IDEATOR: Jailbreaking and Benchmarking Large Vision-Language Models Using Themselves

<!DOCTYPE html>     <html class="no-js" lang="en">  <head> <title>IDEATOR: Jailbreaking and Benchmarking Large Vision-Language Models Using Themselves - ADS</title>  <link rel="apple-touch-icon" sizes="180x180" href="//styles/favicon/apple-touch-icon.png" /> <link rel="icon" type="image/png" sizes="32x32" href="//styles/favicon/favicon-32x32.png" /> <link rel="icon" type="image/png" sizes="16x16" href="//styles/favicon/favicon-16x16.png" /> <link rel="manifest" href="//styles/favicon/site.webmanifest" /> <link rel="mask-icon" href="//styles/favicon/safari-pinned-tab.svg" color="#5bbad5" /> <meta name="apple-mobile-web-app-title" content="ADS" /> <meta name="application-name" content="ADS" /> <meta name="msapplication-TileColor" content="#ffc40d" /> <meta name="theme-color" content="#ffffff" />   <script>(function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start': new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0], j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src= 'https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f); })(window,document,'script','dataLayer','GTM-NT2453N');</script>  <link rel="stylesheet" href="/styles/css/styles.css"> <meta name="robots" content="noarchive"> <link rel="canonical" href="http://ui.adsabs.harvard.edu/abs/2024arXiv241100827W/abstract"/> <meta name="description" content="As large Vision-Language Models (VLMs) gain prominence, ensuring their safe deployment has become critical. Recent studies have explored VLM robustness against jailbreak attacks-techniques that exploit model vulnerabilities to elicit harmful outputs. However, the limited availability of diverse multimodal data has constrained current approaches to rely heavily on adversarial or manually crafted images derived from harmful text datasets, which often lack effectiveness and diversity across different contexts. In this paper, we propose IDEATOR, a novel jailbreak method that autonomously generates malicious image-text pairs for black-box jailbreak attacks. IDEATOR is grounded in the insight that VLMs themselves could serve as powerful red team models for generating multimodal jailbreak prompts. Specifically, IDEATOR leverages a VLM to create targeted jailbreak texts and pairs them with jailbreak images generated by a state-of-the-art diffusion model. Extensive experiments demonstrate IDEATOR's high effectiveness and transferability, achieving a 94% attack success rate (ASR) in jailbreaking MiniGPT-4 with an average of only 5.34 queries, and high ASRs of 82%, 88%, and 75% when transferred to LLaVA, InstructBLIP, and Chameleon, respectively. Building on IDEATOR's strong transferability and automated process, we introduce the VLBreakBench, a safety benchmark comprising 3,654 multimodal jailbreak samples. Our benchmark results on 11 recently released VLMs reveal significant gaps in safety alignment. For instance, our challenge set achieves ASRs of 46.31% on GPT-4o and 19.65% on Claude-3.5-Sonnet, underscoring the urgent need for stronger defenses.">  <meta property="og:type" content="eprint"> <meta property="og:title" content="IDEATOR: Jailbreaking and Benchmarking Large Vision-Language Models Using Themselves"> <meta property="og:site_name" content="ADS"> <meta property="og:description" content="As large Vision-Language Models (VLMs) gain prominence, ensuring their safe deployment has become critical. Recent studies have explored VLM robustness against jailbreak attacks-techniques that exploit model vulnerabilities to elicit harmful outputs. However, the limited availability of diverse multimodal data has constrained current approaches to rely heavily on adversarial or manually crafted images derived from harmful text datasets, which often lack effectiveness and diversity across different contexts. In this paper, we propose IDEATOR, a novel jailbreak method that autonomously generates malicious image-text pairs for black-box jailbreak attacks. IDEATOR is grounded in the insight that VLMs themselves could serve as powerful red team models for generating multimodal jailbreak prompts. Specifically, IDEATOR leverages a VLM to create targeted jailbreak texts and pairs them with jailbreak images generated by a state-of-the-art diffusion model. Extensive experiments demonstrate IDEATOR's high effectiveness and transferability, achieving a 94% attack success rate (ASR) in jailbreaking MiniGPT-4 with an average of only 5.34 queries, and high ASRs of 82%, 88%, and 75% when transferred to LLaVA, InstructBLIP, and Chameleon, respectively. Building on IDEATOR's strong transferability and automated process, we introduce the VLBreakBench, a safety benchmark comprising 3,654 multimodal jailbreak samples. Our benchmark results on 11 recently released VLMs reveal significant gaps in safety alignment. For instance, our challenge set achieves ASRs of 46.31% on GPT-4o and 19.65% on Claude-3.5-Sonnet, underscoring the urgent need for stronger defenses."