检索结果-内蒙古大学图书馆

SCANS: Mitigating the Exaggerated Safety for LLMs via Safety-Conscious Activation Steering

学校读者我要写书评

暂无评论

arXiv 2024年

作者： Cao, Zouying Yang, Yifei Zhao, Hai Department of Computer Science and Engineering Shanghai Jiao Tong University Key Laboratory of Shanghai Education Commission for Intelligent Interaction and Cognitive Engineering Shanghai Jiao Tong University Shanghai Key Laboratory of Trusted Data Circulation and Governance in Web3 China

Safety alignment is indispensable for Large Language Models (LLMs) to defend threats from malicious instructions. However, recent researches reveal safety-aligned LLMs prone to reject benign queries due to the exaggerated safety issue, limiting their helpfulness. In this paper, we propose a Safety-Conscious Activation Steering (SCANS) method to mitigate the exaggerated safety concerns in aligned LLMs. First, SCANS extracts the refusal steering vectors within the activation space and utilizes vocabulary projection to anchor some specific safety-critical layers which influence model refusal behavior. Second, by tracking the hidden state transition, SCANS identifies the steering direction and steers the model behavior accordingly, achieving a balance between exaggerated safety and adequate safety. Experiments show that SCANS achieves new state-of-the-art performance on XSTest and OKTest benchmarks, without impairing their defense capability against harmful queries and maintaining almost unchanged model capability. Copyright © 2024, The Authors. All rights reserved.

关键词： Benchmarking

EM Pre-training for Multi-party Dialogue Response Generation

学校读者我要写书评

暂无评论

arXiv 2023年

作者： Li, Yiyang Zhao, Hai Department of Computer Science and Engineering Shanghai Jiao Tong University China Key Laboratory of Shanghai Education Commission for Intelligent Interaction and Cognitive Engineering Shanghai Jiao Tong University China

Dialogue response generation requires an agent to generate a response according to the current dialogue history, in terms of which two-party dialogues have been well studied, but leaving a great gap for multi-party dialogues at the same time. Different from two-party dialogues where each response is a direct reply to its previous utterance, the addressee of a response utterance should be specified before it is generated in the multi-party scenario. Thanks to the huge amount of two-party conversational data, various pre-trained language models for two-party dialogue response generation have been proposed. However, due to the lack of annotated addressee labels in multiparty dialogue datasets, it is hard to use them to pre-train a response generation model for multi-party dialogues. To tackle this obstacle, we propose an Expectation-Maximization (EM) approach that iteratively performs the expectation steps to generate addressee labels, and the maximization steps to optimize a response generation model. Theoretical analyses and extensive experiments have justified the feasibility and effectiveness of our proposed method. The official implementation of this paper is available at https://***/EricLee8/MPDRG. © 2023, CC BY.

关键词： Iterative methods

Multi-Turn Dialogue Comprehension from a Topic-Aware Perspective

学校读者我要写书评

暂无评论

SSRN

SSRN 2023年

作者： Ma, Xinbei Xu, Yi Zhao, Hai Zhang, Zhuosheng Department of Computer Science and Engineering Shanghai Jiao Tong University China Key Laboratory of Shanghai Education Commission for Intelligent Interaction and Cognitive Engineering Shanghai Jiao Tong University China

Dialogue Machine Reading Comprehension requires language models to effectively decouple and model multi-turn dialogue passages. As a dialogue development goes after the intentions of participants, its topic may not remain constant throughout the whole passage. Hence, it is non-trivial to detect and leverage the topic shift in dialogue modeling. Topic modeling, although has been widely studied in plain text, deserves far more utilization in dialogue reading comprehension. This paper proposes to model multi-turn dialogues from a topic-aware perspective. We start with a dialogue segmentation algorithm to split a dialogue passage into topic-concentrated fragments in an unsupervised way. Then we use these fragments as topic-aware language processing units in further dialogue comprehension. On one hand, the split segments indict specific topics rather than mixed intentions, thus showing convenience on in-domain topic detection and location. For this task, we design a clustering system with a self-training auto-encoder, and we build two constructed datasets for evaluation. On the other hand, the split segments are an appropriate element of multi-turn dialogue response selection. For this purpose, we further present a novel model, Topic-Aware Dual-Attention Matching (TADAM) Network, which takes topic segments as processing elements and matches response candidates with a dual cross-attention. Empirical studies on three public benchmarks show great improvements over baselines. Our work continues the previous studies on document topic, and brings the dialogue modeling to a novel topic-aware perspective with exhaustive experiments and analyses. © 2023, The Authors. All rights reserved.

