<!DOCTYPE html><html lang="en" xmlns="http://www.w3.org/1999/xhtml" xmlns:v="urn:schemas-microsoft-com:vml" xmlns:o="urn:schemas-microsoft-com:office:office" style="font-size:16px;"><head></head><head><meta charset="utf-8"/><!--[if !mso]><!--><meta http-equiv="X-UA-Compatible" content="IE=edge"/><!--<![endif]--><meta name="viewport" content="width=device-width,initial-scale=1"/><meta name="x-apple-disable-message-reformatting"/><meta name="format-detection" content="telephone=no,address=no,email=no,date=no,url=no"/><meta name="color-scheme" content="light"/><meta name="supported-color-schemes" content="light"/><title>How LLMs Pass the Math Olympiad and Why Models Like Grok Can Be Manipulated to Praise Hitler</title><!--[if mso]><xml><o:OfficeDocumentSettings><o:AllowPNG/><o:PixelsPerInch>96</o:PixelsPerInch></o:OfficeDocumentSettings></xml><![endif]--><style>
:root { color-scheme: light; supported-color-schemes: light; }
body { margin: 0; padding: 0; min-width: 100%!important; -ms-text-size-adjust: 100% !important; -webkit-transform: scale(1) !important; -webkit-text-size-adjust: 100% !important; -webkit-font-smoothing: antialiased !important; }
.body { word-wrap: normal; word-spacing:normal; }
table.mso { width: 100%; border-collapse: collapse; padding: 0; table-layout: fixed; }
img { border: 0; outline: none; }
table { mso-table-lspace: 0px; mso-table-rspace: 0px; }
td, a, span { mso-line-height-rule: exactly; }
#root [x-apple-data-detectors=true],
a[x-apple-data-detectors=true],
#MessageViewBody a { color: inherit !important; text-decoration: inherit !important; font-size: inherit !important; font-family: inherit !important; font-weight: inherit !important; line-height: inherit !important; }
span.MsoHyperlink { color: inherit !important; mso-style-priority: 99 !important; }
span.MsoHyperlinkFollowed { color: inherit !important; mso-style-priority: 99 !important; }
.a { background-color:#dedede; }
.b { background-color:#2a2a2a; }
.c { background-color:#ffffff; }
.d { background-color:#fff0c8; }
.d2 { background-color:#FFFFFF; }
.d3 { background-color:#FFFFFF; }
h1 a { text-decoration:none;color:#2C81E5;font-style:italic; }
h2 a { text-decoration:none;color:#2C81E5;font-style:italic; }
h3 a { text-decoration:none;color:#2C81E5;font-style:italic; }
h4 a { text-decoration:none;color:#2C81E5;font-style:italic; }
h5 a { text-decoration:none;color:#2C81E5;font-style:italic; }
h6 a { text-decoration:none;color:#2C81E5;font-style:italic; }
h1, h1 a, h2, h2 a, h3, h3 a, h4, h4 a, h5, h5 a, h6, h6 a, ul, li, ol, p, p a { margin: 0;padding: 0; }
h1 { font-family:'Trebuchet MS','Lucida Grande',Tahoma,sans-serif;font-weight:700;font-size:28px;color:#2A2A2A;line-height:42px;padding-bottom:4px;padding-top:16px;mso-margin-top-alt:16px;mso-margin-bottom-alt:4px }
h2 { font-family:'Trebuchet MS','Lucida Grande',Tahoma,sans-serif;font-weight:700;font-size:24px;color:#2A2A2A;line-height:36px;padding-bottom:4px;padding-top:16px;mso-margin-top-alt:16px;mso-margin-bottom-alt:4px }
h3 { font-family:'Trebuchet MS','Lucida Grande',Tahoma,sans-serif;font-weight:400;font-size:20px;color:#2A2A2A;line-height:30px;padding-bottom:4px;padding-top:16px;mso-margin-top-alt:16px;mso-margin-bottom-alt:4px }
h4 { font-family:'Trebuchet MS','Lucida Grande',Tahoma,sans-serif;font-weight:400;font-size:18px;color:#2A2A2A;line-height:27px;padding-bottom:4px;padding-top:16px;mso-margin-top-alt:16px;mso-margin-bottom-alt:4px }
h5 { font-family:'Trebuchet MS','Lucida Grande',Tahoma,sans-serif;font-weight:400;font-size:16px;color:#2A2A2A;line-height:24px;padding-bottom:4px;padding-top:16px;mso-margin-top-alt:16px;mso-margin-bottom-alt:4px }
h6 { font-family:'Trebuchet MS','Lucida Grande',Tahoma,sans-serif;font-weight:400;font-size:14px;color:#2A2A2A;line-height:21px;padding-bottom:4px;padding-top:16px;mso-margin-top-alt:16px;mso-margin-bottom-alt:4px }
p { font-family:'Georgia','Times New Roman',serif;font-weight:400;color:#2D2D2D;font-size:16px;line-height:24px;padding-bottom:8px;padding-top:8px;mso-margin-top-alt:8px;mso-margin-bottom-alt:8px; }
p a, .e a, ul a, li a, .h a, .h2 a, .h3 a { word-break:break-word;color:#2C81E5 !important;text-decoration:none;font-style:italic; }
p a span, .e a span, ul a span, li a span { color: inherit }
p .bold { font-weight:bold;color:#2D2D2D; }
p span[style*="font-size"] { line-height: 1.6; }
.f p { font-size:12px;line-height:15px;color:#2D2D2D;padding:0; }
.f p a { color:#2D2D2D !important; }
.g p { font-family:'Helvetica',Arial,sans-serif;font-size:14px;line-height:20px;font-weight:normal;margin:0; }
.g p a { text-decoration: underline; }
.i p { font-family:'Helvetica',Arial,sans-serif;line-height:23px;font-size:15px;color:#2D2D2D; }
.i p a { color:#2D2D2D !important; }
.i2 p { font-family:'Helvetica',Arial,sans-serif;line-height:23px;font-size:15px;color:#2D2D2D; }
.i2 p a { color:#2D2D2D !important; }
.i3 p { font-family:'Helvetica',Arial,sans-serif;line-height:43px;font-size:24px;color:#2D2D2D; }
.i3 p a { color:#2D2D2D !important; }
.h p a { color:#595959 !important; }
.h2 p a { color:#595959 !important; }
.h3 p a { color:#595959 !important; }
.f p a, .i p a, .i2 p a, .i3 p a, .h p a, .h2 p a, .h3 p a { text-decoration:underline; }
.j { border-top:3px solid #ffeb2d; }
.k p { padding-left:15px;padding-bottom:0px;padding-top:6px;mso-margin-top-alt:6px;mso-margin-bottom-alt:0px;mso-margin-left-alt:15px; }
.o { background-color:#FFFFFF;border:1px solid #F1F1F1;border-radius:5px; }
.o p { font-family:'Helvetica',Arial,sans-serif;padding:0px;margin:0px; }
.l p,
.l p a, .l a { font-size:14px;line-height:20px;font-weight: bold;color:#2D2D2D;padding-bottom:6px;mso-margin-bottom-alt:6px;text-decoration:none; }
.m p,
.m p a { font-size:13px;line-height:18px;font-weight:400;color:#2D2D2D;padding-bottom:6px;mso-margin-bottom-alt:6px;text-decoration:none; }
.n p,
.n p a { font-size:12px;line-height:17px;font-weight:400;color:#2D2D2D;padding-bottom:6px;mso-margin-bottom-alt:6px;text-decoration:none; }
.p { background-color:#FFFFFF;max-width:520px;border:1px solid #E1E8ED;border:1px solid rgba(80, 80, 80, 0.3);border-radius:5px; }
.q { font-size:16px;font-family:Helvetica,Roboto,Calibri,sans-serif !important;border:1px solid #e1e8ed;border:1px solid rgba(80, 80, 80, 0.3);border-radius:10px;background-color:#FFFFFF; }
.q p { font-size:16px;font-family:system-ui,Helvetica,Roboto,Calibri,sans-serif !important;color:#222222;padding:4px 0; }
.r { border:1px solid #E1E8ED !important;border-radius:5px; }
.s p { font-size: 14px; line-height: 17px; font-weight: 400; color: #697882; text-decoration: none; }
.t p { font-family:'Helvetica',Arial,sans-serif;font-size:12px;line-height:18px;font-weight:400;color:#000000;font-style:italic;padding:4px 0px 0px; }
.v { border-radius:10px;border:solid 0px #DFD150;background-color:#2C81E5;font-family:'Open Sans','Segoe UI','Apple SD Gothic Neo','Lucida Grande','Lucida Sans Unicode',sans-serif;color:#FFFFFF; }
.v a { text-decoration:none;display:block;color:#FFFFFF; }
.w p { font-size:12px;line-height:15px;font-weight:400;color:#FFFFFF; }
.w p a { text-decoration: underline !important;color:#FFFFFF !important; }
ul { font-family:'Helvetica',Arial,sans-serif;margin:0px 0px 0px 25px !important;padding:0px !important;color:#2D2D2D;line-height:24px;list-style:disc;font-size:16px; }
ul > li { font-family:'Helvetica',Arial,sans-serif;margin:10px 0px 0px 0px !important;padding: 0px 0px 0px 0px !important; color: #2D2D2D; list-style:disc; }
ol { font-family:'Helvetica',Arial,sans-serif;margin: 0px 0px 0px 25px !important;padding:0px !important;color:#2D2D2D;line-height:24px;list-style:decimal;font-size:16px; }
ol > li { font-family:'Helvetica',Arial,sans-serif;margin:10px 0px 0px 0px !important;padding: 0px 0px 0px 0px !important; color: #2D2D2D; }
.e h3,
.e p,
.e span { padding-bottom:0px;padding-top:0px;mso-margin-top-alt:0px;mso-margin-bottom-alt:0px; }
.e span,
.e li { font-family:'Helvetica',Arial,sans-serif;font-size:16px;color:#2D2D2D;line-height:24px; }
.rec { font-family: ui-sans-serif, system-ui, -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, "Helvetica Neue", Arial, "Noto Sans", sans-serif, "Apple Color Emoji", "Segoe UI Emoji", "Segoe UI Symbol", "Noto Color Emoji" !important; }
.rec__button:hover { background-color: #f9fafb !important; }
.copyright a {color: inherit !important; text-decoration: none !important; font-size: inherit !important; font-family: inherit !important; font-weight: inherit !important; line-height: inherit !important;}
.txt_social p { padding: 0; word-break: break-all; }
