<div dir="ltr"><div>Corrijo fecha: Viernes 19 de septiembre, 09:00 hs (ARG).</div><br><div class="gmail_quote gmail_quote_container"><div dir="ltr" class="gmail_attr">El vie, 12 sept 2025 a las 15:39, Agustín Martinez Suñé (<<a href="mailto:agusmartinez92@gmail.com">agusmartinez92@gmail.com</a>>) escribió:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div dir="ltr"><span id="m_-1684557672578225134m_8209265489153951116gmail-docs-internal-guid-5fa1d324-7fff-0dfb-48f7-f0ebb6ae2df1" style="color:rgb(0,0,0)"><p dir="ltr" style="line-height:1.38;margin-top:12pt;margin-bottom:12pt"><span style="font-size:11pt;font-family:Arial,sans-serif;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">Desde el Programa de Becas AISAR en AI Safety tenemos el placer de invitarlos a la próxima charla de nuestro seminario online, con la participación de investigadores del área.</span></p><p dir="ltr" style="line-height:1.38;margin-top:12pt;margin-bottom:12pt"><span style="font-size:11pt;font-family:Arial,sans-serif;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">📌 </span><span style="font-size:11pt;font-family:Arial,sans-serif;font-weight:700;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">Fecha y hora:</span><span style="font-size:11pt;font-family:Arial,sans-serif;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap"> Viernes 19 de septiembre, 09:00 hs (ARG).</span><span style="font-size:11pt;font-family:Arial,sans-serif;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap"><br></span><span style="font-size:11pt;font-family:Arial,sans-serif;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">🎤 </span><span style="font-size:11pt;font-family:Arial,sans-serif;font-weight:700;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">Orador:</span><span style="font-size:11pt;font-family:Arial,sans-serif;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap"> Paul C. Bogdan – </span><span style="font-size:11pt;font-family:Arial,sans-serif;font-style:italic;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">Postdoctoral Researcher, Duke University</span><span style="font-size:11pt;font-family:Arial,sans-serif;font-style:italic;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap"><br></span><span style="font-size:11pt;font-family:Arial,sans-serif;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">📖 </span><span style="font-size:11pt;font-family:Arial,sans-serif;font-weight:700;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">Título:</span><span style="font-size:11pt;font-family:Arial,sans-serif;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap"> </span><span style="font-size:11pt;font-family:Arial,sans-serif;font-style:italic;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">Thought Anchors: Which LLM Reasoning Steps Matter?</span></p><p dir="ltr" style="line-height:1.38;margin-top:12pt;margin-bottom:12pt"><span style="font-size:11pt;font-family:Arial,sans-serif;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">👉 </span><span style="font-size:11pt;font-family:Arial,sans-serif;font-weight:700;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">Inscripción:</span><span style="font-size:11pt;font-family:Arial,sans-serif;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap"> Para asistir a la charla, por favor indicá tu nombre en el siguiente formulario (No es necesario que completes este formulario si ya indicaste "Quiero que me avisen por correo electrónico cuando haya nuevas charlas de AISAR" en un formulario previo): </span><a href="https://forms.gle/wAyCczqeAH7WmwjXA" style="text-decoration:none" target="_blank"><span style="font-size:11pt;font-family:Arial,sans-serif;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;text-decoration:underline;vertical-align:baseline;white-space:pre-wrap">https://forms.gle/wAyCczqeAH7WmwjXA</span></a></p><p dir="ltr" style="line-height:1.38;margin-top:12pt;margin-bottom:12pt"><span style="font-size:11pt;font-family:Arial,sans-serif;font-weight:700;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">Abstract:</span><span style="font-size:11pt;font-family:Arial,sans-serif;font-weight:700;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap"><br></span><span style="font-size:11pt;font-family:Arial,sans-serif;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">Reasoning large language models have recently achieved state-of-the-art performance in many fields. However, their long-form chain-of-thought reasoning creates interpretability challenges as each generated token depends on all previous ones, making the computation harder to decompose. We argue that analyzing reasoning traces at the sentence level is a promising approach to understanding reasoning processes. We present three complementary attribution methods: (1) a black-box method measuring each sentence's counterfactual importance by comparing final answers across 100 rollouts conditioned on the model generating that sentence or one with a different meaning; (2) a white-box method of aggregating attention patterns between pairs of sentences, which identified "broadcasting" sentences that receive disproportionate attention from all future sentences via "receiver" attention heads; (3) a causal attribution method measuring logical connections between sentences by suppressing attention toward one sentence and measuring the effect on each future sentence's tokens. Each method provides evidence for the existence of thought anchors, reasoning steps that have outsized importance and that disproportionately influence the subsequent reasoning process. These thought anchors are typically planning or backtracking sentences. We provide an open-source tool (this http URL) for visualizing the outputs of our methods, and present a case study showing converging patterns across methods that map how a model performs multi-step reasoning. The consistency across methods demonstrates the potential of sentence-level analysis for a deeper understanding of reasoning models.</span></p><p dir="ltr" style="line-height:1.38;margin-top:12pt;margin-bottom:12pt"><span style="font-size:11pt;font-family:Arial,sans-serif;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">Encontrá el paper acá: </span><a href="https://arxiv.org/abs/2506.19143" style="text-decoration:none" target="_blank"><span style="font-size:11pt;font-family:Arial,sans-serif;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;text-decoration:underline;vertical-align:baseline;white-space:pre-wrap">https://arxiv.org/abs/2506.19143</span></a></p><p dir="ltr" style="line-height:1.38;margin-top:12pt;margin-bottom:12pt"><span style="font-size:11pt;font-family:Arial,sans-serif;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">Equipo AISAR</span><span style="font-size:11pt;font-family:Arial,sans-serif;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap"><br></span><span style="text-decoration:underline;font-size:11pt;font-family:Arial,sans-serif;font-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap"><a href="http://scholarship.aisafety.ar/?utm_source=chatgpt.com" style="text-decoration:none" target="_blank">http://scholarship.aisafety.ar/</a></span></p></span></div>
</div>
</blockquote></div></div>