关键词：

Design lessons from AI’s two grand goals: human emulation and useful applications

学校读者我要写书评

暂无评论

IEEE Transactions on Technology and Society

IEEE Transactions on Technology and Society 2020年第2期1卷 73-82页

作者： Shneiderman, Ben Department of Computer Science University of Maryland at College Park College ParkMD20742 United States Human–Computer Interaction Laboratory University of Maryland at College Park College ParkMD20742 United States

Researchers’ goals shape the questions they raise, collaborators they choose, methods they use, and outcomes of their work. This article offers a fresh vision of artificial intelligence (AI) research by suggesting a simplification to two goals: 1) emulation to understand human abilities to build systems that perform tasks as well as or better than humans and 2) application of AI methods to build widely used products and services. Researchers and developers for each goal can fruitfully work along their desired paths, but this article is intended to limit the problems that arise when assumptions from one goal are used to drive work on the other goal. For example, autonomous humanoid robots are prominent with emulation researchers, but application developers avoid them, in favor of tool-like appliances or teleoperated devices for widely used commercial products and services. This article covers four such mismatches in goals that affect AI-guided application development: 1) intelligent agent or powerful tool;2) simulated teammate or teleoperated device;3) autonomous system or supervisory control;and 4) humanoid robot or mechanoid appliance. This article clarifies these mismatches to facilitate the discovery of workable compromise designs that will accelerate human-centered AI applications research. A greater emphasis on human-centered AI could reduce AI’s existential threats and increase benefits for users and society, such as in business, education, healthcare, environmental preservation, and community safety. © 2020 IEEE. All right reserved.

关键词： Anthropomorphic robots

Evaluating the Benefits of the Immersive Space to Think

学校读者我要写书评

暂无评论

Evaluating the Benefits of the Immersive Space to Think

2020 IEEE Conference on Virtual Reality and 3D User Interfaces, VRW 2020

作者： Lisle, Lee Chen, Xiaoyu Edward Gitre, J.K. North, Chris Bowman, Doug A. Center for Human-Computer Interaction Department of Computer Science Virginia Tech United States Virginia Tech Grado Department of Industrial and Systems Engineering United States Center for Human-Computer Interaction Department of History Virginia Tech United States

ISBN: (纸本)9781728165325

Sensemaking with large multimedia dataset is a cognitively intensive task that requires analysts to understand the underlying stories that the dataset tells. Often, analysts use tools in order to offload cognition as well as convey their new understanding of the dataset;however, existing tools are limited by their underlying technologies. We have proposed a novel virtual reality tool to support sensemaking called the Immersive Space to Think (IST). IST can aid the process of analyzing multimedia data, but it remains unproven whether IST improves sensemaking performance over a traditional desktop setting. In a study performed over six weeks, one participant used both IST and traditional methods of sensemaking with a dataset of 100 text documents of transcribed survey responses from World War 2 soldiers to perform historical analysis. The participant was asked guided questions that produced three essays with their understanding of the data. After conducting a blind evaluation of the participant's interpretation of the data, a team of three experts in historical analysis concluded that the essays written with IST displayed a better understanding of the dataset. Furthermore, the participant gave positive feedback on IST, and also suggested possible improvements. © 2020 IEEE.

关键词： Surveys

Toward Efficient Automated Feature Engineering 39

学校读者我要写书评

暂无评论

Toward Efficient Automated Feature Engineering

39th IEEE International Conference on Data Engineering, ICDE 2023

作者： Wang, Kafeng Wang, Pengyang Xu, Chengzhong Shenzhen Institute of Advanced Technology Chinese Academy of Sciences China University of Chinese Academy of Sciences University of Macau China Guangdong-Hong Kong-Macao Joint Laboratory of Human-Machine Intelligence-Synergy Systems Shenzhen China State Key Laboratory of Internet of Things for Smart City Department of Computer and Information Science University of Macau China

ISBN: (纸本)9798350322279

Automated Feature Engineering (AFE) refers to automatically generate and select optimal feature sets for downstream tasks, which has achieved great success in real-world applications. Current AFE methods mainly focus on improving the effectiveness of the produced features, but ignoring the low-efficiency issue for large-scale deployment. Therefore, in this work, we propose a generic framework to improve the efficiency of AFE. Specifically, we construct the AFE pipeline based on reinforcement learning setting, where each feature is assigned an agent to perform feature transformation and selection, and the evaluation score of the produced features in downstream tasks serve as the reward to update the policy. We improve the efficiency of AFE in two perspectives. On the one hand, we develop a Feature Pre-Evaluation (FPE) Model to reduce the sample size and feature size that are two main factors on undermining the efficiency of feature evaluation. On the other hand, we devise a two-stage policy training strategy by running FPE on the pre-evaluation task as the initialization of the policy to avoid training policy from scratch. We conduct comprehensive experiments on 36 datasets in terms of both classification and regression tasks. The results show 2.9% higher performance in average and 2x higher computational efficiency comparing to state-of-the-art AFE methods. © 2023 IEEE.