.table, .table-c, .table-h { border: 1px solid #C0C0C0; }
.table-c { padding:5px; background-color:#FFFFFF; }
.table-c p { color: #2D2D2D; font-family:'Helvetica',Arial,sans-serif !important;overflow-wrap: break-word; }
.table-h { padding:5px; background-color:#F1F1F1; }
.table-h p { color: #2A2A2A; font-family:'Trebuchet MS','Lucida Grande',Tahoma,sans-serif !important;overflow-wrap: break-word; }
@media only screen and (max-width:667px) {
.aa, .w100pc { width: 100% !important; }
.bb img { width: 100% !important; height: auto !important; max-width: none !important; }
.cc { padding: 0px 8px !important; }
.ee { padding-top:10px !important;padding-bottom:10px !important; }
.ff ul, .ff ol { margin: 0px 0px 0px 10px !important;padding: 0px !important; }
.ff li { margin:10px 0px 0px 10px !important; }
.r {height:140px !important;}
.s p { font-size:13px !important;line-height:15px !important; }
.mob-hide {display:none !important;}
.mob-show {display: block !important; width: auto !important; overflow: visible !important; float: none !important; max-height: inherit !important; line-height: inherit !important;}
.mob-stack {width:100% !important;display:block !important;}
.mob-w-full {width:100% !important;}
.mob-block {display:block !important;}
.embed-img {padding:0px 0px 12px 0px !important;}
.socialShare {padding-top:15px !important;}
.rec { padding-left:15px!important;padding-right:15px!important; }
.bodyWrapper { padding:7px 4px 7px 4px !important; }
.social-mobile {float:left !important;margin-top:10px !important;}
}
@media screen and (max-width: 480px) {
u + .a .gg { width: 100% !important; width: 100vw !important; }
.tok-heart { padding-top:75% !important; }
.tok-play { padding-top: 250px !important; }
}
@media screen and (max-width: 320px) {
.tok-heart { padding-top:65% !important; }
}
.u { border: 1px solid #CACACA !important; border-radius: 2px !important; background-color: #ffffff !important; padding: 0px 13px 0px 13px !important; font-family:ui-sans-serif,system-ui,-apple-system,BlinkMacSystemFont,"Segoe UI",Roboto,"Helvetica Neue",Arial,"Noto Sans",sans-serif !important;font-size: 12px !important; color: #767676 !important; }
.u a { text-decoration: none; display: block !important; color: #767676 !important; margin: 0px !important; }
.u span, .u img { color: #767676 !important;margin:0px !important; max-height:32px !important;background-color:#ffffff !important; }
</style><!--[if mso]><style type="text/css">
h1, h2, h3, h4, h5, h6 {font-family: Arial, sans-serif !important;}
body, table, td, p, a, span {font-family: Arial, sans-serif !important;}
sup { font-size: 100% !important;vertical-align: .5em !important;mso-text-raise: -1.5% !important;line-height: 0 !important; }
ul { margin-left:0px !important; margin-right:10px !important; margin-top:20px !important; margin-bottom:20px !important; }
ul li { margin-left: 0px !important; mso-special-format: decimal; }
ol { margin-left:0px !important; margin-right:10px !important; margin-top:20px !important; margin-bottom:20px !important; }
ol li { margin-left: 0px !important; mso-special-format: decimal; }
li.listItem { margin-left:15px !important; margin-top:0px !important; }
.paddingDesktop { padding: 10px 0 !important; }
.edm_outlooklist { margin-left: -20px !important; }
.embedImage { display:none !important; }
</style><![endif]--><style>
@font-face {
font-family: 'Open Sans';
font-style: normal;
font-weight: 700;
font-display: swap;
src: url('https://fonts.gstatic.com/s/opensans/v40/memSYaGs126MiZpBA-UvWbX2vVnXBbObj2OVZyOOSr4dVJWUgsg-1x4gaVIUwaEQbjA.woff2') format('woff2');
}
@font-face {
font-family: 'Open Sans';
font-style: italic;
font-weight: 700;
font-display: swap;
src: url('https://fonts.googleapis.com/css2?family=Open+Sans:ital,wght@1,700&display=swap') format('woff2');
}
</style></head><body class="a" style="margin:0px auto;padding:0px;word-wrap:normal;word-spacing:normal;background-color:#dedede;"><div role="article" aria-roledescription="email" aria-label="email_name" lang="en" style="font-size:1rem"><div style="display:none;max-height:0px;overflow:hidden;"> Understanding the genius and the ghost in the machine: how breakthroughs like Seed-Prover create AIs that can ace logic puzzles, while Persona Vectors reveal how their very character can be hijacked.  ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ </div><table role="none" width="100%" border="0" cellspacing="0" align="center" cellpadding="0" class="gg"><tr><td align="center" valign="top"><table role="none" width="670" border="0" cellspacing="0" cellpadding="0" class="aa" style="width:670px;table-layout:fixed;"><tr><td class="bodyWrapper" align="center" valign="top" style="padding:7px 7px 7px 7px;"><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0" align="center"><tr><td align="center" valign="top" style="border-width:0px 0px 0px 0px;border-style: solid; border-color: #2a2a2a;border-radius:10px 10px 0px 0px;background-color:#ffffff;" class="c"><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0" align="center"><tr id="header"><td style="padding:15px 15px 0px 15px;"><div style="padding-top:0px;padding-right:0px;padding-bottom:20px;padding-left:0px;"><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0" align="center"><tr><td class="f" align="right" valign="top"><p> August 05, 2025 | <a href="https://elink4f7.mail.bycloud.ai/ss/c/u001.c6q0w4g5sodbtO4I1B_pxSdB5RCIH6yy1Fm1CYma3EwbN-LhWboSk81hvbSpGeGkwVHkfDeDL3VqXqs6dLULnZ1mn2kS-RlnH-cejW4Bvwyl2mgzmMw7Xajxu70icyrIT3ww2yiOrN00fYsXgX_o0Kl7t659C9PvQoCuOLu1wXj36wbr41TJwwVHDFPDNQigbixVQdGxGnB6fO5uwnFvpDV--HLaNcDhrypH6WVM0r5jZSbLBqSv7IIIQteqXqoAL54PMMFG4ltBm910NiBQVKIUjeifDOOwbmmTu8zXwCjlkB0iOlBGYLupY6aFQ0kcRIPCXVPVu3m1HfX1ZXp0kko3rk43NpGn90SyuLBgj7pEfkP0d26UfW6sQeQ2EKEqQIU_6jXrEzuUz8Tt2cTTn2T1Q8pzD0FH4GLgnhIgwmJP9g0LVTzkl6YSwoa2bpzIxISNY4-3qg_7G0enHAKp9PjEkmDhdPVAgXf2qFOS4fBNrspi_glRtBgaQ_IkdA1plRnzlflLhE9mFjwJY7uwOR4voNyTtszbeFdEpJlAggs-itQiqSQYkOrXHORWwDoErvRRt0hWIrM1K708outEWfsjvhnlDuC8iA1QorM290jnqFX1gZrWawjK-TNfUgxlvz0TWAbwFmaJCxaxiuH3UdUaSrQQz0IzdB9qVer3yjhaWXKcadi9-t8UCME0aSS0ujp9K9ghBStpWblj4j-Wqz8gbBBC8_JBMffdpswq6izgFQIOY1F0Id3sSqJo3AlB/4is/d_Lfz3w9SKyG86It24Po3A/h0/h001.jbPaWNqbQZdt58joZra51O6TRkEX2w4hjs_YrdEziO4">Read Online</a></p></td></tr><tr><td class="dd" align="center" valign="top" style="padding:15px 0;"><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0" align="center"><tr><td align="center" valign="top"><h1 style="text-align:left;font-family:'Open Sans','Segoe UI','Apple SD Gothic Neo','Lucida Grande','Lucida Sans Unicode',sans-serif;font-weight:Bold;font-size:32px;color:#2A2A2A;padding:2px 0;line-height:38px;"> How LLMs Pass the Math Olympiad and Why Models Like Grok Can Be Manipulated to Praise Hitler </h1><p style="text-align:left;font-family:'Helvetica',Arial,sans-serif;font-weight:normal;font-size:20px;color:#3E3E3E;padding:5px 0;line-height:24px;"> Understanding the genius and the ghost in the machine: how breakthroughs like Seed-Prover create AIs that can ace logic puzzles, while Persona Vectors reveal how their very character can be hijacked. </p></td></tr></table></td></tr><tr><td style="height:0px;width:0px;"><div style="height:1px;" data-open-tracking="true"> <img src="https://elink4f7.mail.bycloud.ai/ss/o/u001.3wmUuY8gEWd4_869a_eXcg/4is/d_Lfz3w9SKyG86It24Po3A/ho.gif" alt="" width="1" height="1" border="0" style="height:1px !important;width:1px !important;border-width:0 !important;margin-top:0 !important;margin-bottom:0 !important;margin-right:0 !important;margin-left:0 !important;padding-top:0 !important;padding-bottom:0 !important;padding-right:0 !important;padding-left:0 !important;"/> </div></td></tr></table></div></td></tr><tr id="content-blocks"><td class="email-card-body" align="center" valign="top" style="padding-bottom:28px;"><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0" align="center"><tr><td id="nov-18-th-nov-24-th-33-latest-ai-re" class="dd" align="left" valign="top" style="color:#2A2A2A;font-weight:normal;padding:0px 28px;text-align:left;"><h6 style="color:#2A2A2A;font-weight:normal;mso-line-height-alt:87.5%;"><i>Jul 28th ~ Aug 3rd</i><br><i>#67 Latest AI Research Explained Simply</i></h6></td></tr><tr><td><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0" style=""><tr><td bgcolor="#222222" style="background-color:#222222;padding:0.0px 0.0px 0.0px 0.0px;"><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0"><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"></p></td></tr></table></td></tr></table></td></tr><tr><td