Instruction tuning is instrumental in enabling Large Language Models~(LLMs) to follow user instructions to complete various open-domain tasks. The success of instruction tuning depends on the availability of high-quality instruction data. Owing to the exorbitant cost and substandard quality of human annotation, recent works have been deeply engaged in the exploration of the utilization of powerful closed-source models to generate instruction data automatically. However, these methods carry potential risks arising from the usage requirements of powerful closed-source models, which strictly forbid the utilization of their outputs to develop machine learning models. To deal with this problem, in this work, we explore alternative approaches to generate high-quality instruction data that do not rely on closed-source models. Our exploration includes an investigation of various existing instruction generation methods, culminating in the integration of the most efficient variant with two novel strategies to enhance the quality further. Evaluation results from two benchmarks and the GPT-4 model demonstrate the effectiveness of our generated instruction data, which can outperform Alpaca, a method reliant on closed-source models. We hope that more progress can be achieved in generating high-quality instruction data without using closed-source models.

Binary Code Embedding (BCE) has important applications in various reverse engineering tasks such as binary code similarity detection, type recovery, control-flow recovery and data-flow analysis. Recent studies have shown that the Transformer model can comprehend the semantics of binary code to support downstream tasks. However, existing models overlooked the prior knowledge of assembly language. In this paper, we propose a novel Transformer-based approach, namely kTrans, to generate knowledge-aware binary code embedding. By feeding explicit knowledge as additional inputs to the Transformer, and fusing implicit knowledge with a novel pre-training task, kTrans provides a new perspective to incorporating domain knowledge into a Transformer framework. We inspect the generated embeddings with outlier detection and visualization, and also apply kTrans to 3 downstream tasks: Binary Code Similarity Detection (BCSD), Function Type Recovery (FTR) and Indirect Call Recognition (ICR). Evaluation results show that kTrans can generate high-quality binary code embeddings, and outperforms state-of-the-art (SOTA) approaches on downstream tasks by 5.2%, 6.8%, and 12.6% respectively. kTrans is publicly available at: https://github.com/Learner0x5a/kTrans-release

Emerging tools bring forth fresh approaches to work, and the field of natural science is no different. In natural science, traditional manual, serial, and labour-intensive work is being augmented by automated, parallel, and iterative processes driven by artificial intelligence-based experimental automation and more. To add new capabilities in natural science, enabling the acceleration and enrichment of automation of the discovery process, we present DARWIN, a series of tailored LLMs for natural science, mainly in physics, chemistry, and material science. This series relies on open-source LLM, incorporating structured and unstructured scientific knowledge from public datasets and literature. We fine-tuned the models using over 60,000 instruction data points, emphasizing factual correctness. During the fine-tuning, we introduce the Scientific Instruction Generation (SIG) model, automating instruction generation from scientific texts. This eliminates the need for manual extraction or domain-specific knowledge graphs and efficiently injects scientific knowledge into the model. We also explore multi-task training strategies, revealing interconnections between scientific tasks. DARWIN series not only achieves state-of-the-art results on various scientific tasks but also diminishes reliance on closed-source AI models. Our research showcases the ability of LLM in the scientific domain, with the overarching goal of fostering prosperity within the broader AI for science community.

In traffic safety research, extracting information from crash narratives using text analysis is a common practice. With recent advancements of large language models (LLM), it would be useful to know how the popular LLM interfaces perform in classifying or extracting information from crash narratives. To explore this, our study has used the three most popular publicly available LLM interfaces- ChatGPT, BARD and GPT4. This study investigated their usefulness and boundaries in extracting information and answering queries related to accidents from 100 crash narratives from Iowa and Kansas. During the investigation, their capabilities and limitations were assessed and their responses to the queries were compared. Five questions were asked related to the narratives: 1) Who is at-fault? 2) What is the manner of collision? 3) Has the crash occurred in a work-zone? 4) Did the crash involve pedestrians? and 5) What are the sequence of harmful events in the crash? For questions 1 through 4, the overall similarity among the LLMs were 70%, 35%, 96% and 89%, respectively. The similarities were higher while answering direct questions requiring binary responses and significantly lower for complex questions. To compare the responses to question 5, network diagram and centrality measures were analyzed. The network diagram from the three LLMs were not always similar although they sometimes have the same influencing events with high in-degree, out-degree and betweenness centrality. This study suggests using multiple models to extract viable information from narratives. Also, caution must be practiced while using these interfaces to obtain crucial safety related information.

