As conversational models become increasingly available to the general public, users are engaging with this technology in social interactions. Such unprecedented interaction experiences may pose considerable social and psychological risks to the users unless the technology is properly controlled. This highlights the need for scalable and robust evaluation metrics for conversational chatbots. Existing evaluation metrics aim to automate offline user evaluation and approximate human judgment of pre-curated dialogs. However, they are limited in their ability to capture subjective perceptions of users who actually interact with the bots and might not generalize to real-world settings. To address this limitation, we propose an approach to approximate online human evaluation leveraging large language models (LLMs) from the GPT family. We introduce a new Dialog system Evaluation framework based on Prompting (DEP), which enables a fully automatic evaluation pipeline that replicates live user studies and achieves an impressive correlation with human judgment (up to Pearson r=0.95 on a system level). The DEP approach involves collecting synthetic chat logs of evaluated bots with an LLM in the other-play setting, where the LLM is carefully conditioned to follow a specific scenario. We further explore different prompting approaches to produce evaluation scores with the same LLM. The best performing prompts, which contain few-shot demonstrations and instructions, show outstanding performance on the tested dataset and demonstrate the ability to generalize to other dialog corpora.
相關內容
As large language models increase in capability, researchers have started to conduct surveys of all kinds on these models with varying scientific motivations. In this work, we examine what we can learn from language models' survey responses on the basis of the well-established American Community Survey (ACS) by the U.S. Census Bureau. Using a de-facto standard multiple-choice prompting technique and evaluating 40 different language models, hundreds of thousands of times each on questions from the ACS, we systematically establish two dominant patterns. First, models have significant position and labeling biases, for example, towards survey responses labeled with the letter "A". Second, when adjusting for labeling biases through randomized answer ordering, models across the board trend towards uniformly random survey responses. In fact, binary classifiers can almost perfectly differentiate between models' responses to the ACS and the responses of the US census. Taken together, our findings suggest caution in treating survey responses from language models as equivalent to those of human populations at present time.
Classifying policy documents into policy issue topics has been a long-time effort in political science and communication disciplines. Efforts to automate text classification processes for social science research purposes have so far achieved remarkable results, but there is still a large room for progress. In this work, we test the prediction performance of an alternative strategy, which requires human involvement much less than full manual coding. We use the GPT 3.5 and GPT 4 models of the OpenAI, which are pre-trained instruction-tuned Large Language Models (LLM), to classify congressional bills and congressional hearings into Comparative Agendas Project's 21 major policy issue topics. We propose three use-case scenarios and estimate overall accuracies ranging from %58-83 depending on scenario and GPT model employed. The three scenarios aims at minimal, moderate, and major human interference, respectively. Overall, our results point towards the insufficiency of complete reliance on GPT with minimal human intervention, an increasing accuracy along with the human effort exerted, and a surprisingly high accuracy achieved in the most humanly demanding use-case. However, the superior use-case achieved the %83 accuracy on the %65 of the data in which the two models agreed, suggesting that a similar approach to ours can be relatively easily implemented and allow for mostly automated coding of a majority of a given dataset. This could free up resources allowing manual human coding of the remaining %35 of the data to achieve an overall higher level of accuracy while reducing costs significantly.
For a robot to personalize physical assistance effectively, it must learn user preferences that can be generally reapplied to future scenarios. In this work, we investigate personalization of household cleanup with robots that can tidy up rooms by picking up objects and putting them away. A key challenge is determining the proper place to put each object, as people's preferences can vary greatly depending on personal taste or cultural background. For instance, one person may prefer storing shirts in the drawer, while another may prefer them on the shelf. We aim to build systems that can learn such preferences from just a handful of examples via prior interactions with a particular person. We show that robots can combine language-based planning and perception with the few-shot summarization capabilities of large language models (LLMs) to infer generalized user preferences that are broadly applicable to future interactions. This approach enables fast adaptation and achieves 91.2% accuracy on unseen objects in our benchmark dataset. We also demonstrate our approach on a real-world mobile manipulator called TidyBot, which successfully puts away 85.0% of objects in real-world test scenarios.
An important aspect in developing language models that interact with humans is aligning their behavior to be useful and unharmful for their human users. This is usually achieved by tuning the model in a way that enhances desired behaviors and inhibits undesired ones, a process referred to as alignment. In this paper, we propose a theoretical approach called Behavior Expectation Bounds (BEB) which allows us to formally investigate several inherent characteristics and limitations of alignment in large language models. Importantly, we prove that within the limits of this framework, for any behavior that has a finite probability of being exhibited by the model, there exist prompts that can trigger the model into outputting this behavior, with probability that increases with the length of the prompt. This implies that any alignment process that attenuates an undesired behavior but does not remove it altogether, is not safe against adversarial prompting attacks. Furthermore, our framework hints at the mechanism by which leading alignment approaches such as reinforcement learning from human feedback make the LLM prone to being prompted into the undesired behaviors. This theoretical result is being experimentally demonstrated in large scale by the so called contemporary "chatGPT jailbreaks", where adversarial users trick the LLM into breaking its alignment guardrails by triggering it into acting as a malicious persona. Our results expose fundamental limitations in alignment of LLMs and bring to the forefront the need to devise reliable mechanisms for ensuring AI safety.