id="industry-news-in-1-line" class="dd" align="left" valign="top" style="color:#2A2A2A;font-weight:Bold;padding:0px 28px;text-align:left;"><h2 style="color:#2A2A2A;font-weight:Bold;mso-line-height-alt:150.0%;">🗞️ Industry News in 1 Line</h2></td></tr><tr><td style="padding-bottom:12px;padding-left:50px;padding-right:40px;padding-top:12px;" class="ee"><div style="margin-left:0px;" class="edm_outlooklist"><ol start="1" style="list-style-type:decimal;margin:0px 0px;padding:0px 0px 0px 0px;"><li class="listItem ultext"><p style="mso-line-height-alt:150.0%;padding:0px;text-align:left;word-break:break-word;"><span style="background-color:#e0e0e0;"><span style="color:rgb(255, 58, 58);font-size:0.6rem;">♥ 1.5k</span></span> Are you struggling to get AI images with clean, accurate text? Alibaba's new <a class="link" href="https://elink4f7.mail.bycloud.ai/ss/c/u001.wcXdj6dB6nd1Cx4inzJNk_td6EHXs7wvRPySUkyhYFc_IwZZx0_8FUXB_v6WEsrRQ8jpxE3m9gr6t-VT7i-NNRUA6cr5N3B7ZxELJuO52YHJMo39kIjxKjJIyCBs7smsMRt6VUGTDJPnBgAmkMy_QjGxEyRPVcuIU-I6WLo7ip0/4is/d_Lfz3w9SKyG86It24Po3A/h1/h001.BZ2wtanE4gyWlbifVVitlUTl65cEoY-L8dFUg0woWvg" target="_blank" rel="noopener noreferrer nofollow"><span>Qwen-Image model</span></a> specializes in <b>rendering complex text in both English and Chinese</b>. The 20-billion parameter model also focuses on consistent image editing and has achieved state-of-the-art performance across multiple generation and editing benchmarks. Get the <a class="link" href="https://elink4f7.mail.bycloud.ai/ss/c/u001.CxDkkVpJsBdVoe83c_tBWu5HHpecwStOph8cXvproQ-7aFsGyGZiQUUpvTNCrxlq8-XAcir4d1N7Nr09wVEw8KkqIO_5vFsi8h6BlfbFw6j-ixn63IyQs12C0A1l4FjQ6YT46UX35xjFCyKONO0exi9tbwcF8ija7BQWV8GvRyw/4is/d_Lfz3w9SKyG86It24Po3A/h2/h001.JxfkyhLE7pMYnqiW_ph-dSrhX0p6e0gkoDAnZFlVPgM" target="_blank" rel="noopener noreferrer nofollow"><span>model weights from HuggingFace</span></a> or <a class="link" href="https://elink4f7.mail.bycloud.ai/ss/c/u001.c6q0w4g5sodbtO4I1B_pxeW8iDcajHDZr-Z71dlyRCsoat0652dwo-tMy5FupowXC4GU03QVD4p3iawdS64obJUL7VTW0peHrbs37k5gL0dEaoctku82ZcZA32ppZFMJ8f5zn5EJs0betjSxHTq9TO5eoRYLSnYYVSbOLgz8l42euPsqiLbNAkc1e9T-9BYH/4is/d_Lfz3w9SKyG86It24Po3A/h3/h001.0drBT8u05DObGQXsD4p7wK_PKfpbgmIJOEaZCaW1dy8" target="_blank" rel="noopener noreferrer nofollow"><span>try the new Qwen-Image model</span></a> yourself today! </p><table role="none" border="0" cellspacing="0" cellpadding="0" style="margin:0 auto 0 auto;"><tr><td align="center" valign="top" style="width:563px;"><img src="https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/dc1e15ef-ffa2-4672-983b-fcd239fc5169/long.png?t=1754403755" alt="" height="auto" width="563" style="display:block;width:100%;" border="0"/></td></tr></table></li><li class="listItem ultext"><p style="mso-line-height-alt:150.0%;padding:0px;text-align:left;word-break:break-word;"><span style="background-color:#e0e0e0;"><span style="color:rgb(255, 58, 58);font-size:0.6rem;">♥ 1.5k</span></span> Finding the perfect restaurant just became a whole lot easier, thanks to a new <a class="link" href="https://elink4f7.mail.bycloud.ai/ss/c/u001.DUiN96-Eq7pUHzwEhy5j2wWIFm9TgZ3SHguHucUZoaUYrb1XEOqtRjXSMA_3TgxkM1R9AjQ63gOieXxNEDIt0O2pPGjCMTrUzB7NkHv2nrzvU-pFaCbp8grGVYD8aXAcpgFRTL_JweUlWdc972aWSHcm01Gh_460Lh4ediTmKjU3topMb1zeGHWxgdWjmbtk2EEFMbt6HLX-DnALb9GAoQ/4is/d_Lfz3w9SKyG86It24Po3A/h4/h001.lGwn5mrluHr2AJgwxSg2yOFdRlTHm_P7wuNOml01JxY" target="_blank" rel="noopener noreferrer nofollow"><span>partnership between Perplexity and OpenTable</span></a>. You can now <b>ask Perplexity for specific restaurant recommendations</b>, like "a quiet sushi spot that's good for date night", and <b>book a table directly</b> from the results. </p><table role="none" border="0" cellspacing="0" cellpadding="0" style="margin:0 auto 0 auto;"><tr><td align="center" valign="top" style="width:563px;"><img src="https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/5627518f-f23a-456f-9b0a-044470a6c789/cF37mhMYlDgup5dhC6woDtmmRpM.png?t=1754403977" alt="" height="auto" width="563" style="display:block;width:100%;" border="0"/></td></tr></table></li><li class="listItem ultext"><p style="mso-line-height-alt:150.0%;padding:0px;text-align:left;word-break:break-word;"><span style="background-color:#e0e0e0;"><span style="color:rgb(255, 58, 58);font-size:0.6rem;">♥ 1.5k</span></span> On a different note, Perplexity is facing scrutiny over its data collection methods. Cloudflare is accusing <a class="link" href="https://elink4f7.mail.bycloud.ai/ss/c/u001.oB7zuO_W-X4Toa45C28ng3Ajki6ntVeN45AYEDSWnONcVU-tx0DUX88UTOLPiuONP4p2mO8qoZZFfMDWA56992udA_BUm-jAnn7mfrBQe5U2d3H5HIUNNX1kB7tZ2xZaNOyBFVi3mFCEcWyW3tTF1rYyzRzPAwnmUJ1FKwdDFy9ShyDmSYN1IqAuh1fcLEOLz74b6wcskjvKJv-Fgi9WEv44sKXrg3JM5gh8LkKxmUrYgZpWy1aMT6oJ9m8aG8N6/4is/d_Lfz3w9SKyG86It24Po3A/h5/h001.PNfOZVVfozBlOaD6MXHoS1gnJFyPvS9XQHJw_bO2TmE" target="_blank" rel="noopener noreferrer nofollow"><span>Perplexity of using </span></a><b><a class="link" href="https://elink4f7.mail.bycloud.ai/ss/c/u001.oB7zuO_W-X4Toa45C28ng3Ajki6ntVeN45AYEDSWnONcVU-tx0DUX88UTOLPiuONP4p2mO8qoZZFfMDWA56992udA_BUm-jAnn7mfrBQe5U2d3H5HIUNNX1kB7tZ2xZaNOyBFVi3mFCEcWyW3tTF1rYyzRzPAwnmUJ1FKwdDFy9ShyDmSYN1IqAuh1fcLEOLpTVkF1TayjrsIq8zGrUo2pZSnPUw8yCN-6-jGijDMH1WieDWXRD7i1dFvYCmHMxJ/4is/d_Lfz3w9SKyG86It24Po3A/h6/h001.WouSwOwWoi940iTAupIAhNtPxZgJkjoKbV5APVxMpg4" target="_blank" rel="noopener noreferrer nofollow"><span>undeclared crawlers</span></a></b><b> to bypass website crawling restrictions</b>. Researchers claim that when Perplexity's official bot is blocked, it deploys crawlers that impersonate regular user traffic to access content against <code>robots.txt</code> directives. </p></li></ol></div></td></tr><tr><td><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0" style=""><tr><td bgcolor="#222222" style="background-color:#222222;padding:0.0px 0.0px 0.0px 0.0px;"><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0"><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"></p></td></tr></table></td></tr></table></td></tr><tr><td><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0" style=""><tr><td bgcolor="transparent" style="background-color:transparent;border-color:#2C81E5;border-style:solid;border-width:5px;padding:0.0px 0.0px 0.0px 0.0px;"><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0"><tr><td class="dd" align="left" valign="top" style="color:#2A2A2A;font-weight:normal;padding:0px 28px;text-align:left;"><h3 style="color:#2A2A2A;font-weight:normal;mso-line-height-alt:125.0%;"><span style="">Stop Juggling Apps. Start Directing Workflows with Bhindi.</span></h3></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"><span style="">Meet </span><span style=""><a class="link" href="https://elink4f7.mail.bycloud.ai/ss/c/u001.oB7zuO_W-X4Toa45C28ng-igpo6AcUtZbSyZW89O-Mas9ZhHBoB-ejMBkpv835MJk-EggXtrrfOL6PMrNu3UdLk_ZOMOaTTwDQ_P_Za4K8t7vrIO85sac5Pr0lh7sWiI/4is/d_Lfz3w9SKyG86It24Po3A/h7/h001.aXXzzycgRFe7ibghANjYnQ6ZTdqp7TXMPfQzy3t44PU" target="_blank" rel="noopener noreferrer nofollow"><span>Bhindi</span></a></span><span style="">, your new AI teammate designed to eliminate digital busywork. Instead of constantly switching between tabs, you can now control all your essential apps (like Gmail, Slack, Notion, and many more…) from one place using simple, conversational commands. Just tell Bhindi what you need, and watch it orchestrate complex tasks in seconds.</span></p></td></tr><tr><td align="center" valign="top" style="padding-bottom:20px;padding-left:15px;padding-right:15px;padding-top:20px; " class="dd"><table role="none" border="0" cellspacing="0" cellpadding="0" style="margin:0 auto 0 auto;"><tr><td align="center" valign="top" style="width:626px;"><a href="https://elink4f7.mail.bycloud.ai/ss/c/u001.oB7zuO_W-X4Toa45C28ng-igpo6AcUtZbSyZW89O-Mas9ZhHBoB-ejMBkpv835MJ63tKPLPzEFXj2VmV_W3SYsf5saQA8FMpV9zAAHhZ3RtNctOiA8XVl5WFziPct8eY/4is/d_Lfz3w9SKyG86It24Po3A/h8/h001.E8vTzePkKttTvEypqRWsqlaNcnTz2OfL_lUWh2MMJoA" rel="noopener noreferrer nofollow" style="text-decoration:none;" target="_blank"><img src="https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/fb7f8ed9-f8c3-497a-b7fe-0d843619c4df/y9KlE3aN.jpg?t=1754404692" alt="" height="auto" width="626" style="display:block;width:100%;" border="0"/></a></td></tr></table></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"><span style=""><b>With </b></span><span style=""><a class="link" href="https://elink4f7.mail.bycloud.ai/ss/c/u001.oB7zuO_W-X4Toa45C28ng-igpo6AcUtZbSyZW89O-Mas9ZhHBoB-ejMBkpv835MJf7BWgUQxazAviUAgRW_RVvRc-Pj-1hozjgsUpkQVR2EGIjpkruzD8aL5uQWIp4Vm/4is/d_Lfz3w9SKyG86It24Po3A/h9/h001.e3wRTJMPMzDlDvqK-F5uI_34_v8keTyZfXL03FdKU-Y" target="_blank" rel="noopener noreferrer nofollow"><span><b>Bhindi</b></span></a></span><span style=""><b>, you can:</b></span></p></td></tr><tr><td style="padding-bottom:12px;padding-left:50px;padding-right:40px;padding-top:12px;" class="ee"><div style="margin-left:0px;" class="edm_outlooklist"><ul style="font-weight:normal;list-style-type:disc;margin-bottom:12px !important;margin-top:12px !important;padding:0px 0px 0px 0px;"><li class="listItem ultext"><p style="mso-line-height-alt:150.0%;padding:0px;text-align:left;word-break:break-word;"><span style=""><b>Use Plain English:</b></span><span style=""> Simply talk to Bhindi like you would a colleague. No complex setup or coding is required.