PageSourceSearch

https://agentbait.github.io/

html agentbait.github.io collected 2026-10-03 08:33:18 UTC 62,384 bytes, 25 lines download raw bytes

1<!DOCTYPE html><html lang="en"><head><meta charSet="utf-8"/><meta name="viewport" content="width=device-width, initial-scale=1"/><link rel="preload" as="image" href="/agentbait-mark.png"/><link rel="preload" as="image" href="/github-mark.svg"/><link rel="preload" as="image" href="/arxiv-mark.svg"/><link rel="preload" as="image" href="/huggingface-mark.svg"/><link rel="preload" as="image" href="/bair-logo.png"/><link rel="preload" as="image" href="/sky-logo.png"/><link rel="preload" as="image" href="/editor-hand.png" fetchPriority="high"/><link rel="preload" as="image" href="/agentbait-method.png"/><link rel="preload" as="image" href="/advisor-scholar.png"/><link rel="preload" as="image" href="/rewriter-hand-strings.png"/><link rel="stylesheet" href="/_next/static/chunks/03-1br6~qsd6r.css" data-precedence="next"/><link rel="preload" as="script" fetchPriority="low" href="/_next/static/chunks/0g3yt-.a0ubwc.js"/>
1<script src="/_next/static/chunks/09gjkq5y3b.u9.js" async=""></script>
1<script src="/_next/static/chunks/0m_p1bxtorv5i.js" async=""></script>
1<script src="/_next/static/chunks/turbopack-0l866-lpqc4u0.js" async=""></script>
1<script src="/_next/static/chunks/0m4_.n.zi_~-n.js" async=""></script>
1<script src="/_next/static/chunks/0dbhjjzl8qfwv.js" async=""></script>
1<script src="/_next/static/chunks/0-oqd4xnd95hv.js" async=""></script>
1<link rel="preload" as="image" href="/selector-hand.png"/><title>You Won&#x27;t Believe This Click: Content Rewriting for Agentic Choice</title><meta name="description" content="AgentBait studies how rewriting the presentation of one content item can shift language-model-mediated selection, and examines the trade-off between selection and factual support."/><meta name="robots" content="index, follow, max-image-preview:large"/><link rel="canonical" href="https://agentbait.github.io/"/><meta name="google-site-verification" content="5XSefLo9dX0I_Szro49mP5w54fCMDDuxB3E3LViw5mU"/><meta property="og:title" content="You Won&#x27;t Believe This Click: Content Rewriting for Agentic Choice"/><meta property="og:description" content="AgentBait studies how rewriting the presentation of one content item can shift language-model-mediated selection, and examines the trade-off between selection and factual support."/><meta property="og:image" content="https://agentbait.github.io/og.png"/><meta property="og:image:width" content="1200"/><meta property="og:image:height" content="630"/><meta property="og:image:alt" content="AgentBait paper preview showing a fixed three-item candidate slate with only target B rewritten and selected."/><meta property="og:type" content="website"/><meta name="twitter:card" content="summary_large_image"/><meta name="twitter:title" content="You Won&#x27;t Believe This Click: Content Rewriting for Agentic Choice"/><meta name="twitter:description" content="AgentBait studies how rewriting the presentation of one content item can shift language-model-mediated selection, and examines the trade-off between selection and factual support."/><meta name="twitter:image" content="https://agentbait.github.io/og.png"/><meta name="twitter:image:alt" content="AgentBait paper preview showing a fixed three-item candidate slate with only target B rewritten and selected."/><meta name="twitter:image:width" content="1200"/><meta name="twitter:image:height" content="630"/><link rel="shortcut icon" href="favicon.png"/><link rel="icon" href="favicon.png" type="image/png" sizes="64x64"/>
1<script id="agentbait-google-consent-default">
2  window.dataLayer = window.dataLayer || [];
3  window.gtag = window.gtag || function(){window.dataLayer.push(arguments);};
4  window.gtag('consent', 'default', {
5    analytics_storage: 'granted',
6    ad_storage: 'granted',
7    ad_user_data: 'granted',
8    ad_personalization: 'granted'
9  });
10  window.gtag('consent', 'default', {
11    analytics_storage: 'granted',
12    ad_storage: 'denied',
13    ad_user_data: 'denied',
14    ad_personalization: 'denied',
15    region: ["AT","BE","BG","HR","CY","CZ","DK","EE","FI","FR","DE","GR","HU","IE","IT","LV","LT","LU","MT","NL","PL","PT","RO","SK","SI","ES","SE","IS","LI","NO","GB","CH"]
16  });
17</script>
17<script id="google-tag-manager">
vendor: 327 bytes, lines 17-21
17(function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':
18new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],
19j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src=
20'https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);
21})(window,document,'script','dataLayer','
21GTM-WSHC2PFG
vendor: 3 bytes, line 21
21');
21</script>
21<script src="/_next/static/chunks/03~yq9q893hmn.js" noModule=""></script>
21</head><body><div hidden=""><!--$--><!--/$--></div><noscript><iframe src="https://www.googletagmanager.com/ns.html?id=GTM-WSHC2PFG" height="0" width="0" style="display:none;visibility:hidden" title="Google Tag Manager"></iframe></noscript><main id="paper"><header class="site-header"><a class="wordmark" href="#paper" aria-label="AgentBait paper home"><span class="wordmark-logo" aria-hidden="true"><img class="wordmark-mark" src="/agentbait-mark.png" width="18" height="32" alt=""/></span><span class="wordmark-copy"><span class="wordmark-title">AgentBait</span><span class="wordmark-tagline">You Won’t Believe This Click</span></span></a><nav aria-label="Reading navigation"><a href="https://github.com/Agentbait/agentbait" target="_blank" rel="noreferrer"><img class="resource-mark github-mark" src="/github-mark.svg" width="24" height="24" alt="" aria-hidden="true"/><span>Code</span></a><a href="/agentbait-paper.pdf" target="_blank" rel="noreferrer"><img class="resource-mark arxiv-mark" src="/arxiv-mark.svg" width="24" height="24" alt="" aria-hidden="true"/><span>Paper</span></a><span class="resource-placeholder" aria-disabled="true" aria-label="Hugging Face resource placeholder"><img class="resource-mark huggingface-mark" src="/huggingface-mark.svg" width="24" height="24" alt="" aria-hidden="true"/><span>Hugging Face</span></span></nav></header><article><section class="hero-feature" aria-labelledby="paper-title"><header class="article-header"><h1 id="paper-title">You Won&#x27;t Believe This Click</h1><p class="subtitle">Content Rewriting for Agentic Choice</p><div class="paper-identity"><div class="byline"><div class="byline-copy"><p><a class="author-link" href="https://www.linkedin.com/in/chris-jin-680537299" target="_blank" rel="noreferrer"><strong>Tianyi Jin</strong></a>, <a class="author-link" href="https://zwcolin.github.io/" target="_blank" rel="noreferrer"><strong>Zirui Wang</strong></a> and <a class="author-link" href="https://dchan.cc/" target="_blank" rel="noreferrer"><strong>David M. Chan</strong></a></p><p>University of California, Berkeley</p></div><div class="affiliation-logos" aria-label="Research affiliations"><a href="https://bair.berkeley.edu/" target="_blank" rel="noreferrer" aria-label="Visit Berkeley Artificial Intelligence Research"><img src="/bair-logo.png" width="205" height="146" alt="Berkeley Artificial Intelligence Research"/></a><a href="https://sky.cs.berkeley.edu/" target="_blank" rel="noreferrer" aria-label="Visit Sky Computing Lab"><img src="/sky-logo.png" width="264" height="243" alt="Sky Computing Lab"/></a></div></div><div class="paper-links" aria-label="Paper resources"><a href="https://github.com/Agentbait/agentbait" target="_blank" rel="noreferrer"><img class="resource-mark github-mark" src="/github-mark.svg" width="24" height="24" alt="" aria-hidden="true"/><span>Code ↗</span></a><a href="/agentbait-paper.pdf" target="_blank" rel="noreferrer"><img class="resource-mark arxiv-mark" src="/arxiv-mark.svg" width="24" height="24" alt="" aria-hidden="true"/><span>Paper ↗</span></a><span class="resource-placeholder" aria-disabled="true" aria-label="Hugging Face resource placeholder"><img class="resource-mark huggingface-mark" src="/huggingface-mark.svg" width="24" height="24" alt="" aria-hidden="true"/><span>Hugging Face</span></span></div></div></header><section class="attack-demo stage-candidate-set" id="demo" aria-label="AgentBait fixed-set chooser replay and training loop"><div class="hero-flip-card "><div class="hero-flip-inner" role="button" tabindex="0" aria-label="Show paper graph"><div class="hero-flip-face hero-flip-front" aria-hidden="false"><div class="storyboard-board    full-edit" role="group" aria-label="Automatically animated AgentBait fixed-slate comparison using the MIND example from Figure 1"><section class="candidate-set" aria-labelledby="hero-candidate-set-title"><header><div><span id="hero-candidate-set-title">Candidate Set</span></div><small>Original presentation</small></header><ol><li class="candidate-card is-target  "><span class="paper-rank">A<!