Large Language Models (LLMs) such as GPT and Llama2 are increasingly adopted in many safety-critical applications. Their security is thus essential. Even with considerable efforts spent on reinforcement learning from human feedback (RLHF), recent studies have shown that LLMs are still subject to attacks such as adversarial perturbation and Trojan attacks. Further research is thus needed to evaluate their security and/or understand the lack of it. In this work, we propose a framework for conducting light-weight causality-analysis of LLMs at the token, layer, and neuron level. We applied our framework to open-source LLMs such as Llama2 and Vicuna and had multiple interesting discoveries. Based on a layer-level causality analysis, we show that RLHF has the effect of overfitting a model to harmful prompts. It implies that such security can be easily overcome by `unusual' harmful prompts. As evidence, we propose an adversarial perturbation method that achieves 100\% attack success rate on the red-teaming tasks of the Trojan Detection Competition 2023. Furthermore, we show the existence of one mysterious neuron in both Llama2 and Vicuna that has an unreasonably high causal effect on the output. While we are uncertain on why such a neuron exists, we show that it is possible to conduct a ``Trojan'' attack targeting that particular neuron to completely cripple the LLM, i.e., we can generate transferable suffixes to prompts that frequently make the LLM produce meaningless responses.
相關內容
大語言模型是基于海量文本數據訓練的深度學習模型。它不僅能夠生成自然語言文本，還能夠深入理解文本含義，處理各種自然語言任務，如文本摘要、問答、翻譯等。2023年，大語言模型及其在人工智能領域的應用已成為全球科技研究的熱點，其在規模上的增長尤為引人注目，參數量已從最初的十幾億躍升到如今的一萬億。參數量的提升使得模型能夠更加精細地捕捉人類語言微妙之處，更加深入地理解人類語言的復雜性。在過去的一年里，大語言模型在吸納新知識、分解復雜任務以及圖文對齊等多方面都有顯著提升。隨著技術的不斷成熟，它將不斷拓展其應用范圍，為人類提供更加智能化和個性化的服務，進一步改善人們的生活和生產方式。
Factual knowledge encoded in Pre-trained Language Models (PLMs) enriches their representations and justifies their use as knowledge bases. Previous work has focused on probing PLMs for factual knowledge by measuring how often they can correctly predict an object entity given a subject and a relation, and improving fact retrieval by optimizing the prompts used for querying PLMs. In this work, we consider a complementary aspect, namely the coherency of factual knowledge in PLMs, i.e., how often can PLMs predict the subject entity given its initial prediction of the object entity. This goes beyond evaluating how much PLMs know, and focuses on the internal state of knowledge inside them. Our results indicate that PLMs have low coherency using manually written, optimized and paraphrased prompts, but including an evidence paragraph leads to substantial improvement. This shows that PLMs fail to model inverse relations and need further enhancements to be able to handle retrieving facts from their parameters in a coherent manner, and to be considered as knowledge bases.
In recent years, Large Language Models (LLMs) have achieved significant success in natural language processing (NLP) and various interdisciplinary areas. However, applying LLMs to chemistry is a complex task that requires specialized domain knowledge. This paper provides a thorough exploration of the nuanced methodologies employed in integrating LLMs into the field of chemistry, delving into the complexities and innovations at this interdisciplinary juncture. Specifically, our analysis begins with examining how molecular information is fed into LLMs through various representation and tokenization methods. We then categorize chemical LLMs into three distinct groups based on the domain and modality of their input data, and discuss approaches for integrating these inputs for LLMs. Furthermore, this paper delves into the pretraining objectives with adaptations to chemical LLMs. After that, we explore the diverse applications of LLMs in chemistry, including novel paradigms for their application in chemistry tasks. Finally, we identify promising research directions, including further integration with chemical knowledge, advancements in continual learning, and improvements in model interpretability, paving the way for groundbreaking developments in the field.
Neural scaling laws (NSL) refer to the phenomenon where model performance improves with scale. Sharma & Kaplan analyzed NSL using approximation theory and predict that MSE losses decay as $N^{-\alpha}$, $\alpha=4/d$, where $N$ is the number of model parameters, and $d$ is the intrinsic input dimension. Although their theory works well for some cases (e.g., ReLU networks), we surprisingly find that a simple 1D problem $y=x^2$ manifests a different scaling law ($\alpha=1$) from their predictions ($\alpha=4$). We opened the neural networks and found that the new scaling law originates from lottery ticket ensembling: a wider network on average has more "lottery tickets", which are ensembled to reduce the variance of outputs. We support the ensembling mechanism by mechanistically interpreting single neural networks, as well as studying them statistically. We attribute the $N^{-1}$ scaling law to the "central limit theorem" of lottery tickets. Finally, we discuss its potential implications for large language models and statistical physics-type theories of learning.
Large language models (LLMs) have garnered significant attention in both the AI community and beyond. Among these, the Generative Pre-trained Transformer (GPT) has emerged as the dominant architecture, spawning numerous variants. However, these variants have undergone pre-training under diverse conditions, including variations in input data, data preprocessing, and training methodologies, resulting in a lack of controlled comparative studies. Here we meticulously examine two prominent open-sourced GPT architectures, GPT-NeoX and LLaMA, leveraging the computational power of Frontier, the world's first Exascale supercomputer. Employing the same materials science text corpus and a comprehensive end-to-end pipeline, we conduct a comparative analysis of their training and downstream performance. Our efforts culminate in achieving state-of-the-art performance on a challenging materials science benchmark. Furthermore, we investigate the computation and energy efficiency, and propose a computationally efficient method for architecture design. To our knowledge, these pre-trained models represent the largest available for materials science. Our findings provide practical guidance for building LLMs on HPC platforms.