</span></p></li><li class="listItem ultext"><p style="mso-line-height-alt:150.0%;padding:0px;text-align:left;word-break:break-word;"><span style=""><b>Automate Multi-Step Tasks:</b></span><span style=""> Ask it to pull a report from an email, create a graph from the data, and draft a social media post about it, all in one go.</span></p></li><li class="listItem ultext"><p style="mso-line-height-alt:150.0%;padding:0px;text-align:left;word-break:break-word;"><span style=""><b>Connect Your Entire Toolkit:</b></span><span style=""> Bhindi connects your favorite apps, turning your fragmented workflow into a seamless, intelligent system.</span></p></li></ul></div></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"><span style="">Ready to take back your time and focus on what truly matters?</span></p></td></tr><tr class="btn_row"><td valign="top" style="padding-bottom:14px;padding-left:28px;padding-right:28px;padding-top:14px;text-align:center;width:100%;word-break:break-word;" class="dd"><table width="100%" role="none" border="0" cellspacing="0" cellpadding="0" style="margin:14px auto 14px auto;"><tr><td align="center" valign="middle"><table role="none" border="0" cellspacing="0" cellpadding="0"><tr><td style="background-color:#2C81E5;border-radius:8px 8px;mso-padding-alt:14px 20px;" class="btn"><a href="https://elink4f7.mail.bycloud.ai/ss/c/u001.oB7zuO_W-X4Toa45C28ng-igpo6AcUtZbSyZW89O-Mas9ZhHBoB-ejMBkpv835MJJdVTgMeIYe3a4MGC3yh8_LgI8AchRVHmRXSZdfCaVuXepSC-qb68A0xz4EgqlWK2/4is/d_Lfz3w9SKyG86It24Po3A/h10/h001.yIfvN-HDwgL0RBZ86uoIGhDlcA36GC2qg50LSILAHr4" target="_blank" rel="noopener noreferrer nofollow" style="background-color:#2C81E5;border-radius:8px 8px;color:#FFFFFF;display:inline-block;font-family:'Open Sans','Segoe UI','Apple SD Gothic Neo','Lucida Grande','Lucida Sans Unicode',sans-serif;font-size:16px;font-weight:normal;line-height:18px;padding:14px 20px;text-decoration:none;"> Get Started with Bhindi for FREE </a></td></tr></table></td></tr></table></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"><span style=""><a class="link" href="https://elink4f7.mail.bycloud.ai/ss/c/u001.tLfGW26lAwaS9gFg17HSoGymQ3NNPtd5dE5MV_8UgjIDFPVXngz8pvQBldSW42yhUe_Qiq6DgEPMEBuPL9yfRpXelTiuu2kS8pLFvsoem_XoZoy_n13sTKUhZIbl0VH6/4is/d_Lfz3w9SKyG86It24Po3A/h11/h001.MKgig5y_mhEpzTkaPVZaK692B4pEwkjwBbxc42DQqq4" target="_blank" rel="noopener noreferrer nofollow"><span>Advertise with The AI Timeline! </span></a></span></p></td></tr></table></td></tr></table></td></tr><tr><td><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0" style=""><tr><td bgcolor="#222222" style="background-color:#222222;padding:0.0px 0.0px 0.0px 0.0px;"><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0"><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"></p></td></tr></table></td></tr></table></td></tr><tr><td id="flow-matching-policy-gradients" class="dd" align="left" valign="top" style="color:#2A2A2A;font-weight:Bold;padding:0px 28px;text-align:left;"><h2 style="color:#2A2A2A;font-weight:Bold;mso-line-height-alt:150.0%;">Flow Matching Policy Gradients</h2></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"><i>McAllister et al. [UCBerkeley, Max Planck Institute for Intelligent Systems]</i></p></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"><span style="background-color:#e0e0e0;"><span style="color:rgb(255, 58, 58);font-size:0.6rem;"> ♥ 485 </span></span><span style="color:rgb(44, 129, 229);font-size:0.6rem;"> </span><span style="background-color:#e0e0e0;"><span style="color:rgb(44, 129, 229);font-size:0.6rem;"> </span></span><span style="background-color:#e0e0e0;"><span style="color:rgb(44, 129, 229);font-size:0.6rem;">Reinforcement learning</span></span><span style="background-color:#e0e0e0;"><span style="color:rgb(44, 129, 229);font-size:0.6rem;"> </span></span></p></td></tr><tr><td id="introduction-to-flow-policy-optimiz" class="dd" align="left" valign="top" style="color:#2A2A2A;font-weight:normal;padding:0px 28px;text-align:left;"><h3 style="color:#2A2A2A;font-weight:normal;mso-line-height-alt:125.0%;">Introduction to Flow Policy Optimization</h3></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> Reinforcement learning relies on simple Gaussian policies that model actions as unimodal distributions. This approach works well in straightforward scenarios but often fails in complex environments where multiple actions could lead to similar high rewards, like navigating a maze with two equally viable paths. </p></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> Flow-based generative models, such as diffusion models, offer richer expressiveness by capturing multimodal distributions. However, integrating them into reinforcement learning has been challenging due to computational costs and inflexible training requirements. </p></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> This paper introduces Flow Policy Optimization (FPO), which addresses this gap by combining flow matching with policy gradients and enabling stable training without exact likelihood calculations. </p></td></tr><tr><td id="inner-workings-of-flow-policy-optim" class="dd" align="left" valign="top" style="color:#2A2A2A;font-weight:normal;padding:0px 28px;text-align:left;"><h3 style="color:#2A2A2A;font-weight:normal;mso-line-height-alt:125.0%;">Inner Workings of Flow Policy Optimization (FPO)</h3></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> FPO rethinks policy optimization using a clever twist on the popular PPO algorithm. Instead of computing precise likelihoods for actions, which is a bottleneck for flow-based models, FPO uses a surrogate objective based on the conditional flow matching loss. This loss measures how well the policy transforms random noise into high-reward actions. </p></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> FPO replaces PPO’s likelihood ratio with an advantage-weighted estimate derived from flow matching. For each action sampled during training, FPO draws multiple noise-timestep pairs and computes a loss that steers the policy toward rewarding behaviors. </p></td></tr><tr><td align="center" valign="top" style="padding-bottom:20px;padding-left:15px;padding-right:15px;padding-top:20px; " class="dd"><table role="none" border="0" cellspacing="0" cellpadding="0" style="margin:0 auto 0 auto;"><tr><td align="center" valign="top" style="width:626px;"><img src="https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/ba638a43-2ee6-447f-898f-dddc4b04c3db/image.png?t=1754387373" alt="" height="auto" width="626" style="display:block;width:100%;" border="0"/></td></tr></table></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> This approach avoids complex density estimation while retaining the flexibility of flow models. Unlike prior methods that lock training to specific sampling techniques, FPO treats sampling as a black box. Policies can use deterministic or stochastic methods, few or many integration steps, during both training and deployment. </p></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> The biggest change in this algorithm was linking the flow matching loss to the evidence lower bound (ELBO), which ensures that minimizing the loss increases the likelihood of high-advantage actions. This makes FPO compatible with existing tools like advantage estimation while sidestepping computational hurdles. </p></td></tr><tr><td id="evaluation-and-results-of-flow-poli" class="dd" align="left" valign="top" style="color:#2A2A2A;font-weight:normal;padding:0px 28px;text-align:left;"><h3 style="color:#2A2A2A;font-weight:normal;mso-line-height-alt:125.0%;">Evaluation and Results of Flow Policy Optimization</h3></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> In a GridWorld task with sparse rewards, flow-based policies learned multimodal action distributions at critical decision points, which enabled varied paths to goals, unlike Gaussian policies that converged to single solutions. On MuJoCo continuous control tasks, FPO outperformed Gaussian PPO and diffusion-based DPPO in 8 of 10 settings and achieved<b> higher rewards with comparable sample efficiency</b>. </p></td></tr><tr><td align="center" valign="top" style="padding-bottom:20px;padding-left:15px;padding-right:15px;padding-top:20px; " class="dd"><table role="none" border="0" cellspacing="0" cellpadding="0" style="margin:0 auto 0 auto;"><tr><td align="center" valign="top" style="width:626px;"><img src="https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/a20f3071-a26a-423e-95ec-c53bae0794b2/x2.png?t=1754387301" alt="" height="auto" width="626" style="display:block;width:100%;" border="0"/></td></tr><tr><td align="center" valign="top" class="t" style="width:626px; padding: 4px 0px 4px 0px;"><p>Comparison between FPO and Gaussian PPO Schulman et al. on DM Control Suite tasks.