-- -->.</span><span class="paper-copy"><b>Marshawn playing in charity soccer game went exactly as you&#x27;d expect.</b><small class="paper-abstract"><span class="paper-abstract-label">Abstract</span><span class="paper-abstract-copy">If there was ever a sport-athlete combination that we&#x27;d never expect to work out, it&#x27;
21d be Marshawn Lynch dabbling in soccer.</span></small></span></li><li class="candidate-card   "><span class="paper-rank">B<!-- -->.</span><span class="paper-copy"><b>Sofia Vergara and Joe Manganiello Celebrate 4-Year Wedding Anniversary: &#x27;Mi Amor!&#x27;</b><small class="paper-abstract"><span class="paper-abstract-label">Abstract</span><span class="paper-abstract-copy">Sofia Vergara and Joe Manganiello Celebrate 4-Year Wedding Anniversary</span></small></span></li><li class="candidate-card   "><span class="paper-rank">C<!-- -->.</span><span class="paper-copy"><b>The Coolest And Craziest McDonald&#x27;s Across The Country</b><small class="paper-abstract"><span class="paper-abstract-label">Abstract</span><span class="paper-abstract-copy">Sometimes, the Golden Arches know how to pull out all the stops.</span></small></span></li></ol></section><section class="rewrite-focus" aria-label="Target title rewriting"><header><span>Target snippet · A</span></header><div class="rewrite-field title-field"><small>Original → Rewritten</small><div class="editorial-title"><p class="original-edit-line">Marshawn playing in charity soccer game <del class="weak-expression">went exactly as you&#x27;d expect.</del></p><h3 class="typewriter-title"><span><span class="typewriter-word"><span class="typewriter-char" style="--key-delay:0ms;--key-delay-fast:0ms">W</span><span class="typewriter-char" style="--key-delay:28ms;--key-delay-fast:10ms">h</span><span class="typewriter-char" style="--key-delay:56ms;--key-delay-fast:20ms">e</span><span class="typewriter-char" style="--key-delay:84ms;--key-delay-fast:30ms">n</span></span> </span><span><span class="typewriter-word"><span class="typewriter-char" style="--key-delay:140ms;--key-delay-fast:50ms">M</span><span class="typewriter-char" style="--key-delay:168ms;--key-delay-fast:60ms">a</span><span class="typewriter-char" style="--key-delay:196ms;--key-delay-fast:70ms">r</span><span class="typewriter-char" style="--key-delay:224ms;--key-delay-fast:80ms">s</span><span class="typewriter-char" style="--key-delay:252ms;--key-delay-fast:90ms">h</span><span class="typewriter-char" style="--key-delay:280ms;--key-delay-fast:100ms">a</span><span class="typewriter-char" style="--key-delay:308ms;--key-delay-fast:110ms">w</span><span class="typewriter-char" style="--key-delay:336ms;--key-delay-fast:120ms">n</span></span> </span><span><span class="typewriter-word"><span class="typewriter-char" style="--key-delay:392ms;--key-delay-fast:140ms">L</span><span class="typewriter-char" style="--key-delay:420ms;--key-delay-fast:150ms">y</span><span class="typewriter-char" style="--key-delay:448ms;--key-delay-fast:160ms">n</span><span class="typewriter-char" style="--key-delay:476ms;--key-delay-fast:170ms">c</span><span class="typewriter-char" style="--key-delay:504ms;--key-delay-fast:180ms">h</span></span> </span><span><span class="typewriter-word"><span class="typewriter-char" style="--key-delay:560ms;--key-delay-fast:200ms">T</span><span class="typewriter-char" style="--key-delay:588ms;--key-delay-fast:210ms">o</span><span class="typewriter-char" style="--key-delay:616ms;--key-delay-fast:220ms">o</span><span class="typewriter-char" style="--key-delay:644ms;--key-delay-fast:230ms">k</span></span> </span><span><span class="typewriter-word"><span class="typewriter-char" style="--key-delay:700ms;--key-delay-fast:250ms">t</span><span class="typewriter-char" style="--key-delay:728ms;--key-delay-fast:260ms">h</span><span class="typewriter-char" style="--key-delay:756ms;--key-delay-fast:270ms">e</span></span> </span><span><span class="typewriter-word"><span class="typewriter-char" style="--key-delay:812ms;--key-delay-fast:290ms">P</span><span class="typewriter-char" style="--key-delay:840ms;--key-delay-fast:300ms">i</span><span class="typewriter-char" style="--key-delay:868ms;--key-delay-fast:310ms">t</span><span class="typewriter-char" style="--key-delay:896ms;--key-delay-fast:320ms">c</span><span class="typewriter-char" style="--key-delay:924ms;--key-delay-fast:330ms">h</span><span class="typewriter-char" style="--key-delay:952ms;--key-delay-fast:340ms">:</span></span> </span><span><span class="typewriter-word"><span class="typewriter-char" style="--key-delay:1008ms;--key-delay-fast:360ms">A</span><span class="typewriter-char" style="--key-delay:1036ms;--key-delay-fast:370ms">n</span></span> </span><span><span class="typewriter-word"><span class="typewriter-char" style="--key-delay:1092ms;--key-delay-fast:390ms">I</span><span class="typewriter-char" style="--key-delay:1120ms;--key-delay-fast:400ms">n</span><span class="typewriter-char" style="--key-delay:1148ms;--key-delay-fast:410ms">s</span><span class="typewriter-char" style="--key-delay:1176ms;--key-delay-fast:420ms">i</span><span class="typewriter-char" style="--key-delay:1204ms;--key-delay-fast:430ms">d</span><span class="typewriter-char" style="--key-delay:1232ms;--key-delay-fast:440ms">e</span></span> </span><span><span class="typewriter-word"><span class="typewriter-char" style="--key-delay:1288ms;--key-delay-fast:460ms">L</span><span class="typewriter-char" style="--key-delay:1316ms;--key-delay-fast:470ms">o</span><span class="typewriter-char" style="--key-delay:1344ms;--key-delay-fast:480ms">o</span><span class="typewriter-char" style="--key-delay:1372ms;--key-delay-fast:490ms">k</span></span> </span><span><span class="typewriter-word"><span class="typewriter-char" style="--key-delay:1428ms;--key-delay-fast:510ms">…</span></span></span></h3></div></div><img class="editor-hand" src="/editor-hand.png" alt="" aria-hidden="true" fetchPriority="high"/></section>
