Large Language Models (LLMs) are often misleadingly recognized as having a personality or a set of values. We argue that an LLM can be seen as a superposition of perspectives with different values and personality traits. LLMs exhibit context-dependent values and personality traits that change based on the induced perspective (as opposed to humans, who tend to have more coherent values and personality traits across contexts). We introduce the concept of perspective controllability, which refers to a model's affordance to adopt various perspectives with differing values and personality traits. In our experiments, we use questionnaires from psychology (PVQ, VSM, IPIP) to study how exhibited values and personality traits change based on different perspectives. Through qualitative experiments, we show that LLMs express different values when those are (implicitly or explicitly) implied in the prompt, and that LLMs express different values even when those are not obviously implied (demonstrating their context-dependent nature). We then conduct quantitative experiments to study the controllability of different models (GPT-4, GPT-3.5, OpenAssistant, StableVicuna, StableLM), the effectiveness of various methods for inducing perspectives, and the smoothness of the models' drivability. We conclude by examining the broader implications of our work and outline a variety of associated scientific questions. The project website is available at //sites.google.com/view/llm-superpositions .
相關內容
Order Dependencies (ODs) have many applications, such as query optimization, data integration, and data cleaning. Although many works addressed the problem of discovering OD (and its variants), they do not consider datasets with missing values, a standard observation in real-world datasets. This paper introduces the novel notion of Embedded ODs (eODs) to deal with missing values. The intuition of eODs is to confirm ODs only on tuples with no missing values on a given embedding (a set of attributes). In this paper, we address the problem of validating a given eOD. If the eOD holds, we return true. Otherwise, we search for an updated embedding such that the updated eOD holds. If such embedding does not exist, we return false. A trivial requirement is to consider an embedding such that the number of ignored tuples is minimized. We show that it is NP-complete to compute such embedding. We therefore propose an efficient heuristic algorithm for validating embedded ODs. We conduct experiments on real-world datasets, and the results confirm the efficiency of our algorithm.
The success of AI assistants based on language models (LLMs) hinges crucially on Reinforcement Learning from Human Feedback (RLHF), which enables the generation of responses more aligned with human preferences. As universal AI assistants, there's a growing expectation for them to perform consistently across various domains. However, previous work shows that Reinforcement Learning (RL) often exploits shortcuts to attain high rewards and overlooks challenging samples. This focus on quick reward gains undermines both the stability in training and the model's ability to generalize to new, unseen data. In this work, we propose a novel approach that can learn a consistent policy via RL across various data groups or domains. Given the challenges associated with acquiring group annotations, our method automatically classifies data into different groups, deliberately maximizing performance variance. Then, we optimize the policy to perform well on challenging groups. Lastly, leveraging the established groups, our approach adaptively adjusts the exploration space, allocating more learning capacity to more challenging data and preventing the model from over-optimizing on simpler data. Experimental results indicate that our approach significantly enhances training stability and model generalization.
This study proposes a method for distilling the knowledge of fine-tuned Large Language Models (LLMs) into a smaller, more efficient, and accurate neural network, specifically targeting the challenge of deploying these models on resource-constrained devices. Our methodology involves training the smaller student model using the prediction probabilities of the LLM, which serves as a teacher model. This is achieved through a specialized loss function tailored to learn from the LLM's output probabilities, ensuring that the student model closely mimics the teacher's performance. To test this approach, we utilized a large dataset, 7T, containing 6,684 student-written responses to science questions and three other datasets with student-written responses. We also compared performance with original neural network (NN) models to validate the accuracy. Results have shown that the NN and distilled student models have comparable accuracy to the teacher model for the 7T dataset; however, other datasets have shown significantly lower accuracy (28% on average) for NN, though our proposed distilled model is still able to achieve 12\% higher accuracy than NN. Furthermore, the student model size ranges from 0.1M to 0.02M, 100 times smaller in terms of parameters and ten times smaller compared with the original output model size. The significance of this research lies in its potential to make advanced AI technologies accessible in typical educational settings, particularly for automatic scoring.
The accurate syllabification of words plays a vital role in various Natural Language Processing applications. Syllabification is a versatile linguistic tool with applications in linguistic research, language technology, education, and various fields where understanding and processing language is essential. In this paper, we present a comprehensive approach to syllabification for the Uzbek language, including rule-based techniques and machine learning algorithms. Our rule-based approach utilizes advanced methods for dividing words into syllables, generating hyphenations for line breaks and count of syllables. Additionally, we collected a dataset for evaluating and training using machine learning algorithms comprising word-syllable mappings, hyphenations, and syllable counts to predict syllable counts as well as for the evaluation of the proposed model. Our results demonstrate the effectiveness and efficiency of both approaches in achieving accurate syllabification. The results of our experiments show that both approaches achieved a high level of accuracy, exceeding 99%. This study provides valuable insights and recommendations for future research on syllabification and related areas in not only the Uzbek language itself, but also in other closely-related Turkic languages with low-resource factor.
Every Model of High-Level Computation (MHC) has an underlying composition mechanism for combining simple computation devices into more complex ones. Composition can be done by (explicitly or implicitly) defining control flow, data flow or any combination thereof. Control flow specifies the order in which individual computations are activated, whereas data flow defines how data is exchanged among them. Unfortunately, traditional MHCs either mix data and control or only consider one dimension explicitly, which makes it difficult to reason about data flow and control flow separately. Reasoning about these dimensions orthogonally is a crucial desideratum for optimisation, maintainability and verification purposes. In this paper, we introduce a novel MHC that explicitly treats data flow and control flow as separate dimensions, while providing modularity. As the model is rooted in category theory, it provides category-theoretic operations for compositionally constructing sequential, parallel, branchial or iterative composites. Compositionality entails that a composite exhibits the same properties as its respective constituents, including separation of concerns and modularity. We conclude the paper by demonstrating how our proposal can be used to model high-level computations in two different application domains: software engineering and artificial intelligence.