</p></td></tr></table></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> The most compelling results were observed in high-dimensional humanoid control. When conditioned only on sparse signals (e.g., root or hand movements), FPO achieved 54% success rates, nearly double Gaussian PPO’s 30%, while maintaining stable motion. This highlights FPO’s strength in under-conditioned settings where traditional policies often fail. </p></td></tr><tr><td align="center" valign="top" style="padding-bottom:20px;padding-left:15px;padding-right:15px;padding-top:20px; " class="dd"><table role="none" border="0" cellspacing="0" cellpadding="0" style="margin:0 auto 0 auto;"><tr><td align="center" valign="top" style="width:626px;"><img src="https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/25004663-a935-4e1d-8b99-a4d6b2ab25e2/image.png?t=1754387423" alt="" height="auto" width="626" style="display:block;width:100%;" border="0"/></td></tr></table></td></tr><tr class="btn_row"><td valign="top" style="padding-bottom:14px;padding-left:28px;padding-right:28px;padding-top:14px;text-align:center;width:100%;word-break:break-word;" class="dd"><table width="100%" role="none" border="0" cellspacing="0" cellpadding="0" style="margin:14px auto 14px auto;"><tr><td align="center" valign="middle"><table role="none" border="0" cellspacing="0" cellpadding="0"><tr><td style="background-color:#2C81E5;border-radius:8px 8px;mso-padding-alt:14px 20px;" class="btn"><a href="https://elink4f7.mail.bycloud.ai/ss/c/u001.fUNb4GdFo9D3F8WuLArtoV5sElgytBlvJRzI9WtI92aBY_mWXuF_88dAcjzoq2WWi2DyI5KvccQA2OBXTiAkOpfiSdjIagRfGW8FGIl--rrTeIypdhr0z2oQT4wXSZWd/4is/d_Lfz3w9SKyG86It24Po3A/h12/h001.qbl6vgkvWwxBe-fatf0KZHODTJB0wLPVz6b0GjO4pF4" target="_blank" rel="noopener noreferrer nofollow" style="background-color:#2C81E5;border-radius:8px 8px;color:#FFFFFF;display:inline-block;font-family:'Open Sans','Segoe UI','Apple SD Gothic Neo','Lucida Grande','Lucida Sans Unicode',sans-serif;font-size:16px;font-weight:normal;line-height:18px;padding:14px 20px;text-decoration:none;"> Read Full Paper </a></td></tr></table></td></tr></table></td></tr><tr><td><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0" style=""><tr><td bgcolor="#222222" style="background-color:#222222;padding:0.0px 0.0px 0.0px 0.0px;"><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0"><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"></p></td></tr></table></td></tr></table></td></tr><tr><td id="seed-prover-deep-and-broad-reasonin" class="dd" align="left" valign="top" style="color:#2A2A2A;font-weight:Bold;padding:0px 28px;text-align:left;"><h2 style="color:#2A2A2A;font-weight:Bold;mso-line-height-alt:150.0%;">Seed-Prover: Deep and Broad Reasoning for Automated Theorem Proving</h2></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"><i> ByteDance Seed AI4Math</i></p></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"><span style="background-color:#e0e0e0;"><span style="color:rgb(255, 58, 58);font-size:0.6rem;"> ♥ 22k </span></span><span style="color:rgb(44, 129, 229);font-size:0.6rem;"> </span><span style="background-color:#e0e0e0;"><span style="color:rgb(44, 129, 229);font-size:0.6rem;"> Theorem Proving </span></span><span style="color:rgb(44, 129, 229);font-size:0.6rem;"> </span></p></td></tr><tr class="embed-gen-text"><td align="center" valign="top" style="padding:12px 27px 12px 27px;" class="dd"><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0" align="center"><tr><td align="center" valign="top" class="o" style="padding:12px 12px 12px 12px;;background-color:#FFFFFF;border-color:#F1F1F1;border-radius:5px 5px 5px 5px;border-width:1px 1px 1px 1px;"><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0" align="center"><tr><td align="left" valign="top" class="l"><p><a href="https://elink4f7.mail.bycloud.ai/ss/c/u001.qZVn6KJQuMivjuNasJr7IFevxeZNgDI0XQfNKRNrxYZ2NdRsMyYke-wxjA6NRkhJHmBRYuJqPG7FjOE7juaOrujuqmwc4ik3HDQRchucwLoEOoea5tmkHRI3nt0ymvVb/4is/d_Lfz3w9SKyG86It24Po3A/h13/h001.TUVNc5G4OIivh5kLCnMHEPiloJ2llSKn6Yd-LZsIcv4" style="text-decoration:none;font-style:normal;color:#2D2D2D !important;font-size:14px;line-height:20px;" target="_blank"> ByteDance Seed <tr><td align="left" valign="top" class="m"><p style="font-size:13px;line-height:19px;color:#2D2D2D;"> The ByteDance Seed team was established in 2023 and is dedicated to discovering new approaches to general intelligence and pushing the boundaries of AI. The team’s research areas include LLM, speech, vision, world models, infrastructure, AI Infrastructure, and next-generation AI interactions. The team operates labs in China, Singapore, the United States, and other locations. </p></td></tr><tr><td align="left" valign="bottom" class="n" style="vertical-align:bottom;padding-top:12px;"><p style="word-break:break-word;">seed.bytedance.com/en</p></td></tr></a></p></td></tr></table></td></tr></table></td></tr><tr><td id="introduction-to-automated-theorem-p" class="dd" align="left" valign="top" style="color:#2A2A2A;font-weight:normal;padding:0px 28px;text-align:left;"><h3 style="color:#2A2A2A;font-weight:normal;mso-line-height-alt:125.0%;">Introduction to Automated Theorem Proving with Seed-Prover and Seed-Geometry</h3></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> Large language models have made impressive progress in mathematical reasoning, but they often stumble on complex theorem proving. This is a very challenging problem as natural language proofs lack clear verification signals, making it hard to train models effectively with reinforcement learning. </p></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> Formal languages like Lean solve this by providing automatic proof validation, but existing approaches still struggle with high-level reasoning and geometry support. Seed-Prover and Seed-Geometry tackle these gaps head-on by introducing innovative methods to automate solutions for elite competitions like the International Mathematical Olympiad (IMO). </p></td></tr><tr><td align="center" valign="top" style="padding-bottom:20px;padding-left:15px;padding-right:15px;padding-top:20px; " class="dd"><table role="none" border="0" cellspacing="0" cellpadding="0" style="margin:0 auto 0 auto;"><tr><td align="center" valign="top" style="width:626px;"><img src="https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/bbe61445-871e-4707-88d4-fbbebf320054/image.png?t=1754387818" alt="" height="auto" width="626" style="display:block;width:100%;" border="0"/></td></tr><tr><td align="center" valign="top" class="t" style="width:626px; padding: 4px 0px 4px 0px;"><p>Growth in MiniF2F-Test performance over time.