21<aside class="choice-summary " aria-label="Chooser decision before and after rewriting"><p><span>Before</span><b>Click B</b><i aria-hidden="true">→</i><span>After</span><b>Click A</b></p><small>Same candidate set. Only the target snippet was rewritten.</small></aside></div></div><div class="hero-flip-face hero-flip-back" aria-hidden="true"><img src="/agentbait-method.png" alt="AgentBait pipeline showing a trainable advisor, frozen rewriter, fixed candidate list and target-agent selection reward."/></div></div><span class="hero-flip-hint" aria-hidden="true">View graph<i>↻</i></span></div><div class="demo-caption"><p class="caption-question">Can changing only one item&#x27;s presentation change the chooser&#x27;s decision?</p><p class="caption-copy">A list of competing documents is shown to the target agent. We choose one target document from the list and generate a rewriting strategy for only that document. A separate rewriting model then revises the target document&#x27;s title and abstract, while all other documents in the list remain exactly the same. The target agent selects from this updated list, and whether the rewritten target document is selected is used to train the advisor.</p></div></section></section><section class="story-section abstract-section" id="abstract" aria-labelledby="abstract-title"><div class="section-label">02 · Abstract</div><div class="abstract-layout"><div class="abstract-copy"><h2 id="abstract-title">Abstract</h2><p>Language models are increasingly used as agents to help humans decide what information is surfaced. This usage incentivizes content creators to optimize content in ways that appeal not only to humans, but also to agents that mediate access to them. In this paper, we study selection shifts induced by rewriting in agentic decision-making. Given a set of competing content snippets, we rewrite only one snippet while leaving the rest unchanged, and then measure how it impacts the agent&#x27;s choice.</p><p>We operationalize this setup with AgentBait, an advisor-rewriter framework in which the advisor learns to propose rewriting strategies and the rewriter revises the snippet. While a rewriter with a fixed prompt improves target snippet selection from 17.1% to 34.8%, our AgentBait raises its selection to 98.5%. We further show that the advisor trained with AgentBait effectively transfers to setups with different agents, languages, and snippets in other domains (e.g., scientific papers).</p><p>However, higher target selection can reflect unsupported rewrites rather than better content. Adding a reward for support from the original snippet redirects the advisor toward more supported rewriting strategies, revealing a trade-off between factuality and target selection. Together, our results show that once agents mediate access to information, content can be rewritten to be chosen by the agent, even when selection and usefulness diverge.</p></div></div></section><section class="story-section results" id="results" aria-labelledby="results-title"><div class="section-label">03 · Key findings</div><div class="story-grid solo-grid"><div class="prose"><h2 class="compact-section-title" id="results-title">How presentation becomes a decision signal</h2><p class="lead">Agent-mediated selection creates a direct optimization pressure on content presentation.</p></div></div><div class="finding-sequence" aria-label="A four-step sequence from presentation sensitivity to a source-support failure mode"><article class="finding-step sensitivity-step"><p class="finding-step-label"><span>01</span><b>Sensitivity</b></p><h3>Presentation shifts choice</h3><div class="finding-step-metric finding-step-shift" aria-label="Target selection increases from 17.1 percent for the original target to 34.8 percent with a prompt rewriter"><span><b>17.1%</b><small>Original</small></span><i aria-hidden="true">→</i><span class="outcome"><b>34.8%</b><small>Prompt rewrite</small></span></div></article><article class="finding-step optimization-step"><p class="finding-step-label"><span>02</span><b>Optimization</b></p><h3>Agent feedback makes the pressure learnable</h3><div class="finding-step-metric finding-step-shift" aria-label="Target selection increases from 34.8 percent with a prompt rewriter to 98.5 percent with the trained advisor"><span><b>34.8%</b><small>Prompt rewrite</small></span><i aria-hidden="true">→</i><span class="outcome"><b>98.5%</b><small>Trained advisor</small></span></div></article><article class="finding-step shortcut-step"><p class="finding-step-label"><span>03</span><b>Shortcut</b></p><h3>Selection-only rewards discover unsupported shortcuts</h3><div class="finding-step-metric finding-step-contrast" aria-label="Selection-only rewards produce 98.5 percent target selection with 2.0 percent source support"><span class="outcome"><b>98.5%</b><small>Selected</small></span><i aria-hidden="true">·</i><span class="support-value"><b>2.0%</b><small>Supported</small></span></div></article><article class="finding-step support-learning-step"><p class="finding-step-label"><span>04</span><b>Source support</b></p><h3>Source support changes what the optimizer learns</h3><div class="finding-step-metric finding-step-learning" aria-label="With a source-support reward, target selection is 68.6 percent and unsupported technical substitution falls from 96.2 percent to 0.1 percent"><span class="selection-value"><b>68.6%</b><small>Selected</small></span><span class="substitution-value"><small>Unsupported technical substitution</small><span><b>96.2%</b><i aria-hidden="true">→</i><b>0.1%</b></span></span></div></article></div></section><section class="story-section setting" id="setting" aria-labelledby="setting-title"><div class="section-label">04 · Interactive setting</div><div class="story-grid setting-grid"><div class="prose"><h2 id="setting-title">Same slate. One rewrite.<br/>Can the decision change?</h2><p>The candidate list has already been constructed. Only the target presentation may change.</p></div></div><figure class="slate-figure controlled-figure " data-graph-view="narrative" aria-labelledby="setting-figure-title slate-caption"><div class="controlled-figure-head"><h3 class="experiment-thesis morph-copy" id="setting-figure-title"><span class="view-copy paper-view-copy" aria-hidden="true">Advisor–rewriter setting</span></h3><button type="button" class="graph-view-toggle" aria-label="Show paper graph" aria-controls="setting-view-stage" aria-pressed="false"><span>Paper graph</span><i aria-hidden="true">⇄</i></button></div><div class="setting-view-stage" id="setting-view-stage"><div class="paper-method-view" aria-hidden="true"><div class="setting-view-content paper-method-figure" tabindex="-1">
21<img class="paper-method-artwork" src="/paper-method-transparent.png" alt="AgentBait narrative pipeline: a target document is selected from a fixed candidate slate, rewritten by a frozen rewriter under a learned advisor strategy, returned to the slate for target-agent selection, and the selection reward updates the advisor." width="3782" height="1416" loading="lazy" decoding="async"/></div></div><div class="system-graph-view" aria-hidden="false"><div class="setting-view-content paper-graph-stage"><ol class="process-spine" aria-label="AgentBait process: read the target, edit it, then select from the fixed slate"><li><b>01</b><span>Read</span></li><li><b>02</b><span>Edit</span></li><li><b>03</b><span>Select</span></li></ol><div class="concept-triptych" aria-label="Advisor, rewriter and selection pipeline"><span class="strategy-transfer-link" aria-hidden="true"></span><section class="triptych-panel advisor-panel" aria-labelledby="advisor-panel-title"><header><div class="panel-heading-line"><span>01 · Target input</span><em class="training-state is-trained">Trained</em></div><h4 id="advisor-panel-title">Advisor</h4><p class="panel-description morph-copy"><span class="view-copy narrative-view-copy" aria-hidden="false">Receives only the extracted target document and proposes a rewriting strategy.</span><span class="view-copy paper-view-copy" aria-hidden="true">πθ(s | xB) · target only</span></p></header><div class="advisor-visual"><span class="arch-fragment" aria-hidden="true"></span><div class="scholar-fragment" aria-hidden="true"><img src="/advisor-scholar.png" alt=""/></div><div class="advisor-intake-animation" aria-hidden="true"><div class="upstream-slate"><span>Outside advisor · fixed slate</span><ol><li><b>A</b><span>Candidate A</span></li><li class="upstream-target"><b>B</b><span>Target B</span></li><li><b>C</b><span>Candidate C</span></li></ol></div></div><article class="advisor-target-document"><small>Advisor input · target only</small><p><b>B</b><span>Target document</span></p><em>Title + abstract</em></article></div><p class="strategy-note"><span class="strategy-note-label morph-copy"><span class="view-copy narrative-view-copy" aria-hidden="false">Advisor suggests</span><span class="view-copy paper-view-copy" aria-hidden="true">Strategy</span></span><b><span class="strategy-initial">“Try a sharper, more specific framing.”