> <meta property="og:url" content="https://ui.adsabs.harvard.edu/abs/2024arXiv241100827W/abstract"> <meta property="og:image" content="https://ui.adsabs.harvard.edu/styles/img/transparent_logo.svg"> <meta property="article:published_time" content="10/2024"> <meta property="article:author" content="Wang, Ruofan"> <meta property="article:author" content="Li, Juncheng"> <meta property="article:author" content="Wang, Yixu"> <meta property="article:author" content="Wang, Bo"> <meta property="article:author" content="Wang, Xiaosen"> <meta property="article:author" content="Teng, Yan"> <meta property="article:author" content="Wang, Yingchun"> <meta property="article:author" content="Ma, Xingjun"> <meta property="article:author" content="Jiang, Yu-Gang">  <meta name="citation_journal_title" content="arXiv e-prints"> <meta name="citation_authors" content="Wang, Ruofan;Li, Juncheng;Wang, Yixu;Wang, Bo;Wang, Xiaosen;Teng, Yan;Wang, Yingchun;Ma, Xingjun;Jiang, Yu-Gang"> <meta name="citation_title" content="IDEATOR: Jailbreaking and Benchmarking Large Vision-Language Models Using Themselves"> <meta name="citation_date" content="10/2024"> <meta name="citation_firstpage" content="arXiv:2411.00827"> <meta name="citation_doi" content="10.48550/arXiv.2411.00827"> <meta name="citation_language" content="en"> <meta name="citation_keywords" content="Computer Science - Computer Vision and Pattern Recognition"> <meta name="citation_keywords" content="Computer Science - Artificial Intelligence"> <meta name="citation_abstract_html_url" content="https://ui.adsabs.harvard.edu/abs/2024arXiv241100827W/abstract"> <meta name="citation_publication_date" content="10/2024"> <meta name="citation_arxiv_id" content="arXiv:2411.00827" /> <link title="schema(PRISM)" rel="schema.prism" href="http://prismstandard.org/namespaces/1.2/basic/" /> <meta name="prism.publicationDate" content="10/2024" /> <meta name="prism.publicationName" content="arXiv" /> <meta name="prism.startingPage" content="arXiv:2411.00827" /> <link title="schema(DC)" rel="schema.dc" href="http://purl.org/dc/elements/1.1/" /> <meta name="dc.identifier" content="doi:10.48550/arXiv.2411.00827" /> <meta name="dc.date" content="10/2024" /> <meta name="dc.source" content="arXiv" /> <meta name="dc.title" content="IDEATOR: Jailbreaking and Benchmarking Large Vision-Language Models Using Themselves" /> <meta name="dc.creator" content="Wang, Ruofan"> <meta name="dc.creator" content="Li, Juncheng"> <meta name="dc.creator" content="Wang, Yixu"> <meta name="dc.creator" content="Wang, Bo"> <meta name="dc.creator" content="Wang, Xiaosen"> <meta name="dc.creator" content="Teng, Yan"> <meta name="dc.creator" content="Wang, Yingchun"> <meta name="dc.creator" content="Ma, Xingjun"> <meta name="dc.creator" content="Jiang, Yu-Gang">  <meta name="twitter:card" content="summary_large_image"/> <meta name="twitter:description" content="As large Vision-Language Models (VLMs) gain prominence, ensuring their safe deployment has become critical. Recent studies have explored VLM robustness against jailbreak attacks-techniques that exploit model vulnerabilities to elicit harmful outputs. However, the limited availability of diverse multimodal data has constrained current approaches to rely heavily on adversarial or manually crafted images derived from harmful text datasets, which often lack effectiveness and diversity across different contexts. In this paper, we propose IDEATOR, a novel jailbreak method that autonomously generates malicious image-text pairs for black-box jailbreak attacks. IDEATOR is grounded in the insight that VLMs themselves could serve as powerful red team models for generating multimodal jailbreak prompts. Specifically, IDEATOR leverages a VLM to create targeted jailbreak texts and pairs them