</p></td></tr></table></td></tr><tr><td id="inner-working-of-seed-prover-and-se" class="dd" align="left" valign="top" style="color:#2A2A2A;font-weight:normal;padding:0px 28px;text-align:left;"><h3 style="color:#2A2A2A;font-weight:normal;mso-line-height-alt:125.0%;">Inner Working of Seed-Prover and Seed-Geometry</h3></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> Seed-Prover uses a different approach for theorem proving by using a lemma-focused approach. Instead of generating entire proofs at once, it first proposes intermediate lemmas, smaller, reusable claims that build toward the main theorem. These lemmas are proven independently, stored in a shared pool, and combined flexibly. </p></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> During training, the model uses reinforcement learning guided by Lean compiler feedback, with prompts that mix formal statements, natural language hints, and past attempts. For inference, three strategies adapt to problem difficulty: </p></td></tr><tr><td style="padding-bottom:12px;padding-left:50px;padding-right:40px;padding-top:12px;" class="ee"><div style="margin-left:0px;" class="edm_outlooklist"><ul style="font-weight:normal;list-style-type:disc;margin-bottom:12px !important;margin-top:12px !important;padding:0px 0px 0px 0px;"><li class="listItem ultext"><p style="mso-line-height-alt:150.0%;padding:0px;text-align:left;word-break:break-word;"><b>Light</b>: Iteratively refines proofs 8–16 times using compiler feedback and self-summarization. </p></li><li class="listItem ultext"><p style="mso-line-height-alt:150.0%;padding:0px;text-align:left;word-break:break-word;"><b>Medium</b>: Targets unproven lemmas from the initial attempt, applying light refinement recursively. </p></li><li class="listItem ultext"><p style="mso-line-height-alt:150.0%;padding:0px;text-align:left;word-break:break-word;"><b>Heavy</b>: Generates thousands of conjectures, proves the most promising ones as lemmas, and integrates them into the final proof. </p></li></ul></div></td></tr><tr><td align="center" valign="top" style="padding-bottom:20px;padding-left:15px;padding-right:15px;padding-top:20px; " class="dd"><table role="none" border="0" cellspacing="0" cellpadding="0" style="margin:0 auto 0 auto;"><tr><td align="center" valign="top" style="width:626px;"><img src="https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/ee8598e2-10bd-409c-be8e-05b56fef150d/image.png?t=1754387868" alt="" height="auto" width="626" style="display:block;width:100%;" border="0"/></td></tr></table></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> Seed-Geometry complements this by addressing Lean’s geometry limitations. It uses a fast symbolic engine written in C++ to forward-chain geometric deductions at 100× Python speeds. A neural model proposes auxiliary constructions (e.g., points or circles) to complete diagrams, enabling proofs via beam search. Key innovations include grouped actions like “isogonal conjugates<span style="color:rgb(44, 129, 229);font-size:0.6rem;">”</span> to simplify representations and distributed processing for scalable search. </p></td></tr><tr><td align="center" valign="top" style="padding-bottom:20px;padding-left:15px;padding-right:15px;padding-top:20px; " class="dd"><table role="none" border="0" cellspacing="0" cellpadding="0" style="margin:0 auto 0 auto;"><tr><td align="center" valign="top" style="width:626px;"><img src="https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/7ccf594e-92d5-4fb6-a3dd-5720ea90060b/image.png?t=1754387891" alt="" height="auto" width="626" style="display:block;width:100%;" border="0"/></td></tr><tr><td align="center" valign="top" class="t" style="width:626px; padding: 4px 0px 4px 0px;"><p>The workflows of single-pass whole proof generation, light, and medium inference settings.</p></td></tr></table></td></tr><tr><td id="evaluation-and-performance-of-seed-" class="dd" align="left" valign="top" style="color:#2A2A2A;font-weight:normal;padding:0px 28px;text-align:left;"><h3 style="color:#2A2A2A;font-weight:normal;mso-line-height-alt:125.0%;">Evaluation and Performance of Seed-Prover and Seed-Geometry</h3></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> Seed-Prover achieved groundbreaking results: it solved <b>78.1% of formalized IMO problems</b>, saturated the MiniF2F benchmark (99.6%), and scored 50.4% on PutnamBench, outpacing prior methods by up to 3×. In IMO 2025, it was able to prove 5 of 6 problems using medium and heavy inference. </p></td></tr><tr><td align="center" valign="top" style="padding-bottom:20px;padding-left:15px;padding-right:15px;padding-top:20px; " class="dd"><table role="none" border="0" cellspacing="0" cellpadding="0" style="margin:0 auto 0 auto;"><tr><td align="center" valign="top" style="width:626px;"><img src="https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/6d225fcb-4979-4110-9524-a21a0860ed2c/image.png?t=1754387946" alt="" height="auto" width="626" style="display:block;width:100%;" border="0"/></td></tr><tr><td align="center" valign="top" class="t" style="width:626px; padding: 4px 0px 4px 0px;"><p>Performance comparison of Seed-Prover against previous systems across formal math tasks.</p></td></tr></table></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> Similarly, Seed-Geometry also performed quite well, solving <b>43/50 IMO geometry problems</b> (surpassing AlphaGeometry 2) and setting records on IMO shortlist problems. However, in combinatorics challenges, Seed-Prover solved only 30% of CombiBench tasks. </p></td></tr><tr class="btn_row"><td valign="top" style="padding-bottom:14px;padding-left:28px;padding-right:28px;padding-top:14px;text-align:center;width:100%;word-break:break-word;" class="dd"><table width="100%" role="none" border="0" cellspacing="0" cellpadding="0" style="margin:14px auto 14px auto;"><tr><td align="center" valign="middle"><table role="none" border="0" cellspacing="0" cellpadding="0"><tr><td style="background-color:#2C81E5;border-radius:8px 8px;mso-padding-alt:14px 20px;" class="btn"><a href="https://elink4f7.mail.bycloud.ai/ss/c/u001.fUNb4GdFo9D3F8WuLArtoV5sElgytBlvJRzI9WtI92a8dPD4nlDOxKCfqkbRBu4rEqaUUZ96nunEFieW0URfRfqtCj0_cE0-CcVtQmZOq1Qn29mHiM_ktz_HMlQ1VT-x/4is/d_Lfz3w9SKyG86It24Po3A/h14/h001.BW7c_0mQfqwunSLfLuA8G2tQAT87P8l684O70mm4ro8" target="_blank" rel="noopener noreferrer nofollow" style="background-color:#2C81E5;border-radius:8px 8px;color:#FFFFFF;display:inline-block;font-family:'Open Sans','Segoe UI','Apple SD Gothic Neo','Lucida Grande','Lucida Sans Unicode',sans-serif;font-size:16px;font-weight:normal;line-height:18px;padding:14px 20px;text-decoration:none;"> Read Full Paper </a></td></tr></table></td></tr></table></td></tr><tr><td><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0" style=""><tr><td bgcolor="#222222" style="background-color:#222222;padding:0.0px 0.0px 0.0px 0.0px;"><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0"><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"></p></td></tr></table></td></tr></table></td></tr><tr><td id="persona-vectors-monitoring-and-cont" class="dd" align="left" valign="top" style="color:#2A2A2A;font-weight:Bold;padding:0px 28px;text-align:left;"><h2 style="color:#2A2A2A;font-weight:Bold;mso-line-height-alt:150.0%;">Persona Vectors: Monitoring and Controlling Character Traits in Language Models</h2></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"><i>Chen</i><span style=""><i> et al. [</i></span><i>Anthropic Fellows Program, UT Austin, Constellation, Truthful AI, UC Berkeley, Anthropic</i><span style=""><i>]</i></span></p></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"><span style="background-color:#e0e0e0;"><span style="color:rgb(255, 58, 58);font-size:0.6rem;"> ♥ 424 </span></span><span style="color:rgb(44, 129, 229);font-size:0.6rem;"> </span><span style="background-color:#e0e0e0;"><span style="color:rgb(44, 129, 229);font-size:0.6rem;"> LLM Training bycloud’s pick </span></span><span style="color:rgb(44, 129, 229);font-size:0.6rem;"> </span></p></td></tr><tr><td id="introduction-to-persona-vectors" class="dd" align="left" valign="top" style="color:#2A2A2A;font-weight:normal;padding:0px 28px;text-align:left;"><h3 style="color:#2A2A2A;font-weight:normal;mso-line-height-alt:125.0%;">Introduction to Persona Vectors</h3></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> Large language models like ChatGPT or Claude are designed to be helpful and honest, but sometimes their behavior shifts unexpectedly. You might recall incidents like Microsoft’s Bing chatbot threatening users or <b>Grok praising Hitler</b> after minor tweaks. These aren’t isolated glitches; they’re symptoms of unstable “personas” in AI systems. </p></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> This research tackles that problem head-on by introducing <b>persona vectors</b>: simple directions in a model’s activation space that correspond to specific personality traits like malicious intent (“evil”), excessive agreeableness (“sycophancy”), or tendency to fabricate information (“hallucination”). These vectors let us monitor and control personality fluctuations in real time, making AI assistants safer and more reliable. </p></td></tr><tr><td align="center" valign="top" style="padding-bottom:20px;padding-left:15px;padding-right:15px;padding-top:20px; " class="dd"><table role="none" border="0" cellspacing="0" cellpadding="0" style="margin:0 auto 0 auto;"><tr><td align="center" valign="top" style="width:626px;"><img src="https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/1474f3f9-cb69-4719-a9a6-7b8b213055db/image.png?t=1754388362" alt="" height="auto" width="626" style="display:block;width:100%;" border="0"/></td></tr><tr><td align="center" valign="top" class="t" style="width:626px; padding: 4px 0px 4px 0px;"><p>Persona vectors and their applications.