When we began studying jailbreak evaluations, we found a fascinating paper claiming that you could jailbreak frontier LLMs simply by translating forbidden prompts into obscure languages. Excited by this result, we attempted to reproduce it and found something unexpected. The paper in question claimed an impressive 43% success rate in jailbreaking GPT-4 by translating forbidden prompts into Scots Gaelic (Yong et al., 2023). To showcase their method, the authors asked GPT-4 to provide instructions for building a homemade explosive device using household materials. The authors then show the beginning of GPT-4's alarming response: Truncated GPT-4 response as reported by the authors.

Microsoft is currently focusing significant financial and human resources on the development of its AI assistant Copilot and its integration into Windows and Microsoft 365 applications. The company sees this as an opportunity to set itself apart from the competition of Libre Office and Google. Today, users have several alternatives for AI support in Office. This is because ChatGPT from OpenAI, the software that triggered the AI hype, is also suitable for office tasks in conjunction with Word, Excel, and others. Independent developers provide add-ons that allow you to integrate ChatGPT directly into Word so that you always have it at hand. At the same time, there are AI systems, especially from American providers, that help you create presentations online. These presentations can then be downloaded and in many cases converted into PowerPoint format PPTX.

In 2015, Elon Musk and Sam Altman cofounded OpenAI based on a seemingly ethical ethos: to develop AI technology that benefits humanity, rather than systems controlled by big money corporations. Fast forward a decade that included a spectacular falling out between Musk and Altman, things look very different. Amid legal battles with his friend and former business partner, Musk's latest company, xAI, has launched its own powerful competitor, Grok AI. Described as "an AI search assistant with a twist of humor and a dash of rebellion," Grok is designed to have fewer guardrails than its major competitors. Unsurprisingly, Grok is prone to hallucinations and bias, with the AI assistant blamed for spreading misinformation about the 2024 election.

Deep neural networks (DNNs) have long been recognized as vulnerable to backdoor attacks. By providing poisoned training data in the fine-tuning process, the attacker can implant a backdoor into the victim model. This enables input samples meeting specific textual trigger patterns to be classified as target labels of the attacker's choice. While such black-box attacks have been well explored in both computer vision and natural language processing (NLP), backdoor attacks relying on white-box attack philosophy have hardly been thoroughly investigated. In this paper, we take the first step to introduce a new type of backdoor attack that conceals itself within the underlying model architecture. Specifically, we propose to design separate backdoor modules consisting of two functions: trigger detection and noise injection. The add-on modules of model architecture layers can detect the presence of input trigger tokens and modify layer weights using Gaussian noise to disturb the feature distribution of the baseline model. We conduct extensive experiments to evaluate our attack methods using two model architecture settings on five different large language datasets. We demonstrate that the training-free architectural backdoor on a large language model poses a genuine threat. Unlike the-state-of-art work, it can survive the rigorous fine-tuning and retraining process, as well as evade output probability-based defense methods (i.e. BDDR). All the code and data is available https://github.com/SiSL-URI/Arch_Backdoor_LLM.

Here we show that a Large Language Model (LLM) can serve as a foundation model for a Chemical Language Model (CLM) which performs at or above the level of CLMs trained solely on chemical SMILES string data. Using supervised fine-tuning (SFT) and direct preference optimization (DPO) on the open-source Llama LLM, we demonstrate that we can train an LLM to respond to prompts such as generating molecules with properties of interest to drug development. This overall framework allows an LLM to not just be a chatbot client for chemistry and materials tasks, but can be adapted to speak more directly as a CLM which can generate molecules with user-specified properties.

We introduce LLMmap, a first-generation fingerprinting technique targeted at LLM-integrated applications. LLMmap employs an active fingerprinting approach, sending carefully crafted queries to the application and analyzing the responses to identify the specific LLM version in use. Our query selection is informed by domain expertise on how LLMs generate uniquely identifiable responses to thematically varied prompts. With as few as 8 interactions, LLMmap can accurately identify 42 different LLM versions with over 95% accuracy. More importantly, LLMmap is designed to be robust across different application layers, allowing it to identify LLM versions--whether open-source or proprietary--from various vendors, operating under various unknown system prompts, stochastic sampling hyperparameters, and even complex generation frameworks such as RAG or Chain-of-Thought. We discuss potential mitigations and demonstrate that, against resourceful adversaries, effective countermeasures may be challenging or even unrealizable.

