjustments. Specifically, we first tailor a node adaptive low-rank layer comprising multiple trainable low-rank matrices. Additionally, we devise a multi-layer residual fusion stacking module, injecting the low-rank adapters into predictor modules of various models. Across six real-world traffic datasets and six different types of spatio-temporal prediction models, our approach minimally increases the parameters and training time of the original models by less than 4%, still achieving consistent and sustained performance enhancement.