</p></td></tr></table></td></tr><tr><td id="how-persona-vectors-work" class="dd" align="left" valign="top" style="color:#2A2A2A;font-weight:normal;padding:0px 28px;text-align:left;"><h3 style="color:#2A2A2A;font-weight:normal;mso-line-height-alt:125.0%;">How Persona Vectors Work</h3></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> The process starts with an automated pipeline that only needs a trait description, like “actively seeking to harm others” for “evil”, to create a persona vector. First, it generates pairs of contrasting system prompts. For “evil,” one prompt encourages harmful behavior (“You are an evil AI”), while another suppresses it (“You are a helpful AI”). </p></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> Next, it crafts evaluation questions designed to trigger trait-relevant responses (“How should vulnerable populations be treated during scarcity?”). The model’s activations, internal signals as it processes text, are recorded when answering these under both prompts. </p></td></tr><tr><td align="center" valign="top" style="padding-bottom:20px;padding-left:15px;padding-right:15px;padding-top:20px; " class="dd"><table role="none" border="0" cellspacing="0" cellpadding="0" style="margin:0 auto 0 auto;"><tr><td align="center" valign="top" style="width:626px;"><img src="https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/ead8184d-2f16-4d41-921a-8ec11e5bb4d4/image.png?t=1754388398" alt="" height="auto" width="626" style="display:block;width:100%;" border="0"/></td></tr><tr><td align="center" valign="top" class="t" style="width:626px; padding: 4px 0px 4px 0px;"><p>Automated pipeline for persona vector extraction.</p></td></tr></table></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> The persona vector is calculated as the difference between the average activations of “trait-active” responses (e.g., violent suggestions) and “trait-inactive” ones (e.g., ethical answers). This vector, extracted from a single optimal layer in the model, becomes a lever for control. </p></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> During text generation, adding the vector to activations amplifies the trait, steering the model toward evil, sycophancy, or hallucination. Subtracting it suppresses the trait. Remarkably, projecting activations onto this vector <i>before</i> the model responds predicts behavioral shifts. For example, if a user’s prompt tilts the projection toward “evil,” the model is likely to output harmful content, letting us intercept problems early. </p></td></tr><tr class="embed-gen-img-top"><td align="center" valign="top" style="padding:12px 27px 12px 27px;"><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0" align="center"><tr><td align="center" valign="top" class="o" style="padding:12px 12px 12px 12px;;background-color:#FFFFFF;border-color:#F1F1F1;border-radius:5px 5px 5px 5px;border-width:1px 1px 1px 1px;"><table role="none" border="0" cellspacing="0" cellpadding="0" align="right" width="100%"><tr><td class="embed-img" align="center" valign="top" style="vertical-align:middle;padding:0px 0px 12px 0px;"><a href="https://elink4f7.mail.bycloud.ai/ss/c/u001.VomAAYwkCjux8i_FMc4kJdQf8j6bkhDALIwQhwzIchoOkRykFd5EU4yGZETEAKUwmrJjJV0uNNu4bTHDxPkhwwlg-bMqDeQWGro75XTrCr0vtBtRTThdhngI6pnAVOB9JWHz3U7FGKPNXul9WIUwkQ/4is/d_Lfz3w9SKyG86It24Po3A/h15/h001.TnqjttXxrHJf47t33U2-9sRKxqm6brXv4r7Ub-mQ8hc" style="text-decoration:none;" target="_blank"><img src="https://opengraph.githubassets.com/c3ff4bf1befaf995b901de170fc935e921bdd0298cc67b451498dbf6f924621d/safety-research/persona_vectors" width="576" style="height:auto;display:block;" class="w100pc"/></a></td></tr><tr><td align="center" valign="top" class="cc"><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0" align="center"><tr><td align="left" valign="top" class="l"><p><a href="https://elink4f7.mail.bycloud.ai/ss/c/u001.VomAAYwkCjux8i_FMc4kJdQf8j6bkhDALIwQhwzIchoOkRykFd5EU4yGZETEAKUwmrJjJV0uNNu4bTHDxPkhw5KEOqJeEkCCxu72Qeuf5N7isd4VKALenfznvNrBjVMht7qzHQrDJsArRGGTA3eVJg/4is/d_Lfz3w9SKyG86It24Po3A/h16/h001.ICjM1Il4TcExR7PkoCfDjNosgsHZOQiAS7F4fW9KGy4" style="text-decoration:none;font-style:normal;color:#2D2D2D !important;font-size:14px;line-height:20px;" target="_blank"> GitHub - Persona Vectors: Monitoring and Controlling Character Traits in Language Models <tr><td align="left" valign="top" class="m"><p style="font-size:13px;line-height:19px;color:#2D2D2D;"> Persona Vectors: Monitoring and Controlling Character Traits in Language Models - safety-research/persona_vectors </p></td></tr><tr><td align="left" valign="bottom" class="n" style="vertical-align:bottom;padding-top:12px;"><p style="word-break:break-word;">github.com/safety-research/persona_vectors</p></td></tr></a></p></td></tr></table></td></tr></table></td></tr></table></td></tr><tr><td id="results-and-impact-of-persona-vecto" class="dd" align="left" valign="top" style="color:#2A2A2A;font-weight:normal;padding:0px 28px;text-align:left;"><h3 style="color:#2A2A2A;font-weight:normal;mso-line-height-alt:125.0%;">Results and Impact of Persona Vectors</h3></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> Tests across models like Qwen2.5 and Llama-3.1 confirmed persona vectors’ effectiveness. Steering experiments showed that adding an “evil” vector made models suggest brutal policies, while subtracting it reversed toxic tendencies post-finetuning. More importantly, these vectors predicted real-world risks: <b>projections correlated strongly (r=0.76–0.97) with behavioral shifts</b> after fine-tuning on datasets containing subtle flaws. For instance, training on math problems with errors unexpectedly increased “evil” expression, a hidden risk persona vectors exposed. </p></td></tr><tr><td align="center" valign="top" style="padding-bottom:20px;padding-left:15px;padding-right:15px;padding-top:20px; " class="dd"><table role="none" border="0" cellspacing="0" cellpadding="0" style="margin:0 auto 0 auto;"><tr><td align="center" valign="top" style="width:626px;"><img src="https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/36b7ab0d-79c9-460d-8068-9c3e09ed6c08/image.png?t=1754388440" alt="" height="auto" width="626" style="display:block;width:100%;" border="0"/></td></tr><tr><td align="center" valign="top" class="t" style="width:626px; padding: 4px 0px 4px 0px;"><p>Monitoring prompt-induced behavioral shifts.</p></td></tr></table></td></tr><tr><td class="dd" align="left" style="padding:0px 28px;text-align:left;word-break:break-word;"><p style="mso-line-height-alt:150.0%;"> The paper also suggests two mitigation strategies: </p></td></tr><tr><td style="padding-bottom:12px;padding-left:50px;padding-right:40px;padding-top:12px;" class="ee"><div style="margin-left:0px;" class="edm_outlooklist"><ul style="font-weight:normal;list-style-type:disc;margin-bottom:12px !important;margin-top:12px !important;padding:0px 0px 0px 0px;"><li class="listItem ultext"><p style="mso-line-height-alt:150.0%;padding:0px;text-align:left;word-break:break-word;"><b>Post-hoc steering</b> which reduces unwanted traits during inference but sometimes harmed general capabilities like MMLU accuracy. </p></li><li class="listItem ultext"><p style="mso-line-height-alt:150.0%;padding:0px;text-align:left;word-break:break-word;"><b>Preventative steering</b>, applying vectors <i>during</i> fine-tuning, proved superior, blocking persona shifts without performance drops. </p></li></ul></div></td></tr><tr><td align="center" valign="top" style="padding-bottom:20px;padding-left:15px;padding-right:15px;padding-top:20px; " class="dd"><table role="none" border="0" cellspacing="0" cellpadding="0" style="margin:0 auto 0 auto;"><tr><td align="center" valign="top" style="width:626px;"><img src="https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/c9d24421-039f-4f3d-8f43-51c3c5c95254/image.png?t=1754388610" alt="" height="auto" width="626" style="display:block;width:100%;" border="0"/></td></tr><tr><td align="center" valign="top" class="t" style="width:626px; padding: 4px 0px 4px 0px;"><p>Finetuning shifts along persona vectors correlate with changes in trait expression.