with jailbreak images generated by a state-of-the-art diffusion model. Extensive experiments demonstrate IDEATOR's high effectiveness and transferability, achieving a 94% attack success rate (ASR) in jailbreaking MiniGPT-4 with an average of only 5.34 queries, and high ASRs of 82%, 88%, and 75% when transferred to LLaVA, InstructBLIP, and Chameleon, respectively. Building on IDEATOR's strong transferability and automated process, we introduce the VLBreakBench, a safety benchmark comprising 3,654 multimodal jailbreak samples. Our benchmark results on 11 recently released VLMs reveal significant gaps in safety alignment. For instance, our challenge set achieves ASRs of 46.31% on GPT-4o and 19.65% on Claude-3.5-Sonnet, underscoring the urgent need for stronger defenses."/> <meta name="twitter:title" content="IDEATOR: Jailbreaking and Benchmarking Large Vision-Language Models Using Themselves"/> <meta name="twitter:site" content="@adsabs"/> <meta name="twitter:domain" content="ADS"/> <meta name="twitter:image:src" content="https://ui.adsabs.harvard.edu/styles/img/transparent_logo.svg"/> <meta name="twitter:creator" content="@adsabs"/> <meta charset="utf-8"> <meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no"> <base href="/"> <style> .btn-full-ads { color: #fff !important; background-color: #1a1a1a !important; border-color: #1a1a1a !important; margin-top: 9px !important; padding-bottom: 10px !important; padding-top: 10px !important; } .btn-full-ads:hover, .btn-full-ads:focus, .btn-full-ads:active, .btn-full-ads.active, .open>.dropdown-toggle.btn-full-ads { color: #000 !important; background-color: #ddd !important; border-color: #1a1a1a !important; } .dropdown-toggle:hover .dropdown-menu { display: block; } .navbar-nav.navbar-right:last-child { margin-right: -15px !important; } .navbar-right { @media screen (min-width: $screen-sm) { float: right!important; } } /*the container must be positioned relative:*/ .autocomplete { position: relative; display: inline-block; } .autocomplete-items { position: absolute; border: 1px solid #d4d4d4; border-bottom: none; border-top: none; z-index: 99; /*position the autocomplete items to be the same width as the container:*/ top: 100%; left: 0; right: 0; } .autocomplete-items div { padding: 10px; cursor: pointer; background-color: #fff; border-bottom: 1px solid #d4d4d4; } /*when hovering an item:*/ .autocomplete-items div:hover { background-color: #e9e9e9; } /*when navigating through the items using the arrow keys:*/ .autocomplete-active { background-color: #d7dfec !important; color: #000000; } </style> </head> <body>  <noscript><iframe src="https://www.googletagmanager.com/ns.html?id=GTM-NT2453N" height="0" width="0" style="display:none;visibility:hidden"></iframe></noscript>  <div id="aria-announcement-container">Now on home page</div> <div id="app-container"> <div id="body-template-container"> <div class="s-master-page-manager"> <div id="navbar-container"> <div data-widget="NavbarWidget"> <nav class="navbar navbar-inverse"> <div class="container-fluid">  <div class=""> <ul class="nav navbar-nav navbar-left"> <li> <a class="navbar-brand" href="/"> <img class="s-ads-icon" src="/styles/img/transparent_logo.svg" alt="ads icon"/> <h1> <b>ads</b></h1> </a> </li> </ul> </div>  <div class="">  <ul class="nav navbar-nav navbar-right"> <li data-match-route="/"> <a href="/core/never/abs/2024arXiv241100827W/abstract" style="transition: none; " class="btn btn-full-ads"> <i class="fa fa-refresh"></i> Enable full ADS </a> </li> </ul> </div> </div> </nav> </div> </div> <div id="content-container"> <div class="dynamic-container s-dynamic-container"> <div id="abstract-page-layout" class="s-abstract-page-layout"> <div class="row s-stable-search-bar-height s-results-control-row-container hidden-xs"> </div> <div class="row s-dynamic-page-body" id="dynamic-page-body"> <div class="s-abstract-content"> <div class="col-xs-12 col-sm-3 col-md-2" style="" id="left-column"> <div class="nav-container s-nav-container" style="transform: none; width: 100%; position: relative" id="left-column"> <nav> <div class="s-nav-header s-view-nav"> <i class="icon-list"></i> <h3>view </h3> </div> <a href="/abs/2024arXiv241100827W/abstract" data-widget-id="ShowAbstract"> <div class="abstract-nav s-nav s-nav-selected"> <span class="s-content"> Abstract </span> </div> </a> </a> <a href="/abs/2024arXiv241100827W/citations" aria-disabled="true" data-widget-id="ShowCitations"> <div