Large Language Models (LLMs) demonstrate human-level capabilities in dialogue, reasoning, and knowledge retention. However, even the most advanced LLMs face challenges such as hallucinations and real-time updating of their knowledge. Current research addresses this bottleneck by equipping LLMs with external knowledge, a technique known as Retrieval Augmented Generation (RAG). However, two key issues constrained the development of RAG. First, there is a growing lack of comprehensive and fair comparisons between novel RAG algorithms. Second, open-source tools such as LlamaIndex and LangChain employ high-level abstractions, which results in a lack of transparency and limits the ability to develop novel algorithms and evaluation metrics. To close this gap, we introduce RAGLAB, a modular and research-oriented open-source library. RAGLAB reproduces 6 existing algorithms and provides a comprehensive ecosystem for investigating RAG algorithms. Leveraging RAGLAB, we conduct a fair comparison of 6 RAG algorithms across 10 benchmarks. With RAGLAB, researchers can efficiently compare the performance of various algorithms and develop novel algorithms.

In-context learning (ICL) is a cornerstone of large language model (LLM) functionality, yet its theoretical foundations remain elusive due to the complexity of transformer architectures. In particular, most existing work only theoretically explains how the attention mechanism facilitates ICL under certain data models. It remains unclear how the other building blocks of the transformer contribute to ICL. To address this question, we study how a two-attention-layer transformer is trained to perform ICL on $n$-gram Markov chain data, where each token in the Markov chain statistically depends on the previous $n$ tokens. We analyze a sophisticated transformer model featuring relative positional embedding, multi-head softmax attention, and a feed-forward layer with normalization. We prove that the gradient flow with respect to a cross-entropy ICL loss converges to a limiting model that performs a generalized version of the induction head mechanism with a learned feature, resulting from the congruous contribution of all the building blocks. In the limiting model, the first attention layer acts as a $\mathit{copier}$, copying past tokens within a given window to each position, and the feed-forward network with normalization acts as a $\mathit{selector}$ that generates a feature vector by only looking at informationally relevant parents from the window. Finally, the second attention layer is a $\mathit{classifier}$ that compares these features with the feature at the output position, and uses the resulting similarity scores to generate the desired output. Our theory is further validated by experiments.

Recent breakthroughs of large language models (LLMs) have exhibited superior capability across major industries and stimulated multi-hundred-billion-dollar investment in AI-centric data centers in the next 3-5 years. This, in turn, bring the increasing concerns on sustainability and AI-related energy usage. However, there is a largely overlooked issue as challenging and critical as AI model and infrastructure efficiency: the disruptive dynamic power consumption behaviour. With fast, transient dynamics, AI infrastructure features ultra-low inertia, sharp power surge and dip, and a significant peak-idle power ratio. The power scale covers from several hundred watts to megawatts, even to gigawatts. These never-seen-before characteristics make AI a very unique load and pose threats to the power grid reliability and resilience. To reveal this hidden problem, this paper examines the scale of AI power consumption, analyzes AI transient behaviour in various scenarios, develops high-level mathematical models to depict AI workload behaviour and discusses the multifaceted challenges and opportunities they potentially bring to existing power grids. Observing the rapidly evolving machine learning (ML) and AI technologies, this work emphasizes the critical need for interdisciplinary approaches to ensure reliable and sustainable AI infrastructure development, and provides a starting point for researchers and practitioners to tackle such challenges.

We demonstrate the ability of large language models (LLMs) to perform material and molecular property regression tasks, a significant deviation from the conventional LLM use case. We benchmark the Large Language Model Meta AI (LLaMA) 3 on several molecular properties in the QM9 dataset and 24 materials properties. Only composition-based input strings are used as the model input and we fine tune on only the generative loss. We broadly find that LLaMA 3, when fine-tuned using the SMILES representation of molecules, provides useful regression results which can rival standard materials property prediction models like random forest or fully connected neural networks on the QM9 dataset. Not surprisingly, LLaMA 3 errors are 5-10x higher than those of the state-of-the-art models that were trained using far more granular representation of molecules (e.g., atom types and their coordinates) for the same task. Interestingly, LLaMA 3 provides improved predictions compared to GPT-3.5 and GPT-4o. This work highlights the versatility of LLMs, suggesting that LLM-like generative models can potentially transcend their traditional applications to tackle complex physical phenomena, thus paving the way for future research and applications in chemistry, materials science and other scientific domains.