Recently, there has been growing interest in using Large Language Models (LLMs) for scientific research. Numerous benchmarks have been proposed to evaluate the ability of LLMs for scientific research. However, current benchmarks are mostly based on pre-collected objective questions. This design suffers from data leakage problem and lacks the evaluation of subjective Q/A ability. In this paper, we propose SciEval, a comprehensive and multi-disciplinary evaluation benchmark to address these issues. Based on Bloom's taxonomy, SciEval covers four dimensions to systematically evaluate scientific research ability. In particular, we design a "dynamic" subset based on scientific principles to prevent evaluation from potential data leakage. Both objective and subjective questions are included in SciEval. These characteristics make SciEval a more effective benchmark for scientific research ability evaluation of LLMs. Comprehensive experiments on most advanced LLMs show that, although GPT-4 achieves SOTA performance compared to other LLMs, there is still substantial room for improvement, especially for dynamic questions. The data and codes are now publicly available.

The rapid growth in the usage and applications of Natural Language Processing (NLP) in various sociotechnical solutions has highlighted the need for a comprehensive understanding of bias and its impact on society. While research on bias in NLP has expanded, several challenges persist that require attention. These include the limited focus on sociodemographic biases beyond race and gender, the narrow scope of analysis predominantly centered on models, and the technocentric implementation approaches. This paper addresses these challenges and advocates for a more interdisciplinary approach to understanding bias in NLP. The work is structured into three facets, each exploring a specific aspect of bias in NLP.

Some argue scale is all what is needed to achieve AI, covering even causal models. We make it clear that large language models (LLMs) cannot be causal and give reason onto why sometimes we might feel otherwise. To this end, we define and exemplify a new subgroup of Structural Causal Model (SCM) that we call meta SCM which encode causal facts about other SCM within their variables. We conjecture that in the cases where LLM succeed in doing causal inference, underlying was a respective meta SCM that exposed correlations between causal facts in natural language on whose data the LLM was ultimately trained. If our hypothesis holds true, then this would imply that LLMs are like parrots in that they simply recite the causal knowledge embedded in the data. Our empirical analysis provides favoring evidence that current LLMs are even weak `causal parrots.'

Security critical software, e.g., OpenSSL, comes with numerous side-channel leakages left unpatched due to a lack of resources or experts. The situation will only worsen as the pace of code development accelerates, with developers relying on Large Language Models (LLMs) to automatically generate code. In this work, we explore the use of LLMs in generating patches for vulnerable code with microarchitectural side-channel leakages. For this, we investigate the generative abilities of powerful LLMs by carefully crafting prompts following a zero-shot learning approach. All generated code is dynamically analyzed by leakage detection tools, which are capable of pinpointing information leakage at the instruction level leaked either from secret dependent accesses or branches or vulnerable Spectre gadgets, respectively. Carefully crafted prompts are used to generate candidate replacements for vulnerable code, which are then analyzed for correctness and for leakage resilience. From a cost/performance perspective, the GPT4-based configuration costs in API calls a mere few cents per vulnerability fixed. Our results show that LLM-based patching is far more cost-effective and thus provides a scalable solution. Finally, the framework we propose will improve in time, especially as vulnerability detection tools and LLMs mature.

While most research into speech synthesis has focused on synthesizing high-quality speech for in-dataset speakers, an equally essential yet unsolved problem is synthesizing speech for unseen speakers who are out-of-dataset with limited reference data, i.e., speaker adaptive speech synthesis. Many studies have proposed zero-shot speaker adaptive text-to-speech and voice conversion approaches aimed at this task. However, most current approaches suffer from the degradation of naturalness and speaker similarity when synthesizing speech for unseen speakers (i.e., speakers not in the training dataset) due to the poor generalizability of the model in out-of-distribution data. To address this problem, we propose GZS-TV, a generalizable zero-shot speaker adaptive text-to-speech and voice conversion model. GZS-TV introduces disentangled representation learning for both speaker embedding extraction and timbre transformation to improve model generalization and leverages the representation learning capability of the variational autoencoder to enhance the speaker encoder. Our experiments demonstrate that GZS-TV reduces performance degradation on unseen speakers and outperforms all baseline models in multiple datasets.

Vision-Language Pre-training (VLP) shows remarkable progress with the assistance of extremely heavy parameters, which challenges deployment in real applications. Knowledge distillation is well recognized as the essential procedure in model compression. However, existing knowledge distillation techniques lack an in-depth investigation and analysis of VLP, and practical guidelines for VLP-oriented distillation are still not yet explored. In this paper, we present DLIP, a simple yet efficient Distilling Language-Image Pre-training framework, through which we investigate how to distill a light VLP model. Specifically, we dissect the model distillation from multiple dimensions, such as the architecture characteristics of different modules and the information transfer of different modalities. We conduct comprehensive experiments and provide insights on distilling a light but performant VLP model. Experimental results reveal that DLIP can achieve a state-of-the-art accuracy/efficiency trade-off across diverse cross-modal tasks, e.g., image-text retrieval, image captioning and visual question answering. For example, DLIP compresses BLIP by 1.9x, from 213M to 108M parameters, while achieving comparable or better performance. Furthermore, DLIP succeeds in retaining more than 95% of the performance with 22.4% parameters and 24.8% FLOPs compared to the teacher model and accelerates inference speed by 2.7x.