class="abstract-nav s-nav "> <span class="s-content"> Citations <span class="num-items">(4)</span> </span> </div> </a> <a href="/abs/2024arXiv241100827W/references" aria-disabled="true" data-widget-id="ShowReferences"> <div class="abstract-nav s-nav "> <span class="s-content"> References <span class="num-items">(40)</span> </span> </div> </a> <a href="/abs/2024arXiv241100827W/coreads" aria-disabled="true" aria-disabled="true" data-widget-id="ShowCoreads"> <div class="abstract-nav s-nav "> <span class="s-content"> Co-Reads </span> </div> </a> <a href="/abs/2024arXiv241100827W/similar" aria-disabled="true" data-widget-id="ShowSimilar"> <div class="abstract-nav s-nav "> <span class="s-content"> Similar Papers </span> </div> </a> <div aria-disabled="true" data-widget-id="ShowToc"> <div class="abstract-nav s-nav s-nav-inactive"> <span class="s-content"> Volume Content </span> </div> </div> <div href="#" data-widget-id="ShowGraphics"> <div class="abstract-nav s-nav s-nav-inactive"> <span class="s-content"> Graphics </span> </div> </div> <a href="/abs/2024arXiv241100827W/metrics" data-widget-id="ShowMetrics"> <div class="abstract-nav s-nav"> <span class="s-content"> Metrics </span> </div> <a href="/abs/2024arXiv241100827W/exportcitation" data-widget-id="ShowExportcitation__default"> <div class="abstract-nav s-nav "> <span class="content"> Export Citation </span> </div> </a> </nav> </div> </div> <div class="col-xs-12 col-sm-8 col-md-7 col-lg-7 s-middle-column" id="middle-column" style="padding-bottom: 0%">  <div class="main-content-container s-main-content-container" id="main-content" tabindex="-1" style="margin-bottom: 5px"> <div class="print-visible"> <h2 style="margin-left:6.1%;">ADS</h2> </div> <div id="abstract-title-container" class="s-abstract-title-container"> <div data-widget="ShowAbstract"> <article class="s-abstract-metadata">  <h2 class="s-abstract-title"> IDEATOR: Jailbreaking and Benchmarking Large Vision-Language Models Using Themselves <a href=""></a> </h2> <div id="authors-and-aff" class="s-authors-and-aff"> <ul class="list-inline"> <li class="author"><a href="/search/?q=author%3A%22Wang%2C+Ruofan%22">Wang, Ruofan</a> </li>; <li class="author"><a href="/search/?q=author%3A%22Li%2C+Juncheng%22">Li, Juncheng</a> </li>; <li class="author"><a href="/search/?q=author%3A%22Wang%2C+Yixu%22">Wang, Yixu</a> </li>; <li class="author"><a href="/search/?q=author%3A%22Wang%2C+Bo%22">Wang, Bo</a> </li>; <li class="author"><a href="/search/?q=author%3A%22Wang%2C+Xiaosen%22">Wang, Xiaosen</a> </li>; <li class="author"><a href="/search/?q=author%3A%22Teng%2C+Yan%22">Teng, Yan</a> </li>; <li class="author"><a href="/search/?q=author%3A%22Wang%2C+Yingchun%22">Wang, Yingchun</a> </li>; <li class="author"><a href="/search/?q=author%3A%22Ma%2C+Xingjun%22">Ma, Xingjun</a> </li>; <li class="author"><a href="/search/?q=author%3A%22Jiang%2C+Yu-Gang%22">Jiang, Yu-Gang</a> </li> </ul> </div> <div class="s-abstract-text"> <h4 class="sr-only">Abstract</h4> <p> As large Vision-Language Models (VLMs) gain prominence, ensuring their safe deployment has become critical. Recent studies have explored VLM robustness against jailbreak attacks-techniques that exploit model vulnerabilities to elicit harmful outputs. However, the limited availability of diverse multimodal data has constrained current approaches to rely heavily on adversarial or manually crafted images derived from harmful text datasets, which often lack effectiveness and diversity across different contexts. In this paper, we propose IDEATOR, a novel jailbreak method that autonomously generates malicious image-text pairs for black-box jailbreak attacks. IDEATOR is grounded in the insight that VLMs themselves could serve as powerful red team models for generating multimodal jailbreak prompts. Specifically, IDEATOR leverages a VLM to create targeted jailbreak texts and pairs them with jailbreak images generated by a state-of-the-art diffusion model. Extensive experiments demonstrate IDEATOR's high effectiveness and transferability, achieving a 94% attack success rate (ASR) in jailbreaking MiniGPT-4 with an average of only 5.34 queries, and high ASRs of 82%, 88%, and 75% when transferred to LLaVA, InstructBLIP, and Chameleon, respectively. Building on IDEATOR's strong transferability and automated process, we introduce the VLBreakBench, a safety