Recent advances have greatly increased the capabilities of large language models (LLMs), but our understanding of the models and their safety has not progressed as fast. In this paper we aim to understand LLMs deeper by studying their individual neurons. We build upon previous work showing large language models such as GPT-4 can be useful in explaining what each neuron in a language model does. Specifically, we analyze the effect of the prompt used to generate explanations and show that reformatting the explanation prompt in a more natural way can significantly improve neuron explanation quality and greatly reduce computational cost. We demonstrate the effects of our new prompts in three different ways, incorporating both automated and human evaluations.
Deep learning models have achieved state-of-the-art performances in various domains, while they are vulnerable to the inputs with well-crafted but small perturbations, which are named after adversarial examples (AEs). Among many strategies to improve the model robustness against AEs, Projected Gradient Descent (PGD) based adversarial training is one of the most effective methods. Unfortunately, the prohibitive computational overhead of generating strong enough AEs, due to the maximization of the loss function, sometimes makes the regular PGD adversarial training impractical when using larger and more complicated models. In this paper, we propose that the adversarial loss can be approximated by the partial sum of Taylor series. Furthermore, we approximate the gradient of adversarial loss and propose a new and efficient adversarial training method, adversarial training with gradient approximation (GAAT), to reduce the cost of building up robust models. Additionally, extensive experiments demonstrate that this efficiency improvement can be achieved without any or with very little loss in accuracy on natural and adversarial examples, which show that our proposed method saves up to 60\% of the training time with comparable model test accuracy on MNIST, CIFAR-10 and CIFAR-100 datasets.
Modern software heavily relies on the use of components. Those components are usually published in central repositories, and managed by build systems via dependencies. Due to issues around vulnerabilities, licenses and the propagation of bugs, the study of those dependencies is of utmost importance, and numerous software composition analysis tools have emerged for this purpose. A particular challenge are hidden dependencies that are the result of cloning or shading where code from a component is "inlined", and, in the case of shading, moved to different namespaces. We present a novel approach to detect vulnerable clones in the Maven repository. Our approach is lightweight in that it does not require the creation and maintenance of a custom index. Starting with 29 vulnerabilities with assigned CVEs and proof-of-vulnerability projects, we retrieve over 53k potential vulnerable clones from Maven Central. After running our analysis on this set, we detect 727 confirmed vulnerable clones (86 if versions are aggregated) and synthesize a testable proof-of-vulnerability project for each of those. We demonstrate that existing SCA tools often miss those exposures. At the time of submission those results have led to changes to the entries for six CVEs in the GitHub Security Advisory Database (GHSA) via accepted pull requests, with more pending.
As artificial intelligence (AI) models continue to scale up, they are becoming more capable and integrated into various forms of decision-making systems. For models involved in moral decision-making, also known as artificial moral agents (AMA), interpretability provides a way to trust and understand the agent's internal reasoning mechanisms for effective use and error correction. In this paper, we provide an overview of this rapidly-evolving sub-field of AI interpretability, introduce the concept of the Minimum Level of Interpretability (MLI) and recommend an MLI for various types of agents, to aid their safe deployment in real-world settings.
Graph neural networks (GNNs) have been demonstrated to be a powerful algorithmic model in broad application fields for their effectiveness in learning over graphs. To scale GNN training up for large-scale and ever-growing graphs, the most promising solution is distributed training which distributes the workload of training across multiple computing nodes. However, the workflows, computational patterns, communication patterns, and optimization techniques of distributed GNN training remain preliminarily understood. In this paper, we provide a comprehensive survey of distributed GNN training by investigating various optimization techniques used in distributed GNN training. First, distributed GNN training is classified into several categories according to their workflows. In addition, their computational patterns and communication patterns, as well as the optimization techniques proposed by recent work are introduced. Second, the software frameworks and hardware platforms of distributed GNN training are also introduced for a deeper understanding. Third, distributed GNN training is compared with distributed training of deep neural networks, emphasizing the uniqueness of distributed GNN training. Finally, interesting issues and opportunities in this field are discussed.
Autonomic computing investigates how systems can achieve (user) specified control outcomes on their own, without the intervention of a human operator. Autonomic computing fundamentals have been substantially influenced by those of control theory for closed and open-loop systems. In practice, complex systems may exhibit a number of concurrent and inter-dependent control loops. Despite research into autonomic models for managing computer resources, ranging from individual resources (e.g., web servers) to a resource ensemble (e.g., multiple resources within a data center), research into integrating Artificial Intelligence (AI) and Machine Learning (ML) to improve resource autonomy and performance at scale continues to be a fundamental challenge. The integration of AI/ML to achieve such autonomic and self-management of systems can be achieved at different levels of granularity, from full to human-in-the-loop automation. In this article, leading academics, researchers, practitioners, engineers, and scientists in the fields of cloud computing, AI/ML, and quantum computing join to discuss current research and potential future directions for these fields. Further, we discuss challenges and opportunities for leveraging AI and ML in next generation computing for emerging computing paradigms, including cloud, fog, edge, serverless and quantum computing environments.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191