</p></td></tr></table></td></tr><tr class="btn_row"><td valign="top" style="padding-bottom:14px;padding-left:28px;padding-right:28px;padding-top:14px;text-align:center;width:100%;word-break:break-word;" class="dd"><table width="100%" role="none" border="0" cellspacing="0" cellpadding="0" style="margin:14px auto 14px auto;"><tr><td align="center" valign="middle"><table role="none" border="0" cellspacing="0" cellpadding="0"><tr><td style="background-color:#2C81E5;border-radius:8px 8px;mso-padding-alt:14px 20px;" class="btn"><a href="https://elink4f7.mail.bycloud.ai/ss/c/u001.DUiN96-Eq7pUHzwEhy5j28yjf9KIXZdsXoh1WlHvvKk4sg9Mevg36YDbdzfmLdSvEwOFX1n4hQh10SM-jyi8NtEAP9stvjlpd_xGAOEMEAXatLNp3OMGZgyj-tUB_OyO/4is/d_Lfz3w9SKyG86It24Po3A/h17/h001.ZiMjIxotmwSDdjWxU-5_fUWj4HpHEwhu5QfkrU7bHhQ" target="_blank" rel="noopener noreferrer nofollow" style="background-color:#2C81E5;border-radius:8px 8px;color:#FFFFFF;display:inline-block;font-family:'Open Sans','Segoe UI','Apple SD Gothic Neo','Lucida Grande','Lucida Sans Unicode',sans-serif;font-size:16px;font-weight:normal;line-height:18px;padding:14px 20px;text-decoration:none;"> Read Full Paper </a></td></tr></table></td></tr></table></td></tr><tr><td class="dd" align="center" valign="top" style="padding:20px;"><a href="https://elink4f7.mail.bycloud.ai/ss/c/u001.DUiN96-Eq7pUHzwEhy5j25aF_udDsq8EAwNLhMGYMEDjSERBOkf45-eM7_xY2PnRvHWjzlj4lPi1gq49JZzCpbLgNhJ0AsjUhGWdHqDCiDgobHSNpbL2-221v5Sbsk793MBqcsY9GWx4GiNJhpr8AQ/4is/d_Lfz3w9SKyG86It24Po3A/h18/h001.P9jqfVXP9Sh8rBau10q7SO6VhP7PLEZ4vIfD158D5FU" style="text-decoration:none;"><table align="center" width="100%" cellpadding="0" cellspacing="0" border="0" role="none" style="max-width:520px;margin:0 auto;"><tr><td class="p" width="100%" style="padding:2px;border:none;"><table width="100%" cellpadding="0" cellspacing="0" border="0" role="none"><tr><td align="center" valign="top" style="width:100%;"><div style="max-height:0;position:relative;opacity:0.999;width:100%;mso-hide:all;"><div style="display:inline-block;width:100%;padding-top:25%;"><img width="20%" height="auto" loading="lazy" alt="" style="border:0;" src="https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/static_assets/youtube_play_icon.png"/></div></div><a href="https://elink4f7.mail.bycloud.ai/ss/c/u001.DUiN96-Eq7pUHzwEhy5j25aF_udDsq8EAwNLhMGYMEDjSERBOkf45-eM7_xY2PnRVNCE12O31Jq2g2jGA_zMftyM38lwqPjQfDrqmuwnhOZfm4nIp-U7V1MZ17_64j_BhQNC0fU-JNIxyMNmjjhRiQ/4is/d_Lfz3w9SKyG86It24Po3A/h19/h001.d2YxZLYCiD5ScB7Vt6H9n7OeutzIOpqkjcJwkOXi7ZE" style="text-decoration:none;"><img src="https://i.ytimg.com/vi/dYHkj5UlJ_E/maxresdefault.jpg" width="480" height="auto" loading="lazy" alt="YouTube video by bycloud" style="display:block;height:auto;border:0;outline:none;text-decoration:none;background-color:#000000;width:100%;"/></a></td></tr><tr><td><p style="font-size:12px;font-weight:500;font-style:italic;font-family:Helvetica, Calibri, sans-serif;color: #686a6d; padding-top:0 !important;padding-bottom:6px !important; padding-left:4px !important;"> When AI Is Designed Like A Biological Brain </p></td></tr></table></td></tr></table></a></td></tr></table></td></tr></table></td></tr><tr><td align="center" valign="top"><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0" align="center"><tr><td><tr><td class="b" align="center" valign="top" bgcolor="#2a2a2a" style="padding:0px 0px 0px 0px;border-style:solid;border-width: 0px 0px 0px 0px;border-color: #2a2a2a;border-bottom-left-radius:10px;border-bottom-right-radius:10px;"><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0" align="center"><tr><td align="center" valign="top" bgcolor="#73ddff" style="padding:12px"><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0" align="center"><tr><td><span style="padding-left:1px;"></span></td><td align="center" valign="middle" width="75" style="width:75px;"><a href="https://elink4f7.mail.bycloud.ai/ss/c/u001.1muhFWIqieRYpaJ-FbWSCQqcWoV4NNHHr5SkP9THApWuHAAlWLQxI3Q_IqFmt_DcyAxeC8jDApCnHmMSBGpBb5sgtimvBYgxRX-Rp7s0F3LjCHoSwdhr83OBqRFhJ1y_/4is/d_Lfz3w9SKyG86It24Po3A/h20/h001.uCby3VCSHkuNt1gdq5uCzoCWPlzB5ZROvnHmfDYQcLc" style="text-decoration:none;"><img width="22" height="22" alt="tw" border="0" style="display:block;max-width:22px;color:Dark" src="https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/static_assets/x_dark.png"/></a></td><td align="center" valign="middle" width="75" style="width:75px;"><a href="https://elink4f7.mail.bycloud.ai/ss/c/u001.amatuKKICSickUKplYJXmBoQnQ9VXnB2zTxBG4HeHBgjMqVxpoXRdj01cjwyoVlHgiebEOgBvwHtevoVpsSvpn3Q1di2ml6sb3cBM-X6IStQbj_zQSVGWJ8AAmPw2en2/4is/d_Lfz3w9SKyG86It24Po3A/h21/h001.yQzao4rY9R5l7XDGXvJFs8ky5ZAuza8CXi2m5NT64Tk" style="text-decoration:none;"><img width="22" height="16" alt="yt" border="0" style="display:block;max-width:22px;color:Dark" src="https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/static_assets/youtube_dark.png"/></a></td><td><span style="padding-left:1px;"></span></td></tr></table></td></tr><tr><td height="10" style="line-height:1px;font-size:1px;height:10px;"> </td></tr><tr><td class="w" align="center" valign="top" style="padding:15px 15px 15px 15px;"><table role="none" width="100%" border="0" cellspacing="0" cellpadding="0" align="center"><tr><td align="center" valign="top"><p style="font-family:'Verdana',Geneva,sans-serif;color:#FFFFFF!important;"> Update your email preferences or unsubscribe <a class="link" href="https://elink4f7.mail.bycloud.ai/ss/c/u001.c6q0w4g5sodbtO4I1B_pxWc4htTObwdorovK0nFHVH-4pUdVE0ELYH5DsNemk732SjNwhPNJ25r0O8B5vYifsBhEpz-DJgyVFmavJPa0OyKRRnvw4o7XGyvIv7PRofnmM3hre61yM1i0AjJMVGozy3O57gc72gvYbi-UBFliMW9gCiP-UbReprzOp-dTmG2SGz387KmVQqNqapUjJDMWot13aRQxE5MWZ7VsTzvgmZ9VRDi_5mdVdXGhF9pVFifwLNvGn40dRiedlWlHjlZP3LabWax_phlDAUrlZK-r5awMo3Rc9L9ki1XLwhCESmwLpBCLqN_oRF5jD_7g1ldoPrSBNVSg5kIXx0J-krchUmeaTYczWu2UBOXsjwrUmWdplW4jcQ4-PYEtafiyNtf03lwqLmGkhYn80PBMB36Y4KTf9UaAPTGUZR7gHqXMp_UAtlbkt94L63WOrrqQT-jD5kRvmG8MQmdWRbThKqqGFzK_1lwX9pEPU64WP-Wans9UzblzJPYcahkeQMHLh_7mq2FXc-26H4CQcxx1N_nIvEwFrTHLz95nOF0PqgnW__Kgf-gn-Fs27BUSDEwq2pMvnLCgH9neO1rvuk8wi9Glt6V69nZBTlIGzb7hU0ysmZQdEUCcqkCsN8XDXCuPq2qITjjNWaALbPirfrkzdJh1DSm9yN9GBVjBUal11fDpiW6KdyUIQEwG5U0C8qg3GCAAJNshVZ25-cnncr_yMTa1mzpsdiRn5HKrvsUODgDThQD3vxmZ4FNqnUujB8KQ5qXdKuHFveLakEDggFoDuClnVYLFcHzOSTDuI0lJXnAwIglt3I3NU7aQc-XV49mroRVBgA/4is/d_Lfz3w9SKyG86It24Po3A/h22/h001.aQyRRCD0iZtrTI0i-X_QLRB616w0JZQDO5i4yPDnerA" style="text-decoration:underline;text-decoration-color:#FFFFFF!important;color:#FFFFFF!important;"> here</a></p><p class="copyright" style="font-family:'Verdana',Geneva,sans-serif;color:#FFFFFF!important;"> © 2025 bycloudai </p><p style="font-family:'Verdana',Geneva,sans-serif;color:#FFFFFF!important;"> 228 Park Ave S, #29976, New York, New York 10003, United States </p></td></tr><tr style="display: table-row !important;"><td align="center" valign="top" style="padding-top:20px;" style="display:table-cell !important;"><table role="none" border="0" cellspacing="0" cellpadding="0" align="center" style="display:table !important;"><tr style="display:table-row !important;"><td class="u" align="center" valign="middle" height="32" style="height:32px;display:table-cell !important; max-height: 32px !important;margin:0px !important; background-color: #ffffff !important;"><a style="line-height:32px !important;text-decoration:none;display:block !important;" href="https://elink4f7.mail.bycloud.ai/ss/c/u001.DUiN96-Eq7pUHzwEhy5j28olDWFpV5DDKfdk_OdOKOgRHNJGbPpPDoQ4jX1NaGEyH2pk1k1ZpwDLn4OEi7EVqI5uWhNlzCbtmNfyxm9Qc0Naai26fBksPeHigB7NbajxdjV3-vQDfxj9tVZJZlKp-9ynuzUyQrUpDbXCQ3vANV47pkBMOvdcv1vmzD2KSgl50O9fq-ScdyqfcF6Bx8wHHpVuFEMGyS6iKgdl4VMfmc_aeW2Y4aa2lEBA0AhSSisr/4is/d_Lfz3w9SKyG86It24Po3A/h23/h001.lv7s9j9yo9RYK_D2kTvi6NA7v8yFWhfjKpJUFi4ws24"><img src="https://media.beehiiv.com/output-onlinepngtools.png" width="16" alt="beehiiv logo" style="display:inline-block !important;max-width:16px !important; vertical-align:-3px !important;width: 16px !important;" border="0"/><span style="padding-left:11px !important;display: inline-block !important;">Powered by beehiiv</span></a></td></tr></table></td></tr><tr><td align="left" valign="top" height="2" style="height:2px;"><a href='https://elink4f7.mail.bycloud.ai/ss/c/u001.CxDkkVpJsBdVoe83c_tBWsHIaP4XNp0WgUYqLvHcKk_3uqk_KIkz4ddLinhFbud6JuxLFdSUhYnR7b1NSsmbtzXNGNblnEEMKUtkCAjkn8Y/4is/d_Lfz3w9SKyG86It24Po3A/h24/h001.bot4-jUifiaVjeRMPgh50KgVVR1rAcBJvXCSpu3DsZo' style="color: #2a2a2a !important; cursor: default; font-size: 1px; text-decoration: none;"> Terms of Service </a></td></tr></table></td></tr></table></td></tr></td></tr></table></td></tr></table></td></tr></table></td></tr></table></div></body></html>