关键词： Reinforcement learning

Let's Influence Algorithms Together: How Millions of Fans Build Collective Understanding of Algorithms and Organize Coordinated Algorithmic Actions

学校读者我要写书评

暂无评论

arXiv 2024年

作者： Xiao, Qing Zheng, Yuhang Fan, Xianzhe Zhang, Bingbing Lu, Zhicong Human-Computer Interaction Institute Carnegie Mellon University PittsburghPA United States University of Amsterdam Amsterdam Netherlands Tsinghua University Beijing China School of Journalism and Mass Communication University of Iowa Iowa CityIA United States Department of Computer Science George Mason University FairfaxVA United States

Previous research pays attention to how users strategically understand and consciously interact with algorithms but mainly focuses on an individual level, making it difficult to explore how users within communities could develop a collective understanding of algorithms and organize collective algorithmic actions. Through a two-year ethnography of online fan activities, this study investigates 43 core fans who always organize large-scale fans collective actions and their corresponding general fan groups. This study aims to reveal how these core fans mobilize millions of general fans through collective algorithmic actions. These core fans reported the rhetorical strategies used to persuade general fans, the steps taken to build a collective understanding of algorithms, and the collaborative processes that adapt collective actions across platforms and cultures. Our findings highlight the key factors that enable computer-supported collective algorithmic actions and extend collective action research into the large-scale domain targeting algorithms. Copyright © 2024, The Authors. All rights reserved.

关键词： Consensus algorithm

BatGPT-Chem: A Foundation Large Model For Retrosynthesis Prediction

学校读者我要写书评

暂无评论

arXiv 2024年

作者： Yang, Yifei Shi, Runhan Li, Zuchao Jiang, Shu Lu, Bao-Liang Yang, Yang Zhao, Hai Department of Computer Science and Engineering Key Laboratory of Shanghai Education Commission for Intelligent Interaction and Cognitive Engineering Shanghai Jiao Tong University 800 Dongchuan Road Shanghai200240 China Wuhan University China Nantong University China

Retrosynthesis analysis is pivotal yet challenging in drug discovery and organic chemistry. Despite the proliferation of computational tools over the past decade, AI-based systems often fall short in generalizing across diverse reaction types and exploring alternative synthetic pathways. This paper presents BatGPT-Chem, a large language model with 15 billion parameters, tailored for enhanced retrosynthesis prediction. Integrating chemical tasks via a unified framework of natural language and SMILES notation, this approach synthesizes extensive instructional data from an expansive chemical database. Employing both autoregressive and bidirectional training techniques across over one hundred million instances, BatGPT-Chem captures a broad spectrum of chemical knowledge, enabling precise prediction of reaction conditions and exhibiting strong zero-shot capabilities. Superior to existing AI methods, our model demonstrates significant advancements in generating effective strategies for complex molecules, as validated by stringent benchmark tests. BatGPT-Chem not only boosts the efficiency and creativity of retrosynthetic analysis but also establishes a new standard for computational tools in synthetic design. This development empowers chemists to adeptly address the synthesis of novel compounds, potentially expediting the innovation cycle in drug manufacturing and materials science. We release our trial platform at https://***/dapp/chem. Copyright © 2024, The Authors. All rights reserved.

关键词： Synthesis (chemical)

Real-Time Imitation of human Head Motions, Blinks and Emotions by Nao Robot: A Closed-Loop Approach

学校读者我要写书评

暂无评论

Real-Time Imitation of Human Head Motions, Blinks and Emotio...

RSI/ISM International Conference on Robotics and Mechatronics (ICROM)

作者： Keyhan Rayati Amirhossein Feizi Alireza Beigy Pourya Shahverdi Mehdi Tale Masouleh Ahmad Kalhor Wing-Yue Geoffrey Louie Human and Robot Interaction Lab Electrical and Computer Engineering University of Tehran Tehran Iran Department of Electrical Engineering Shahrood University of Technology Shahrood Iran Intelligent Robotics Laboratory Oakland University Michigan USA

This paper introduces a novel approach for enabling real-time imitation of human head motion by a Nao robot, with a primary focus on elevating human-robot interactions. By using the robust capabilities of the MediaPipe as a computer vision library and the DeepFace as an emotion recognition library, this research endeavors to capture the subtleties of human head motion, including blink actions and emotional expressions, and seamlessly incorporate these indicators into the robot’s responses. The result is a comprehensive framework which facilitates precise head imitation within human-robot interactions, utilizing a closed-loop approach that involves gathering real-time feedback from the robot’s imitation performance. This feedback loop ensures a high degree of accuracy in modeling head motion, as evidenced by an impressive R2 score of 96.3 for pitch and 98.9 for yaw. Notably, the proposed approach holds promise in improving communication for children with autism, offering them a valuable tool for more effective interaction. In essence, proposed work explores the integration of real-time head imitation and real-time emotion recognition to enhance human-robot interactions, with potential benefits for individuals with unique communication needs.

关键词：