benchmark comprising 3,654 multimodal jailbreak samples. Our benchmark results on 11 recently released VLMs reveal significant gaps in safety alignment. For instance, our challenge set achieves ASRs of 46.31% on GPT-4o and 19.65% on Claude-3.5-Sonnet, underscoring the urgent need for stronger defenses. </p> </div> <br> <dl class="s-abstract-dl-horizontal"> <dt>Publication:</dt> <dd> <div id="article-publication">arXiv e-prints</div> </dd> <dt>Pub Date:</dt> <dd>October 2024</dd> <dt>DOI:</dt> <dd> <p class="doi-p"> <a href="/link_gateway/2024arXiv241100827W/doi:10.48550/arXiv.2411.00827" target="_blank" rel="noreferrer noopener">10.48550/arXiv.2411.00827</a> <i class="fa fa-external-link"></i> </p> </dd> <dt>arXiv:</dt> <dd> <span> <a href="/link_gateway/2024arXiv241100827W/arXiv:2411.00827" target="_blank" rel="noreferrer noopener">arXiv:2411.00827</a> <i class="fa fa-external-link"></i> </span> </dd> <dt>Bibcode:</dt> <dd> <a href="/abs/2024arXiv241100827W/abstract"> 2024arXiv241100827W </a> <i class="icon-help" title="The bibcode is assigned by the ADS as a unique identifier for the paper."></i> </dd> <dt>Keywords:</dt> <dd> <ul class="list-inline"> <li>Computer Science - Computer Vision and Pattern Recognition;</li> <li>Computer Science - Artificial Intelligence</li> </ul> </dd> </dl> </article> </div> <div data-widget="ShowCitations"></div> <div data-widget="ShowReferences"></div> <div data-widget="ShowCoreads"></div> <div data-widget="ShowSimilar"></div> <div data-widget="ShowTableofcontents"></div> <div data-widget="ShowGraphics"></div> <div data-widget="ShowExportcitation" data-origin="abstract"></div> <div data-widget="ShowMetrics" data-allow-redirect="false"></div> <div data-widget="MetaTagsWidget"></div> </div> </div> </div> <div class="s-right-col-container col-xs-12 col-sm-12 col-md-3 col-lg-2 s-right-column" id="right-col-container" > <div data-widget="ShowResources"> <div data-reactroot="" class="s-right-col-widget-container" style="padding: 10px" > <div> <div class="resources__container"> <div class="resources__full__list"> <div class="resources__header__row"> <i class="fa fa-file-text-o" aria-hidden="true"> </i> <div class="resources__header__title">full text sources</div> </div> <div class="resources__content"> <div class="resources__content__title">Preprint</div> <div class="resources__content__links"> <span> <a href="/link_gateway/2024arXiv241100827W/EPRINT_PDF" rel="noopener" class="resources__content__link unlock" > <i class="fa fa-file-pdf-o" aria-hidden="true"> </i> </a> <div class="resources__content__link__separator">|</div> </span> <span> <a href="/link_gateway/2024arXiv241100827W/EPRINT_HTML" rel="noopener" class="resources__content__link unlock" > <i class="fa fa-file-text" aria-hidden="true"> </i> </a> </span> </div> </div> </div> </div> <div data-widget="ShowAssociated"> </div> </div> </div> </div> <div data-widget="ShowGraphicsSidebar"> </div> </div> </div> </div> </div> </div> </div> <div id="footer-container"> <div data-widget="FooterWidget"> <div class="footer s-footer"> <footer> <div class="__footer_wrapper"> <div class="__footer_brand"> 漏 The SAO Astrophysics Data System <div class="__footer_brand_extra"> <p> <i class="fa fa-envelope"></i> adshelp[at]cfa.harvard.edu </p> <p> The ADS is operated by the Smithsonian Astrophysical Observatory under NASA Cooperative Agreement <em>80NSSC21M0056</em> </p> </div> <div class="__footer_brand_logos"> <div class="logo1"> <a href="http://www.si.edu" target="_blank" rel="noreferrer noopener"> <img id="smithsonian-logo" src="/styles/img/smithsonian-logo.svg" alt="Smithsonian logo" /> </a> </div> <div class="logo2"> <a href="https://www.cfa.harvard.edu/" target="_blank" rel="noreferrer noopener"> <img src="/styles/img/cfa.png" alt="Harvard Center for Astrophysics logo" id="cfa-logo" /> </a> </div> <div class="logo3"> <a href="http://www.nasa.gov" target="_blank" rel="noreferrer noopener"> <img src="/styles/img/nasa-partner.svg" alt="NASA logo" id="nasa-logo" /> </a> </div> </div> <div class="__footer_brand_disclaimer"> *The material contained in this document is based upon work supported by a National Aeronautics and Space Administration (NASA) grant or cooperative agreement. Any opinions, findings, conclusions or recommendations expressed in this material are those of the author and do not necessarily reflect the views of NASA. </div> </div> <div