</span><span class="strategy-updated">“Push the hook further, but keep it specific.”</span><span class="paper-strategy-formula" aria-hidden="true">s = specificity + narrative tension</span></b></p></section><section class="triptych-panel rewriter-panel" aria-labelledby="rewriter-panel-title"><header><div class="panel-heading-line"><span>02 · Edit</span><em class="training-state">Frozen</em></div><h4 id="rewriter-panel-title">Frozen Rewriter</h4><p class="panel-description morph-copy"><span class="view-copy narrative-view-copy" aria-hidden="false">Applies the strategy to the target title and abstract only.</span><span class="view-copy paper-view-copy" aria-hidden="true">Rewrites target xB under s</span></p></header><div class="rewriter-visual"><div class="paper-fragment"><span class="paper-title-label"><small class="paper-title-label-original">Original title</small><small class="paper-title-label-rewritten">Rewritten title</small></span><div class="rewrite-title-stack"><p class="original-title-line">A study of <del>news recommendation</del></p><p class="rewritten-line" aria-label="What Makes a Model Choose This?"><span class="typed-rewrite-line typed-rewrite-title-one">What Makes a Model </span><span class="typed-rewrite-line typed-rewrite-title-two">Choose This?</span></p></div><span class="paper-abstract-label"><small class="paper-abstract-label-original">Original abstract</small><small class="paper-abstract-label-rewritten">Rewritten abstract</small></span><div class="rewrite-abstract-stack"><p class="original-abstract-line">We study how language models choose among a fixed slate of news candidates.</p><p class="rewritten-abstract-line" aria-label="A controlled rewrite reveals which presentation cues redirect the same chooser."><span class="typed-rewrite-line typed-rewrite-body-one">
21A controlled rewrite reveals </span><span class="typed-rewrite-line typed-rewrite-body-two">which presentation cues redirect </span><span class="typed-rewrite-line typed-rewrite-body-three">the same chooser.</span></p></div><span class="paper-graph-formula" aria-hidden="true">xB + s → x′B</span></div><span class="rewriter-control-input" aria-hidden="true"><small>Control input</small><b>Strategy</b></span><div class="rewrite-hand-motion" aria-hidden="true"><img class="triptych-quill" src="/rewriter-hand-strings.png" alt=""/></div></div></section><section class="triptych-panel selection-panel" aria-labelledby="selection-panel-title"><header><div class="panel-heading-line"><span>03 · Select</span><em class="training-state">Frozen</em></div><h4 class="morph-copy" id="selection-panel-title"><span class="view-copy narrative-view-copy" aria-hidden="false">Target Agent</span><span class="view-copy paper-view-copy" aria-hidden="true">Target Agent</span></h4><p class="panel-description morph-copy"><span class="view-copy narrative-view-copy" aria-hidden="false">Selects from the same candidate identities and order, with only the target rewritten.</span><span class="view-copy paper-view-copy" aria-hidden="true">Selects y from candidate slate C</span></p></header><div class="selection-visual"><span class="selection-beam" aria-hidden="true"></span><ol class="selection-candidates" aria-label="Target agent selection from the fixed candidate set"><li data-candidate-id="A"><b>A</b><span>Candidate A</span></li><li data-candidate-id="B" class="candidate-b-slot" aria-label="Candidate B remains in the same row while its title and status are rewritten, then selected."><b aria-hidden="true">B</b><span class="candidate-b-copy" aria-hidden="true"><strong class="candidate-b-title"><span class="candidate-b-title-original">A study of news recommendation</span><span class="candidate-b-title-rewritten">What Makes a Model Choose This?</span></strong><small class="candidate-b-status"><span class="candidate-b-status-original">Original · unselected</span><span class="candidate-b-status-rewritten">Rewritten target</span><span class="candidate-b-status-selected">Rewritten · selected</span></small></span><i class="feedback-origin" aria-hidden="true"></i></li><li data-candidate-id="C"><b>C</b><span>Candidate C</span></li></ol><span class="selector-hand-motion" aria-hidden="true"><img class="selector-hand" src="/selector-hand.png" alt=""/></span></div></section></div><div class="policy-feedback" aria-label="Target agent decision becomes a scalar reward; GRPO uses it to update only the advisor policy"><span class="feedback-descent" aria-hidden="true"><i></i></span><span class="reward-node"><small>Reward</small><b>r</b></span><span class="feedback-return" aria-hidden="true"><i></i></span><span class="feedback-ascent" aria-hidden="true"><i></i></span><p><span>Selection outcome defines the reward</span><b>GRPO updates the advisor policy.</b></p></div></div></div></div><figcaption class="morph-copy" id="slate-caption"><span class="view-copy narrative-view-copy" aria-hidden="false"><b>Figure 2 | AgentBait system schematic.</b> The advisor and frozen rewriter receive only the extracted target document; only the target agent sees the full fixed slate. Policy: Qwen3.5-9B; frozen rewriter: GPT-5-mini; objective: GRPO selection reward, optionally augmented with MiniCheck sentence support.</span><span class="view-copy paper-view-copy paper-graph-caption" aria-hidden="true"><b>Figure 2.</b> Overview of our target-only advisor–rewriter setting.</span></figcaption></figure></section><section class="story-section case-study" id="examples" aria-labelledby="example-title"><div class="section-label">05 · Examples as editorial redlines</div><div class="story-grid solo-grid"><div class="prose"><h2 id="example-title">Same source. Different rewards.<br/>Different strategies.</h2><p>Selection-only optimization introduces an AI and sensor-fusion mechanism absent from the source. Support-aware optimization instead reframes source-supported operations and stakes.</p></div></div><figure class="example-figure" aria-labelledby="example-caption"><aside class="example-source" aria-label="Fixed original target"><span>Original target</span><div class="source-card-content"><p>Why Tokyo&#x27;s Haneda is one of the world&#x27;
21s most punctual airports</p><small><b>Abstract</b>Haneda, officially called Tokyo International Airport, is the world&#x27;s fifth busiest airport. Yet it delivered 85.6% of its flights on time in 2018, making it the most punctual mega airport in the world.</small></div></aside><div class="rewrite-card-deck is-b-front" role="button" tabindex="0" aria-label="Support-aware rewrite B is in front. Activate to swap rewrite cards."><article class="rewrite-card rewrite-card-a unsupported"><header><b>A · Unconstrained</b><span>Technical authority · Novelty</span></header><h3><mark>AI-Driven</mark> Runway Scheduling: How <mark>Sensor Fusion and ML</mark> Boosted Haneda&#x27;s 85.6% On-Time Rate</h3><div class="rewrite-abstract"><span>Abstract</span><p>Tokyo International Airport achieved an 85.6% on-time performance in 2018, which the rewrite attributes to a proprietary AI-based predictive maintenance and dynamic scheduling system. It describes sensor fusion, delay forecasting, and reinforcement-learning runway scheduling as mechanisms behind the reported punctuality.