In patent prosecution, timely and effective responses to Office Actions (OAs) are crucial for acquiring patents, yet past automation and AI research have scarcely addressed this aspect. To address this gap, our study introduces the Patent Office Action Response Intelligence System (PARIS) and its advanced version, the Large Language Model Enhanced PARIS (LE-PARIS). These systems are designed to expedite the efficiency of patent attorneys in collaboratively handling OA responses. The systems' key features include the construction of an OA Topics Database, development of Response Templates, and implementation of Recommender Systems and LLM-based Response Generation. Our validation involves a multi-paradigmatic analysis using the USPTO Office Action database and longitudinal data of attorney interactions with our systems over six years. Through five studies, we examine the constructiveness of OA topics (studies 1 and 2) using topic modeling and the proposed Delphi process, the efficacy of our proposed hybrid recommender system tailored for OA (both LLM-based and non-LLM-based) (study 3), the quality of response generation (study 4), and the practical value of the systems in real-world scenarios via user studies (study 5). Results demonstrate that both PARIS and LE-PARIS significantly meet key metrics and positively impact attorney performance.
Many problems in robotics involve creating or breaking multiple contacts nearly simultaneously or in an indeterminate order. We present a novel general purpose numerical integrator based on the theory of Event Selected Systems (ESS). Many multicontact models are ESS, which has recently been shown to imply that despite a discontinuous vector field, the flow of these systems is continuous, piecewise smooth, and has a well defined orbital derivative for all trajectories, which can be rapidly computed. We provide an elementary proof that our integrator is first-order accurate and verify numerically that it is in fact second-order accurate as its construction anticipated. We also compare our integrator, implemented in NumPy, to a MuJoCo simulation on models with 2 to 100 contacts, and confirm that the increase in simulation time per contact is nearly identical. The results suggest that this novel integrator can be invaluable for modelling and control in many robotics applications.
The number of Language Models (LMs) dedicated to processing scientific text is on the rise. Keeping pace with the rapid growth of scientific LMs (SciLMs) has become a daunting task for researchers. To date, no comprehensive surveys on SciLMs have been undertaken, leaving this issue unaddressed. Given the constant stream of new SciLMs, appraising the state-of-the-art and how they compare to each other remain largely unknown. This work fills that gap and provides a comprehensive review of SciLMs, including an extensive analysis of their effectiveness across different domains, tasks and datasets, and a discussion on the challenges that lie ahead.
Full-duplex (FD) wireless can significantly enhance spectrum efficiency but requires effective self-interference (SI) cancellers. RF SI cancellation (SIC) via frequency-domain equalization (FDE), where bandpass filters channelize the SI, is suited for integrated circuits (ICs). In this paper, we explore the limits and higher layer challenges associated with using such cancellers. We evaluate the performance of a custom FDE-based canceller using two testbeds; one with mobile FD radios and the other with upgraded, static FD radios in the PAWR COSMOS testbed. The latter is a lasting artifact for the research community, alongside a dataset containing baseband waveforms captured on the COSMOS FD radios, facilitating FD-related experimentation at the higher networking layers. We evaluate the performance of the FDE-based FD radios in both testbeds, with experiments showing 95 dB overall achieved SIC (52 dB from RF SIC) across 20 MHz bandwidth, and an average link-level FD rate gain of 1.87x. We also conduct experiments in (i) uplink-downlink networks with inter-user interference, and (ii) heterogeneous networks with half-duplex and FD users. The experimental FD gains in the two types of networks depend on the users' SNR values and the number of FD users, and are 1.14x-1.25x and 1.25x-1.73x, respectively, confirming previous analytical results.
Despite the Internet's continued growth, it increasingly depends on a small set of service providers to support Domain Name System (DNS) and web content hosting. This trend poses many potential threats including susceptibility to outages, failures, and potential censorship by providers. This paper aims to quantify consolidation in terms of popular domains' reliance on a small set of organizations for both DNS and web hosting. We highlight the extent to which a set of relatively few platforms host the authoritative name servers and web content for the top million websites. Our results show that both DNS and web hosting are concentrated, with Cloudflare and Amazon hosting over $30\%$ of the domains for both services. With the addition of Akamai, Fastly, and Google, these five organizations host $60\%$ of index pages in the Tranco top 10K, as well as the majority of external page resources. These trends are consistent across six different global vantage points, indicating that consolidation is happening globally and popular organizations can influence users' online experience across the world.
Visual Question Answering (VQA) models have struggled with counting objects in natural images so far. We identify a fundamental problem due to soft attention in these models as a cause. To circumvent this problem, we propose a neural network component that allows robust counting from object proposals. Experiments on a toy task show the effectiveness of this component and we obtain state-of-the-art accuracy on the number category of the VQA v2 dataset without negatively affecting other categories, even outperforming ensemble models with our single model. On a difficult balanced pair metric, the component gives a substantial improvement in counting over a strong baseline by 6.6%.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191