class="__footer_list"> <div class="__footer_list_title"> Resources </div> <ul class="__footer_links"> <li> <a href="/about/" target="_blank" rel="noopener"> <i class="fa fa-question-circle"></i> About ADS </a> </li> <li> <a href="//ui.adsabs.harvard.edu/help/" target="_blank" rel="noopener"> <i class="fa fa-info-circle"></i> ADS Help </a> </li> <li> <a href="//ui.adsabs.harvard.edu/help/whats_new/" target="_blank" rel="noopener"> <i class="fa fa-bullhorn"></i> What's New </a> </li> <li> <a href="/about/careers/" target="_blank" rel="noopener"> <i class="fa fa-group"></i> Careers@ADS </a> </li> <li> <a href="//ui.adsabs.harvard.edu/help/accessibility/" target="_blank" rel="noreferrer noopener"> <i class="fa fa-universal-access" aria-hidden="true"></i> Web Accessibility Policy </a> </li> </ul> </div> <div class="__footer_list"> <div class="__footer_list_title"> Social </div> <ul class="__footer_links"> <li> <a href="//twitter.com/adsabs" target="_blank" rel="noopener"> <i class="fa fa-twitter"></i> @adsabs </a> </li> <li> <a href="//ui.adsabs.harvard.edu/blog/" target="_blank" rel="noopener"> <i class="fa fa-newspaper-o"></i> ADS Blog </a> </li> </ul> </div> <div class="__footer_list"> <div class="__footer_list_title"> Project </div> <ul class="__footer_links"> <li> <a href="/core/never">Switch to full ADS</a> </li> <li> <a href="https://adsisdownorjustme.herokuapp.com/" target="_blank" rel="noopener">Is ADS down? (or is it just me...)</a> </li> <li> <a href="http://www.si.edu" target="_blank" rel="noopener">Smithsonian Institution</a> </li> <li> <a href="http://www.si.edu/Privacy" target="_blank" rel="noopener">Smithsonian Privacy Notice</a> </li> <li> <a href="http://www.si.edu/Termsofuse" target="_blank" rel="noopener">Smithsonian Terms of Use</a> </li> <li> <a href="http://www.cfa.harvard.edu/sao" target="_blank" rel="noopener">Smithsonian Astrophysical Observatory</a> </li> <li> <a href="http://www.nasa.gov" target="_blank" rel="noopener">NASA</a> </li> </ul> </div> </div> </footer> </div> </div> </div> </div> </div> </div> <div id="darkSwitch" class="darkmode-toggle hidden" title="Turn on dark mode">馃寭</div> <script> function autocomplete(searchBox, autoValues) { // Arguments: the text field element and an array of possible autocompleted values var currentFocus; // selected autocomplete option // Function to be run when the user types searchBox.addEventListener("input", function(e) { var a, b, i, val = this.value; // close any list of autocomplete values closeAllLists(); if (!val) { return false;} val = val.split(/\s+/); val = val[val.length - 1]; if (!val) { return false;} currentFocus = -1; // Create a DIV element that will contain the items (values): a = document.createElement("DIV"); a.setAttribute("id", this.id + "autocomplete-list"); a.setAttribute("class", "autocomplete-items"); // Append the DIV element as a child of the autocomplete container: this.parentNode.appendChild(a); for (i = 0; i < autoValues.length; i++) { // Check if the item starts with the same letters as the text field value: if (autoValues[i].match.substr(0, val.length).toUpperCase() == val.toUpperCase()) { // Create a DIV element for each matching element: b = document.createElement("DIV"); b.innerHTML = autoValues[i].label; if ("desc" in autoValues[i]) { b.innerHTML += " <i>" + autoValues[i].desc + "</i>"; } if (autoValues[i].value.startsWith(autoValues[i].match) ) { b.innerHTML += " | <strong>" + autoValues[i].match.substr(0, val.length) + "</strong>"; b.innerHTML += autoValues[i].match.substr(val.length); } // Insert a input field that will hold the current array item's value: b.innerHTML += "<input type='hidden' value='" + autoValues[i].value + "'>"; // Listen to clicks on the item value (DIV element): b.addEventListener("click", function(e) { var terms = searchBox.value.split(/\s+/); // Remove the current part of the input used for matching terms.pop(); // Insert the value for the autocomplete text field: terms.push(this.getElementsByTagName("input")[0].value); searchBox.value = terms.join(" "); // Move cursor position inside quotes/parenthesis if needed searchBox.focus(); if (searchBox.value[searchBox.value.length-1] === '"' || searchBox.value[searchBox.value.length-1] === ')') { searchBox.setSelectionRange(searchBox.value.length-1, searchBox.value.length-1); } // Close the list of autocompleted values closeAllLists(); }); a.appendChild(b); } } if (a.children.length > 0) { // By default, enter will select the first entry currentFocus = 0; addActive(a.children); } }); /*execute a function presses a key on the keyboard:*/ searchBox.addEventListener("keydown", function(e) { var x = document.getElementById(this.id + "autocomplete-list"); if (x) x = x.getElementsByTagName("div"); if (e.keyCode == 40) { // If the arrow DOWN key is pressed, increase the currentFocus variable: currentFocus++; addActive(x); } else if (e.keyCode == 38) { //up // If the arrow UP key is pressed, decrease the currentFocus variable: currentFocus--; /*and and make the current item more visible:*/ addActive(x); } else if (e.keyCode == 13) { // If the ENTER key is pressed: if (currentFocus > -1) { // Prevent the form from being submitted: e.preventDefault(); // Simulate a click on the "active" item: if (x) x[currentFocus].click(); currentFocus = -1; } } }); function addActive(x) { // Classify an item as "active": if (!x) return false; // Remove the "active" class on all items: removeActive(x); if (currentFocus >= x.length) currentFocus = 0; if (currentFocus < 0) currentFocus = (x.length - 1); // Add class "autocomplete-active": x[currentFocus].classList.add("autocomplete-active"); } function removeActive(x) { // Remove the "active" class from all autocomplete items: for (var i = 0; i < x.length; i++) { x[i].classList.remove("autocomplete-active"); } } function closeAllLists(elmnt) { // Close all autocomplete lists in the document, except the one passed as an argument: var x = document.getElementsByClassName("autocomplete-items"); for (var i = 0; i < x.length; i++) { if (elmnt != x[i] && elmnt != searchBox) { x[i].parentNode.removeChild(x[i]); } } } // Any other clicks in the document: document.addEventListener("click", function (e) { closeAllLists(e.target); }); } var autoList = [ { value: 'author:""', label: 'Author', match: 'author:"' }, { value: 'author:"^"', label: 'First Author', match: 'first author' }, { value: 'author:"^"', label: 'First Author', match: 'author:"^' }, { value: 'bibcode:""', label: 'Bibcode', desc: 'e.g. bibcode:1989ApJ...342L..71R', match: 'bibcode:"' }, { value: 'bibstem:""', label: 'Publication', desc: 'e.g. bibstem:ApJ', match: 'bibstem:"' }, { value: 'bibstem:""', label: 'Publication', desc: 'e.g. bibstem:ApJ', match: 'publication (bibstem)' }, { value: 'arXiv:', label: 'arXiv ID', match: 'arxiv:' }, { value: 'doi:', label: 'DOI', match: 'doi:' }, { value: 'full:""', label: 'Full text search', desc: 'title, abstract, and body', match: 'full:' }, { value: 'full:""', label: 'Full text search', desc: 'title, abstract, and body', match: 'fulltext' }, { value: 'full:""', label: 'Full text search', desc: 'title, abstract, and body', match: 'text' }, { value: 'year:', label: 'Year', match: 'year' }, { value: 'year:1999-2005', label: 'Year Range', desc: 'e.g. 1999-2005', match: 'year range' }, { value: 'aff:""', label: 'Affiliation', match: 'aff:' }, { value: 'abs:""', label: 'Search abstract + title + keywords', match: 'abs:' }, { value: 'database:astronomy', label: 'Limit to papers in the astronomy database', match: 'database:astronomy' }, { value: 'database:physics', label: 'Limit to papers in the physics database', match: 'database:physics' }, { value: 'title:""', label: 'Title', match: 'title:"' }, { value: 'orcid:', label: 'ORCiD identifier', match: 'orcid:' }, { value: 'object:', label: 'SIMBAD object (e.g. object:LMC)', match: 'object:' }, { value: 'property:refereed', label: 'Limit to refereed', desc: '(property:refereed)', match: 'refereed' }, { value: 'property:refereed', label: 'Limit to refereed', desc: '(property:refereed)', match: 'property:refereed' }, { value: 'property:notrefereed', label: 'Limit to non-refereed', desc: '(property:notrefereed)', match: 'property:notrefereed' }, { value: 'property:notrefereed', label: 'Limit to non-refereed', desc: '(property:notrefereed)', match: 'notrefereed' }, { value: 'property:eprint', label: 'Limit to eprints', desc: '(property:eprint)', match: 'eprint' }, { value: 'property:eprint', label: 'Limit to eprints', desc: '(property:eprint)', match: 'property:eprint' }, { value: 'property:openaccess', label: 'Limit to open access', desc: '(property:openaccess)', match: 'property:openaccess' }, { value: 'property:openaccess', label: 'Limit to open access', desc: '(property:openaccess)', match: 'openaccess' }, { value: 