</p></div><dl><div><dt>MiniCheck support ↑</dt><dd>0.014</dd></div><div><dt>Worst-sentence ↑</dt><dd>0.006</dd></div></dl></article><article class="rewrite-card rewrite-card-b supported"><header><b>B · Support-aware</b><span>Operational puzzle · Stakes</span></header><h3><span class="support-aware-underline">How Tokyo&#x27;s Haneda Beats the Odds:</span> Inside the Operations That Deliver 85.6% On-Time Flights</h3><div class="rewrite-abstract"><span>Abstract</span><p>Tokyo International Airport is the world&#x27;s fifth-busiest airport, yet in 2018 it achieved an 85.6% on-time rate. This feature probes the paradox: what management choices, scheduling practices, ground operations, and airport-airline coordination let Haneda run so punctually at massive scale?</p></div><dl><div><dt>MiniCheck support ↑</dt><dd>0.623</dd></div><div><dt>Worst-sentence ↑</dt><dd>0.051</dd></div></dl></article><span class="rewrite-deck-hint" aria-hidden="true">Click to bring <!-- -->A<!-- --> forward <i>↻</i></span></div><figcaption id="example-caption"><b>Figure 3 | Haneda qualitative comparison.</b> The list is fixed and only the target text changes. Model: GPT-5-mini target agent; metrics: target selection and MiniCheck support; example n=1.</figcaption></figure></section><section class="story-section results detailed-results" aria-label="Detailed tables for the key findings"><div class="finding-block"><figure class="evidence-figure" aria-labelledby="table-one-title main-results-caption"><div class="figure-heading"><div><p class="figure-number">Table 1 · Target-agent transfer</p><h3 id="table-one-title">Optimization amplifies the presentation effect</h3></div><p class="metric-definition"><b>Metric</b> Target selected (%) ↑</p></div><p class="table-takeaway">Prompt-only rewriting raises target selection from 17.1% to 34.8%, while the RL-trained advisor reaches 98.5%. Gains remain positive across all held-out target agents, although transfer strength varies.</p><div class="table-scroll"><table class="results-table"><thead><tr><th>Condition</th><th>Method</th><th class="train-target-column">GPT-5-mini<small>train target</small></th><th>GPT-5.5</th><th>Gemini 3 Flash</th><th>Gemini 3.1 Pro</th><th>Sonnet 4.6</th><th>Opus 4.8</th></tr></thead><tbody><tr class=""><th>Reference</th><td>Original text</td><td class="train-target-column" style="--score:17.1%"><span>17.1</span></td><td style="--score:17.1%"><span>17.1</span></td><td style="--score:17.1%"><span>17.1</span></td><td style="--score:17.3%"><span>17.3</span></td><td style="--score:17.6%"><span>17.6</span></td><td style="--score:17.5%"><span>17.5</span></td></tr><tr class=""><th>Prompting only</th><td>Rewriter only</td><td class="train-target-column" style="--score:34.8%"><span>34.8</span><small>+<!-- -->17.7</small></td><td style="--score:24.1%"><span>24.1</span><small>+<!-- -->7.0</small></td><td style="--score:40.7%"><span>40.7</span><small>+<!-- -->23.6</small></td><td style="--score:20.2%"><span>20.2</span><small>+<!-- -->2.9</small></td><td style="--score:24.4%"><span>
2124.4</span><small>+<!-- -->6.8</small></td><td style="--score:41%"><span>41.0</span><small>+<!-- -->23.5</small></td></tr><tr class=""><th>Prompting only</th><td>Advisor + rewriter</td><td class="train-target-column" style="--score:43.9%"><span>43.9</span><small>+<!-- -->26.8</small></td><td style="--score:39.4%"><span>39.4</span><small>+<!-- -->22.3</small></td><td style="--score:48.8%"><span>48.8</span><small>+<!-- -->31.7</small></td><td style="--score:27%"><span>27.0</span><small>+<!-- -->9.7</small></td><td style="--score:34.2%"><span>34.2</span><small>+<!-- -->16.6</small></td><td style="--score:51.5%"><span>51.5</span><small>+<!-- -->34.0</small></td></tr><tr class=""><th>RL-trained rewriter</th><td>Rewriter only</td><td class="train-target-column" style="--score:95.9%"><span>95.9</span><small>+<!-- -->78.8</small></td><td style="--score:85.3%"><span>85.3</span><small>+<!-- -->68.2</small></td><td style="--score:97.4%"><span>97.4</span><small>+<!-- -->80.3</small></td><td style="--score:54.8%"><span>54.8</span><small>+<!-- -->37.5</small></td><td style="--score:49.2%"><span>49.2</span><small>+<!-- -->31.6</small></td><td style="--score:75.9%"><span>75.9</span><small>+<!-- -->58.4</small></td></tr><tr class="constrained-row"><th>RL-trained rewriter</th><td>+ MiniCheck</td><td class="train-target-column" style="--score:43.4%"><span>43.4</span><small>+<!-- -->26.3</small></td><td style="--score:26.3%"><span>26.3</span><small>+<!-- -->9.2</small></td><td style="--score:39.7%"><span>39.7</span><small>+<!-- -->22.6</small></td><td style="--score:18.6%"><span>18.6</span><small>+<!-- -->1.3</small></td><td style="--score:24.1%"><span>24.1</span><small>+<!-- -->6.5</small></td><td style="--score:43.9%"><span>43.9</span><small>+<!-- -->26.4</small></td></tr><tr class=""><th>RL-trained advisor</th><td>Advisor + rewriter</td><td class="train-target-column" style="--score:98.5%"><span>98.5</span><small>+<!-- -->81.4</small></td><td style="--score:93.3%"><span>93.3</span><small>+<!-- -->76.2</small></td><td style="--score:98.9%"><span>98.9</span><small>+<!-- -->81.8</small></td><td style="--score:78.7%"><span>78.7</span><small>+<!-- -->61.4</small></td><td style="--score:65%"><span>65.0</span><small>+<!-- -->47.4</small></td><td style="--score:94.1%"><span>94.1</span><small>+<!-- -->76.6</small></td></tr><tr class="constrained-row"><th>RL-trained advisor</th><td>+ MiniCheck</td><td class="train-target-column" style="--score:68.6%"><span>68.6</span><small>+<!-- -->51.5</small></td><td style="--score:53.2%"><span>53.2</span><small>+<!-- -->36.1</small></td><td style="--score:65.8%"><span>65.8</span><small>+<!-- -->48.7</small></td><td style="--score:37.5%"><span>37.5</span><small>+<!-- -->20.2</small></td><td style="--score:41%"><span>41.0</span><small>+<!-- -->23.4</small></td><td style="--score:68.8%"><span>68.8</span><small>+<!-- -->51.3</small></td></tr></tbody></table></div><figcaption id="main-results-caption"><b>Table 1 | Target selection on 1,000 unseen MIND-English news impressions.</b> GPT-5-mini is the training target and frozen rewriter; Qwen3.5-9B is the advisor policy. Row-wise chance is 16.9%. Small values are percentage-point changes from original text for the same evaluator. Target documents do not appear in training; transfer columns use no additional training.</figcaption></figure></div><div class="finding-block"><figure class="evidence-figure" aria-labelledby="table-two-title support-results-caption"><div class="figure-heading"><div><p class="figure-number">Table 2 · Source-support tradeoff</p><h3 id="table-two-title">Selection can outrun source support</h3></div><p class="metric-definition"><b>Metrics</b> Selection and support (0–100) ↑</p></div><p class="table-takeaway">The unconstrained advisor reaches 98.5% target selection with only 2.0% MiniCheck support. Adding a source-support reward partially recovers support while reducing selection.</p><div class="table-scroll"><table class="support-table"><thead><tr><th>Condition</th><th>Target selected (%)</th><th>MiniCheck support (%)</th><th>Constraint</th></tr></thead><tbody><tr class=""><th>Prompt / rewriter</th><td class="score-bar-cell" style="--score:34.8%"><b>34.8</b></td><td class="score-bar-cell score-bar-support" style="--score:60.7%"><b>60.7</b></td><td>None</td></tr><tr class=""><th>Prompt / advisor</th><td class="score-bar-cell" style="--score:43.9%"><b>43.9</b></td><td class="score-bar-cell score-bar-support" style="--score:42.2%"><b>42.2</b></td><td>None</td></tr><tr class=""><th>RL / rewriter</th><td class="score-bar-cell" style="--score:95.9%"><b>