Foundation models like the Segment Anything Model (SAM) have demonstrated promise in generic object segmentation. However, directly applying SAM to surgical instrument segmentation presents key challenges. First, SAM relies on per-frame point-or-box prompts which complicate surgeon-computer interaction. Also, SAM yields suboptimal performance on segmenting surgical instruments, owing to insufficient surgical data in its pre-training as well as the complex structure and fine-grained details of various surgical instruments. To address these challenges, in this paper, we investigate text promptable surgical instrument segmentation and propose SP-SAM (SurgicalPart-SAM), a novel efficient-tuning approach that integrates surgical instrument structure knowledge with the generic segmentation knowledge of SAM. Specifically, we achieve this by proposing (1) collaborative prompts in the text form "[part name] of [instrument category name]" that decompose instruments into fine-grained parts; (2) a Cross-Modal Prompt Encoder that encodes text prompts jointly with visual embeddings into discriminative part-level representations; and (3) a Part-to-Whole Selective Fusion and a Hierarchical Decoding strategy that selectively assemble the part-level representations into a whole for accurate instrument segmentation. Built upon them, SP-SAM acquires a better capability to comprehend surgical instrument structures and distinguish between various categories. Extensive experiments on both the EndoVis2018 and EndoVis2017 datasets demonstrate SP-SAM's state-of-the-art performance with minimal tunable parameters. Code is at //github.com/wenxi-yue/SurgicalPart-SAM.
Cloud Service Providers, such as Google Cloud Platform, Microsoft Azure, or Amazon Web Services, offer continuously evolving cloud services. It is a growing industry. Businesses, such as Netflix and PayPal, rely on the Cloud for data storage, computing power, and other services. For businesses, the cloud reduces costs, provides flexibility, and allows for growth. However, there are security and privacy concerns regarding the Cloud. Because Cloud services are accessed through the internet, hackers and attackers could possibly access the servers from anywhere. To protect data in the Cloud, it should be encrypted before it is uploaded, it should be protected in storage and also in transit. On the other hand, data owners may need to access their encrypted data. It may also need to be altered, updated, deleted, read, searched, or shared with others. If data is decrypted in the Cloud, sensitive data is exposed and could be exposed and misused. One solution is to leave the data in its encrypted form and use Searchable Encryption (SE) which operates on encrypted data. The functionality of SE has improved since its inception and research continues to explore ways to improve SE. This paper reviews the functionality of Searchable Encryption, mostly related to Cloud services, in the years 2019 to 2023, and evaluates one of its schemes, Fully Homomorphic Encryption. Overall, it seems that research is at the point where SE efficiency is increased as multiple functionalities are aggregated and tested.
Large Language Models (LLMs) have shown excellent generalization capabilities that have led to the development of numerous models. These models propose various new architectures, tweaking existing architectures with refined training strategies, increasing context length, using high-quality training data, and increasing training time to outperform baselines. Analyzing new developments is crucial for identifying changes that enhance training stability and improve generalization in LLMs. This survey paper comprehensively analyses the LLMs architectures and their categorization, training strategies, training datasets, and performance evaluations and discusses future research directions. Moreover, the paper also discusses the basic building blocks and concepts behind LLMs, followed by a complete overview of LLMs, including their important features and functions. Finally, the paper summarizes significant findings from LLM research and consolidates essential architectural and training strategies for developing advanced LLMs. Given the continuous advancements in LLMs, we intend to regularly update this paper by incorporating new sections and featuring the latest LLM models.
Graph Neural Networks (GNNs) have been studied from the lens of expressive power and generalization. However, their optimization properties are less well understood. We take the first step towards analyzing GNN training by studying the gradient dynamics of GNNs. First, we analyze linearized GNNs and prove that despite the non-convexity of training, convergence to a global minimum at a linear rate is guaranteed under mild assumptions that we validate on real-world graphs. Second, we study what may affect the GNNs' training speed. Our results show that the training of GNNs is implicitly accelerated by skip connections, more depth, and/or a good label distribution. Empirical results confirm that our theoretical results for linearized GNNs align with the training behavior of nonlinear GNNs. Our results provide the first theoretical support for the success of GNNs with skip connections in terms of optimization, and suggest that deep GNNs with skip connections would be promising in practice.
While it is nearly effortless for humans to quickly assess the perceptual similarity between two images, the underlying processes are thought to be quite complex. Despite this, the most widely used perceptual metrics today, such as PSNR and SSIM, are simple, shallow functions, and fail to account for many nuances of human perception. Recently, the deep learning community has found that features of the VGG network trained on the ImageNet classification task has been remarkably useful as a training loss for image synthesis. But how perceptual are these so-called "perceptual losses"? What elements are critical for their success? To answer these questions, we introduce a new Full Reference Image Quality Assessment (FR-IQA) dataset of perceptual human judgments, orders of magnitude larger than previous datasets. We systematically evaluate deep features across different architectures and tasks and compare them with classic metrics. We find that deep features outperform all previous metrics by huge margins. More surprisingly, this result is not restricted to ImageNet-trained VGG features, but holds across different deep architectures and levels of supervision (supervised, self-supervised, or even unsupervised). Our results suggest that perceptual similarity is an emergent property shared across deep visual representations.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191