'doctype:software', label: 'Limit to software', desc: '(doctype:software)', match: 'software' }, { value: 'doctype:software', label: 'Limit to software', desc: '(doctype:software)', match: 'doctype:software' }, { value: 'property:inproceedings', label: 'Limit to papers in conference proceedings', desc: '(property:inproceedings)', match: 'proceedings' }, { value: 'property:inproceedings', label: 'Limit to papers in conference proceedings', desc: '(property:inproceedings)', match: 'property:inproceedings' }, { value: 'citations()', label: 'Citations', desc: 'Get papers citing your search result set', match: 'citations(' }, { value: 'references()', label: 'References', desc: 'Get papers referenced by your search result set', match: 'references(' }, { value: 'trending()', label: 'Trending', desc: 'Get papers most read by users who recently read your search result set', match: 'trending(' }, { value: 'reviews()', label: 'Review Articles', desc: 'Get most relevant papers that cite your search result set', match: 'reviews(' }, { value: 'useful()', label: 'Useful', desc: 'Get papers most frequently cited by your search result set', match: 'useful(' }, { value: 'similar()', label: 'Similar', desc: 'Get papers that have similar full text to your search result set', match: 'similar(' }, ]; // initiate the autocomplete function on the "q" element, and pass along the operators array as possible autocomplete values: inputBox = document.getElementById("q") if (inputBox) { inputBox.focus() // autofucs inputBox.setSelectionRange(inputBox.value.length, inputBox.value.length); // bring cursor to the end autocomplete(inputBox, autoList); } </script> <script> (function() { // turn off no-js if we have javascript document.documentElement.className = document.documentElement.className.replace("no-js", "js"); function getCookie(cname) { var name = cname + "="; var decodedCookie = decodeURIComponent(document.cookie); var ca = decodedCookie.split(';'); for (var i = 0; i < ca.length; i++) { var c = ca[i]; while (c.charAt(0) == ' ') { c = c.substring(1); } if (c.indexOf(name) == 0) { return c.substring(name.length, c.length); } } return ""; } (function() { // looks for the cookie, and sets true if its 'always' const coreCookie = getCookie('core') === 'always'; // only load bumblebee if we detect the core cookie and we are on abstract page if (coreCookie || (!(/^\/abs\//.test(document.location.pathname)) && !coreCookie)) { return; } window.__PRERENDERED = true; const addScript = function(args, cb) { const script = document.createElement('script'); Object.keys(args).forEach((key) => { script.setAttribute(key, args[key]); }); script.onload = function() { cb && cb(script); }; document.body.appendChild(script); } window.require = { waitSeconds: 0, baseUrl: '/' }; addScript({ src: '/libs/require.js' }, () => { addScript({ src: '/config/init.js' }); }); })(); })(); </script>  <script src="https://js.sentry-cdn.com/46062cbe0aeb7a3b2bb4c3a9b8cd1ac7.min.js" crossorigin="anonymous"></script> <script> // Set the environment variable window.process = { env: { NODE_ENV: window.ENV ?? 'production' } }; window.getSentry = (cb, errCb) => { if (typeof errCb === 'function') { errCb.call(null, new Error('Failed to get Sentry')); } }; window.sentryOnLoad = function() { try { Sentry.init({ dsn: 'https://46062cbe0aeb7a3b2bb4c3a9b8cd1ac7@o1060269.ingest.us.sentry.io/4507341192036352', // Performance Monitoring tracesSampleRate: window.ENV === 'development' ? 1.0 : 0.75, // Session Replay replaysSessionSampleRate: window.ENV === 'development' ? 1.0 : 0.1, replaysOnErrorSampleRate: 1.0, environment: window.ENV ?? 'production', integrations: [ Sentry.browserTracingIntegration(), Sentry.replayIntegration({ maskAllText: false, blockAllMedia: false, unmask: ['[name=q]'], }), ], tracePropagationTargets: [ 'localhost', 'https://ui.adsabs.harvard.edu', 'https://devui.adsabs.harvard.edu', 'https://qa.adsabs.harvard.edu', ], }); } catch (e) { console.error('Could not initialize Sentry', e); } }; window.getSentry = function(cb, errCb = () => {}) { try { if (!window.Sentry) { throw new Error('Sentry not found'); } return cb.call(window, Sentry); } catch (e) { console.warn('unable to run sentry callback', e); if (typeof errCb === 'function') { errCb.call(window, e); } } }; </script>  </body> </html>

CINXE.COM

IDEATOR: Jailbreaking and Benchmarking Large Vision-Language Models Using Themselves - ADS