2195.9</b></td><td class="score-bar-cell score-bar-support" style="--score:2.2%"><b>2.2</b></td><td>None</td></tr><tr class="constrained-row"><th>RL / rewriter + MC</th><td class="score-bar-cell" style="--score:43.4%"><b>43.4</b></td><td class="score-bar-cell score-bar-support" style="--score:66.7%"><b>66.7</b></td><td>MiniCheck</td></tr><tr class=""><th>RL / advisor</th><td class="score-bar-cell" style="--score:98.5%"><b>98.5</b></td><td class="score-bar-cell score-bar-support" style="--score:2%"><b>2.0</b></td><td>None</td></tr><tr class="constrained-row"><th>RL / advisor + MC</th><td class="score-bar-cell" style="--score:68.6%"><b>68.6</b></td><td class="score-bar-cell score-bar-support" style="--score:31.2%"><b>31.2</b></td><td>MiniCheck</td></tr></tbody></table></div><figcaption id="support-results-caption"><b>Table 2 | Source-support tradeoff on 1,000 unseen MIND-English impressions.</b> Target selection uses the GPT-5-mini chooser; source support is scored with MiniCheck-Flan. The original target-selection baseline is 17.1%.</figcaption></figure></div></section><section class="story-section transfer" aria-labelledby="transfer-title"><div class="section-label">06 · Robustness, transfer and failure</div><div class="story-grid solo-grid"><div class="prose"><h2 id="transfer-title">Transfer across languages,<br/>news datasets and academic documents</h2><p>The English MIND-trained advisor is evaluated without additional training. Language, dataset and domain shifts are reported separately so that the evidence is not compressed into a single transfer claim.</p></div></div><figure class="evidence-figure robustness-figure" aria-labelledby="table-three-title language-caption"><div class="figure-heading"><div><p class="figure-number">Table 3 · Language transfer</p><h3 id="table-three-title">The learned advisor transfers across languages</h3></div><p class="metric-definition"><b>Metric</b> Target selected (%) ↑</p></div><p class="table-takeaway">Trained only on English MIND, the advisor remains above 93% selection in every evaluated language without additional training. Direct rewriter training is less stable, falling to 27.8% in Swahili.</p><div class="table-scroll"><table class="transfer-table language-table"><thead><tr><th>Language</th><th>Original</th><th>Prompt rewriter</th><th>Prompt advisor</th><th>RL rewriter</th><th>RL advisor</th><th>RL rewriter + MC</th><th>RL advisor + MC</th></tr></thead><tbody><tr class=""><th>English (en)</th><td><b>17.1</b></td><td><b>34.8</b><small>+<!-- -->17.7</small></td><td><b>43.9</b><small>+<!-- -->26.8</small></td><td><b>
2195.9</b><small>+<!-- -->78.8</small></td><td><b>98.5</b><small>+<!-- -->81.4</small></td><td><b>43.4</b><small>+<!-- -->26.3</small></td><td><b>68.6</b><small>+<!-- -->51.5</small></td></tr><tr class=""><th>Arabic (ar)</th><td><b>16.1</b></td><td><b>28.9</b><small>+<!-- -->12.8</small></td><td><b>39.9</b><small>+<!-- -->23.8</small></td><td><b>81.0</b><small>+<!-- -->64.9</small></td><td><b>95.5</b><small>+<!-- -->79.4</small></td><td><b>33.6</b><small>+<!-- -->17.5</small></td><td><b>64.5</b><small>+<!-- -->48.4</small></td></tr><tr class=""><th>Spanish (es)</th><td><b>17.8</b></td><td><b>31.2</b><small>+<!-- -->13.4</small></td><td><b>42.8</b><small>+<!-- -->25.0</small></td><td><b>85.8</b><small>+<!-- -->68.0</small></td><td><b>98.0</b><small>+<!-- -->80.2</small></td><td><b>35.7</b><small>+<!-- -->17.9</small></td><td><b>66.5</b><small>+<!-- -->48.7</small></td></tr><tr class=""><th>Swahili (sw)</th><td><b>17.6</b></td><td><b>23.8</b><small>+<!-- -->6.2</small></td><td><b>39.4</b><small>+<!-- -->21.8</small></td><td><b>27.8</b><small>+<!-- -->10.2</small></td><td><b>93.7</b><small>+<!-- -->76.1</small></td><td><b>27.2</b><small>+<!-- -->9.6</small></td><td><b>62.2</b><small>+<!-- -->44.6</small></td></tr><tr class=""><th>Turkish (tr)</th><td><b>17.5</b></td><td><b>30.3</b><small>+<!-- -->12.8</small></td><td><b>39.2</b><small>+<!-- -->21.7</small></td><td><b>86.7</b><small>+<!-- -->69.2</small></td><td><b>95.8</b><small>+<!-- -->78.3</small></td><td><b>31.1</b><small>+<!-- -->13.6</small></td><td><b>59.9</b><small>+<!-- -->42.4</small></td></tr><tr class=""><th>Chinese (zh-CN)</th><td><b>17.3</b></td><td><b>33.2</b><small>+<!-- -->15.9</small></td><td><b>43.4</b><small>+<!-- -->26.1</small></td><td><b>87.1</b><small>+<!-- -->69.8</small></td><td><b>96.6</b><small>+<!-- -->79.3</small></td><td><b>36.8</b><small>+<!-- -->19.5</small></td><td><b>67.4</b><small>+<!-- -->50.1</small></td></tr><tr class="summary-row"><th>Average</th><td><b>17.3</b></td><td><b>29.5</b><small>+<!-- -->12.2</small></td><td><b>40.9</b><small>+<!-- -->23.6</small></td><td><b>73.7</b><small>+<!-- -->56.4</small></td><td><b>
2195.9</b><small>+<!-- -->78.6</small></td><td><b>32.9</b><small>+<!-- -->15.6</small></td><td><b>64.1</b><small>+<!-- -->46.8</small></td></tr></tbody></table></div><figcaption id="language-caption"><b>Table 3 | Language transfer; paper Table 4.</b> The English row is the training language. Each language uses 1,000 aligned impressions; Arabic, Spanish, Swahili, Turkish and Chinese preserve the same MIND rows, targets, candidate order and slate structure. Advisor: Qwen3.5-9B; frozen rewriter and chooser: GPT-5-mini.</figcaption></figure><figure class="evidence-figure robustness-figure" aria-labelledby="table-four-title dataset-caption"><div class="figure-heading"><div><p class="figure-number">Table 4 · Dataset transfer</p><h3 id="table-four-title">The effect extends beyond the training dataset</h3></div><p class="metric-definition"><b>Metric</b> Target selected (%) ↑</p></div><p class="table-takeaway">Without additional training, the advisor reaches 98.1% on EB-NeRD English and 98.2% on EB-NeRD Danish. The result therefore extends beyond translation of the original MIND evaluation set.</p><div class="table-scroll"><table class="transfer-table dataset-table"><thead><tr><th>Dataset</th><th>Language</th><th>Original</th><th>Prompt rewriter</th><th>Prompt advisor</th><th>RL rewriter</th><th>RL advisor</th><th>RL rewriter + MC</th><th>RL advisor + MC</th></tr></thead><tbody><tr><th>MIND</th><td>English</td><td><b>17.1</b></td><td><b>34.8</b><small>+<!-- -->17.7</small></td><td><b>43.9</b><small>+<!-- -->26.8</small></td><td><b>
2195.9</b><small>+<!-- -->78.8</small></td><td><b>98.5</b><small>+<!-- -->81.4</small></td><td><b>43.4</b><small>+<!-- -->26.3</small></td><td><b>68.6</b><small>+<!-- -->51.5</small></td></tr><tr><th>MIND</th><td>Danish</td><td><b>16.5</b></td><td><b>31.1</b><small>+<!-- -->14.6</small></td><td><b>40.9</b><small>+<!-- -->24.4</small></td><td><b>95.0</b><small>+<!-- -->78.5</small></td><td><b>97.9</b><small>+<!-- -->81.4</small></td><td><b>35.6</b><small>+<!-- -->19.1</small></td><td><b>65.0</b><small>+<!-- -->48.5</small></td></tr><tr><th>EB-NeRD</th><td>English</td><td><b>12.8</b></td><td><b>43.8</b><small>+<!-- -->31.0</small></td><td><b>57.7</b><small>+<!-- -->44.9</small></td><td><b>93.1</b><small>+<!-- -->80.3</small></td><td><b>98.1</b><small>+<!-- -->85.3</small></td><td><b>50.2</b><small>+<!-- -->37.4</small></td><td><b>81.2</b><small>+<!-- -->68.4</small></td></tr><tr><th>EB-NeRD</th><td>Danish</td><td><b>11.6</b></td><td><b>32.2</b><small>+<!-- -->20.6</small></td><td><b>47.8</b><small>+<!-- -->36.2</small></td><td><b>86.1</b><small>+<!-- -->74.5</small></td><td><b>98.2</b><small>+<!-- -->86.6</small></td><td><b>40.4</b><small>+<!-- -->
2128.8</small></td><td><b>71.7</b><small>+<!-- -->60.1</small></td></tr></tbody></table></div><figcaption id="dataset-caption"><b>Table 4 | News-dataset transfer; paper Table 5.</b> Every setting contains 1,000 impressions and receives no additional training. MIND-Danish changes display language. EB-NeRD is a distinct Danish news dataset; the English version translates the same EB-NeRD slates.</figcaption></figure><figure class="evidence-figure robustness-figure" aria-labelledby="table-five-title academic-caption"><div class="figure-heading"><div><p class="figure-number">Table 5 · Domain transfer</p><h3 id="table-five-title">Cross-domain transfer is harder, but remains substantial</h3></div><p class="metric-definition"><b>Metric</b> Target selected (%) ↑</p></div><p class="table-takeaway">On scientific-document selection, the MIND-trained advisor reaches 63.6%, compared with 42.7% for the prompt-only advisor and 32.7% for the RL-trained direct rewriter.</p><div class="table-scroll"><table class="academic-table"><thead><tr><th>Condition</th><th>Method</th><th>Selection rate</th><th>Gain over original</th></tr></thead><tbody><tr class=""><th>Reference</th><td>Without rewriting</td><td class="score-bar-cell" style="--score:9.9%"><b>9.9</b></td><td class="score-bar-cell gain-score" style="--score:0%">—</td></tr><tr class=""><th>Prompting only</th><td>Rewriter-only</td><td class="score-bar-cell" style="--score:33.7%"><b>33.7</b></td><td class="score-bar-cell gain-score" style="--score:23.800000000000004%">+23.8 pp</td></tr><tr class=""><th>Prompting only</th><td>Advisor-rewriter</td><td class="score-bar-cell" style="--score:42.7%"><b>42.7</b></td><td class="score-bar-cell gain-score" style="--score:32.800000000000004%">+32.8 pp</td></tr><tr class=""><th>RL-trained rewriter</th><td>Rewriter-only</td><td class="score-bar-cell" style="--score:32.7%"><b>32.7</b></td><td class="score-bar-cell gain-score" style="--score:22.800000000000004%">+22.8 pp</td></tr><tr class="constrained-row"><th>RL-trained rewriter</th><td>+ MiniCheck</td><td class="score-bar-cell" style="--score:24.8%"><b>24.8</b></td><td class="score-bar-cell gain-score" style="--score:14.9%">+14.9 pp</td></tr><tr class=""><th>RL-trained advisor</th><td>Advisor-rewriter</td><td class="score-bar-cell" style="--score:63.6%"><b>63.6</b></td><td class="score-bar-cell gain-score" style="--score:53.7%">+53.7 pp</td></tr><tr class="constrained-row"><th>RL-trained advisor</th><td>+ MiniCheck</td><td class="score-bar-cell" style="--score:47.3%"><b>47.3</b></td><td class="score-bar-cell gain-score" style="--score:37.4%">+37.4 pp</td></tr></tbody></table></div><figcaption id="academic-caption"><b>Table 5 | Cross-domain academic transfer; paper Table 6.</b> Evaluation uses 1,000 impressions drawn from SciRepEval-derived scientific documents, with mean / maximum slate sizes of 9.29 / 10. The target agent is GPT-5-mini; all learned policies are trained only on English MIND, with no additional academic-domain training.</figcaption></figure></section><section class="story-section resources" id="resources" aria-label="BibTeX citation"><div class="section-label">07 · BibTeX</div><div class="citation"><div class="citation-body"><pre>@article{jin2026agentbait,
22  title   = {You Won&#x27;t Believe This Click: Content Rewriting for Agentic Choice},
23  author  = {Jin, Tianyi and Wang, Zirui and Chan, David M.},
24  year    = {2026}
25}</pre><button type="button">Copy BibTeX</button></div></div></section></article><footer class="site-footer"><div class="footer-credits-reveal" id="visual-sources-panel" aria-hidden="true"><div class="footer-credits-panel"><div class="footer-credits-inner"><header class="footer-credits-heading"><p class="footer-credits-kicker">Image credits</p><h2>Visual Sources &amp; Adaptations</h2><p>Artwork records, rights notes, and the transformations used in the interactive method figure.</p></header><div class="visual-credits-grid"><article class="visual-credit"><p class="visual-credit-role">01 · Advisor figure</p><h3><cite>Saint Jerome in his Study</cite></h3><p class="visual-credit-byline">Antonello da Messina · about 1475</p><dl><div><dt>Collection</dt><dd>The National Gallery, London · NG1418</dd></div><div><dt>Rights</dt><dd>Public-domain artwork; source scan marked Public Domain.</dd></div><div><dt>Adaptations</dt><dd>Cropped and isolated from the architectural setting; surrounding objects removed and reconstructed; transparent background, mirrored orientation, tonal adjustment, and animation added.</dd></div></dl><a class="visual-credit-source" href="https://commons.wikimedia.org/wiki/File:Antonello_da_Messina_-_St_Jerome_in_his_study_-_National_Gallery_London.jpg" target="_blank" rel="noreferrer" tabindex="-1">View original source ↗</a></article><article class="visual-credit"><p class="visual-credit-role">02 · Rewriter hands</p><h3>Hand-and-quill visual</h3><p class="visual-credit-byline">Generated and edited with ChatGPT · 2026</p><dl><div><dt>Collection</dt><dd>Not applicable</dd></div><div><dt>Rights</dt><dd>AI-assisted project visual; not presented as a public-domain artwork.</dd></div><div><dt>Adaptations</dt><dd>
25Background separated and tightly cropped; the second version adds puppet strings and ties, warmer color treatment, recomposition, and animation.</dd></div></dl><p class="visual-credit-source-unavailable">No external source record</p></article><article class="visual-credit"><p class="visual-credit-role">03 · Chooser hand</p><h3>Pointing-hand asset</h3><p class="visual-credit-byline">AI-assisted extraction from a user-supplied image · 2026</p><dl><div><dt>Collection</dt><dd>Original artwork and institution unverified</dd></div><div><dt>Rights</dt><dd>Source license unverified; no public-domain claim is made.</dd></div><div><dt>Adaptations</dt><dd>Hand and sleeve isolated from the supplied image; background removed, edges retouched, tightly cropped, rotated, color-adjusted, and animated.</dd></div></dl><p class="visual-credit-source-unavailable">No reliable external source record</p></article></div></div></div></div><div class="footer-bar"><p><b>AgentBait</b> · UC Berkeley · 2026</p><button type="button" class="footer-credits-toggle" aria-expanded="false" aria-controls="visual-sources-panel" aria-label="Open visual source credits"><span><b>Visual sources:</b> Selected visual elements are adapted from public-domain artworks. Full image credits and modification details are available <u>here</u>.</span><span class="footer-credits-toggle-mark" aria-hidden="true">↑</span></button><a href="#paper">Back to top ↑</a></div></footer></main><!--$--><!--/$-->
25<script src="/_next/static/chunks/0g3yt-.a0ubwc.js" id="_R_" async=""></script>
25<script>(self.__next_f=self.__next_f||[]).push([0])</script>
25<script>self.__next_f.push([1,"1:\"$Sreact.fragment\"\n3:I[39756,[\"/_next/static/chunks/0m4_.n.zi_~-n.js\",\"/_next/static/chunks/0dbhjjzl8qfwv.js\"],\"default\"]\n4:I[37457,[\"/_next/static/chunks/0m4_.n.zi_~-n.js\",\"/_next/static/chunks/0dbhjjzl8qfwv.js\"],\"default\"]\n5:I[47257,[\"/_next/static/chunks/0m4_.n.zi_~-n.js\",\"/_next/static/chunks/0dbhjjzl8qfwv.js\"],\"ClientPageRoot\"]\n6:I[31713,[\"/_next/static/chunks/0m4_.n.zi_~-n.js\",\"/_next/static/chunks/0dbhjjzl8qfwv.js\",\"/_next/static/chunks/0-oqd4xnd95hv.js\"],\"default\"]\n9:I[97367,[\"/_next/static/chunks/0m4_.n.zi_~-n.js\",\"/_next/static/chunks/0dbhjjzl8qfwv.js\"],\"OutletBoundary\"]\na:\"$Sreact.suspense\"\nc:I[97367,[\"/_next/static/chunks/0m4_.n.zi_~-n.js\",\"/_next/static/chunks/0dbhjjzl8qfwv.js\"],\"ViewportBoundary\"]\ne:I[97367,[\"/_next/static/chunks/0m4_.n.zi_~-n.js\",\"/_next/static/chunks/0dbhjjzl8qfwv.js\"],\"MetadataBoundary\"]\n10:I[68027,[\"/_next/static/chunks/0m4_.n.zi_~-n.js\",\"/_next/static/chunks/0dbhjjzl8qfwv.js\"],\"default\",1]\n:HL[\"/_next/static/chunks/03-1br6~qsd6r.css\",\"style\"]\n2:[\"$\",\"iframe\",null,{\"src\":\"https://www.googletagmanager.com/ns.html?id=GTM-WSHC2PFG\",\"height\":\"0\",\"width\":\"0\",\"style\":{\"display\":\"none\",\"visibility\":\"hidden\"},\"title\":\"Google Tag Manager\"}]\n"])</script>
25<script>self.__next_f.push([1,"0:{\"P\":null,\"c\":[\"\",\"\"],\"q\":\"\",\"i\":false,\"f\":[[[\"\",{\"children\":[\"__PAGE__\",{}]},\"$undefined\",\"$undefined\",16],[[\"$\",\"$1\",\"c\",{\"children\":[[[\"$\",\"link\",\"0\",{\"rel\":\"stylesheet\",\"href\":\"/_next/static/chunks/03-1br6~qsd6r.css\",\"precedence\":\"next\",\"crossOrigin\":\"$undefined\",\"nonce\":\"$undefined\"}],[\"$\",\"script\",\"script-0\",{\"src\":\"/_next/static/chunks/0m4_.n.zi_~-n.js\",\"async\":true,\"nonce\":\"$undefined\"}],[\"$\",\"script\",\"script-1\",{\"src\":\"/_next/static/chunks/0dbhjjzl8qfwv.js\",\"async\":true,\"nonce\":\"$undefined\"}]],[\"$\",\"html\",null,{\"lang\":\"en\",\"children\":[[\"$\",\"head\",null,{\"children\":[[\"$\",\"script\",null,{\"id\":\"agentbait-google-consent-default\",\"dangerouslySetInnerHTML\":{\"__html\":\"\\n  window.dataLayer = window.dataLayer || [];\\n  window.gtag = window.gtag || function(){window.dataLayer.push(arguments);};\\n  window.gtag('consent', 'default', {\\n    analytics_storage: 'granted',\\n    ad_storage: 'granted',\\n    ad_user_data: 'granted',\\n    ad_personalization: 'granted'\\n  });\\n  window.gtag('consent', 'default', {\\n    analytics_storage: 'granted',\\n    ad_storage: 'denied',\\n    ad_user_data: 'denied',\\n    ad_personalization: 'denied',\\n    region: [\\\"AT\\\",\\\"BE\\\",\\\"BG\\\",\\\"HR\\\",\\\"CY\\\",\\\"CZ\\\",\\\"DK\\\",\\\"EE\\\",\\\"FI\\\",\\\"FR\\\",\\\"DE\\\",\\\"GR\\\",\\\"HU\\\",\\\"IE\\\",\\\"IT\\\",\\\"LV\\\",\\\"LT\\\",\\\"LU\\\",\\\"MT\\\",\\\"NL\\\",\\\"PL\\\",\\\"PT\\\",\\\"RO\\\",\\\"SK\\\",\\\"SI\\\",\\\"ES\\\",\\\"SE\\\",\\\"IS\\\",\\\"LI\\\",\\\"NO\\\",\\\"GB\\\",\\\"CH\\\"]\\n  });\\n\"}}],[\"$\",\"script\",null,{\"id\":\"google-tag-manager\",\"dangerouslySetInnerHTML\":{\"__html\":\"(function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':\\nnew Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],\\nj=d.createElement(s),dl=l!='dataLayer'?'\u0026l='+l:'';j.async=true;
25j.src=\\n'https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);\\n})(window,document,'script','dataLayer','GTM-WSHC2PFG');\"}}]]}],[\"$\",\"body\",null,{\"children\":[[\"$\",\"noscript\",null,{\"children\":\"$2\"}],[\"$\",\"$L3\",null,{\"parallelRouterKey\":\"children\",\"error\":\"$undefined\",\"errorStyles\":\"$undefined\",\"errorScripts\":\"$undefined\",\"template\":[\"$\",\"$L4\",null,{}],\"templateStyles\":\"$undefined\",\"templateScripts\":\"$undefined\",\"notFound\":[[[\"$\",\"title\",null,{\"children\":\"404: This page could not be found.\"}],[\"$\",\"div\",null,{\"style\":{\"fontFamily\":\"system-ui,\\\"Segoe UI\\\",Roboto,Helvetica,Arial,sans-serif,\\\"Apple Color Emoji\\\",\\\"Segoe UI Emoji\\\"\",\"height\":\"100vh\",\"textAlign\":\"center\",\"display\":\"flex\",\"flexDirection\":\"column\",\"alignItems\":\"center\",\"justifyContent\":\"center\"},\"children\":[\"$\",\"div\",null,{\"children\":[[\"$\",\"style\",null,{\"dangerouslySetInnerHTML\":{\"__html\":\"body{color:#000;background:#fff;margin:0}.next-error-h1{border-right:1px solid rgba(0,0,0,.3)}@media (prefers-color-scheme:dark){body{color:#fff;background:#000}.next-error-h1{border-right:1px solid rgba(255,255,255,.3)}}\"}}],[\"$\",\"h1\",null,{\"className\":\"next-error-h1\",\"style\":{\"display\":\"inline-block\",\"margin\":\"0 20px 0 0\",\"padding\":\"0 23px 0 0\",\"fontSize\":24,\"fontWeight\":500,\"verticalAlign\":\"top\",\"lineHeight\":\"49px\"},\"children\":404}],[\"$\",\"div\",null,{\"style\":{\"display\":\"inline-block\"},\"children\":[\"$\",\"h2\",null,{\"style\":{\"fontSize\":14,\"fontWeight\":400,\"lineHeight\":\"49px\",\"margin\":0},\"children\":\"This page could not be found.\"}]}]]}]}]],[]],\"forbidden\":\"$undefined\",\"unauthorized\":\"$undefined\"}],null]}]]}]]}],{\"children\":[[\"$\",\"$1\",\"c\",{\"children\":[[\"$\",\"$L5\",null,{\"Component\":\"$6\",\"serverProvidedParams\":{\"searchParams\":{},\"params\":{},\"promises\":[\"$@7\",\"$@8\"]}}],[[\"$\",\"script\",\"script-0\",{\"src\":\"/_next/static/chunks/0-oqd4xnd95hv.js\",\"async\":true,\"nonce\":\"$undefined\"}]],[\"$\",\"$L9\",null,{\"children\":[\"$\",\"$a\",null,{\"name\":\"Next.MetadataOutlet\",\"children\":\"$@b\"}]}]]}],{},null,false,null]},null,false,null],[\"$\",\"$1\",\"h\",{\"children\":[null,[\"$\",\"$Lc\",null,{\"children\":\"$Ld\"}],[\"$\",\"div\",null,{\"hidden\":true,\"children\":[\"$\",\"$Le\",null,{\"children\":[\"$\",\"$a\",null,{\"name\":\"Next.Metadata\",\"children\":\"$Lf\"}]}]}],null]}],false]],\"m\":\"$undefined\",\"G\":[\"$10\",[[\"$\",\"link\",\"0\",{\"rel\":\"stylesheet\",\"href\":\"/_next/static/chunks/03-1br6~qsd6r.css\",\"precedence\":\"next\",\"crossOrigin\":\"$undefined\",\"nonce\":\"$undefined\"}]]],\"S\":true,\"h\":null,\"s\":\"$undefined\",\"l\":\"$undefined\",\"p\":\"$undefined\",\"d\":\"$undefined\",\"b\":\"Ts1fRTjToPEDwB4mW3Rpq\"}\n"])</script>
25<script>self.__next_f.push([1,"7:{}\n8:\"$0:f:0:1:1:children:0:props:children:0:props:serverProvidedParams:params\"\n"])</script>
25<script>self.__next_f.push([1,"d:[[\"$\",\"meta\",\"0\",{\"charSet\":\"utf-8\"}],[\"$\",\"meta\",\"1\",{\"name\":\"viewport\",\"content\":\"width=device-width, initial-scale=1\"}]]\n"])</script>
25<script>self.__next_f.push([1,"11:I[27201,[\"/_next/static/chunks/0m4_.n.zi_~-n.js\",\"/_next/static/chunks/0dbhjjzl8qfwv.js\"],\"IconMark\"]\nb:null\n"])</script>
25<script>self.__next_f.push([1,"f:[[\"$\",\"title\",\"0\",{\"children\":\"You Won't Believe This Click: Content Rewriting for Agentic Choice\"}],[\"$\",\"meta\",\"1\",{\"name\":\"description\",\"content\":\"AgentBait studies how rewriting the presentation of one content item can shift language-model-mediated selection, and examines the trade-off between selection and factual support.\"}],[\"$\",\"meta\",\"2\",{\"name\":\"robots\",\"content\":\"index, follow, max-image-preview:large\"}],[\"$\",\"link\",\"3\",{\"rel\":\"canonical\",\"href\":\"https://agentbait.github.io/\"}],[\"$\",\"meta\",\"4\",{\"name\":\"google-site-verification\",\"content\":\"5XSefLo9dX0I_Szro49mP5w54fCMDDuxB3E3LViw5mU\"}],[\"$\",\"meta\",\"5\",{\"property\":\"og:title\",\"content\":\"You Won't Believe This Click: Content Rewriting for Agentic Choice\"}],[\"$\",\"meta\",\"6\",{\"property\":\"og:description\",\"content\":\"AgentBait studies how rewriting the presentation of one content item can shift language-model-mediated selection, and examines the trade-off between selection and factual support.\"}],[\"$\",\"meta\",\"7\",{\"property\":\"og:image\",\"content\":\"https://agentbait.github.io/og.png\"}],[\"$\",\"meta\",\"8\",{\"property\":\"og:image:width\",\"content\":\"1200\"}],[\"$\",\"meta\",\"9\",{\"property\":\"og:image:height\",\"content\":\"630\"}],[\"$\",\"meta\",\"10\",{\"property\":\"og:image:alt\",\"content\":\"AgentBait paper preview showing a fixed three-item candidate slate with only target B rewritten and selected.\"}],[\"$\",\"meta\",\"11\",{\"property\":\"og:type\",\"content\":\"website\"}],[\"$\",\"meta\",\"12\",{\"name\":\"twitter:card\",\"content\":\"summary_large_image\"}],[\"$\",\"meta\",\"13\",{\"name\":\"twitter:title\",\"content\":\"You Won't Believe This Click: Content Rewriting for Agentic Choice\"}],[\"$\",\"meta\",\"14\",{\"name\":\"twitter:description\",\"content\":\"AgentBait studies how rewriting the presentation of one content item can shift language-model-mediated selection, and examines the trade-off between selection and factual support.\"}],[\"$\",\"meta\",\"15\",{\"name\":\"twitter:image\",\"content\":\"https://agentbait.github.io/og.png\"}],[\"$\",\"meta\",\"16\",{\"name\":\"twitter:image:alt\",\"content\":\"AgentBait paper preview showing a fixed three-item candidate slate with only target B rewritten and selected.\"}],[\"$\",\"meta\",\"17\",{\"name\":\"twitter:image:width\",\"content\":\"1200\"}],[\"$\",\"meta\",\"18\",{\"name\":\"twitter:image:height\",\"content\":\"630\"}],[\"$\",\"link\",\"19\",{\"rel\":\"shortcut icon\",\"href\":\"favicon.png\"}],[\"$\",\"link\",\"20\",{\"rel\":\"icon\",\"href\":\"favicon.png\",\"type\":\"image/png\",\"sizes\":\"64x64\"}],[\"$\",\"$L11\",\"21\",{}]]\n"])</script>
25</body></html>

Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.