<div dir="ltr"><span id="gmail-docs-internal-guid-de84a13d-7fff-0f53-5f3a-1aa46972366c" style="color:rgb(0,0,0)"><p style="line-height:1.38;margin-top:12pt;margin-bottom:12pt"><span style="font-family:Arial,sans-serif;font-size:11pt;white-space:pre-wrap">Actualizamos la información del próximo seminario: habrá un cambio en el tema de la charla, pero se mantiene el mismo orador.</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"> Miércoles 24 de septiembre, 13:00 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"> Adrià Garriga-Alonso – Research Scientist, </span><a href="https://www.far.ai/" style="text-decoration:none"><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">FAR AI</span><span style="font-size:11pt;font-family:Arial,sans-serif;color:rgb(0,0,0);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></a><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">Reverse-engineering a neural network that plans: a mesa-optimizer model organism</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/XNDf9uskcRoZ6koW6" style="text-decoration:none"><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/XNDf9uskcRoZ6koW6</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-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">We partially reverse-engineer a convolutional recurrent neural network (RNN) trained to play the puzzle game Sokoban with model-free reinforcement learning. Prior work found that this network solves more levels with more test-time compute. Our analysis reveals several mechanisms analogous to components of classic bidirectional search. For each square, the RNN represents its plan in the activations of channels associated with specific directions. These state-action activations are analogous to a value function - their magnitudes determine when to backtrack and which plan branch survives pruning. Specialized kernels extend these activations (containing plan and value) forward and backward to create paths, forming a transition model. The algorithm is also unlike classical search in some ways. State representation is not unified; instead, the network considers each box separately. Each layer has its own plan representation and value function, increasing search depth. Far from being inscrutable, the mechanisms leveraging test-time compute learned in this network by model-free training can be understood in familiar terms.</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><br></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/2506.10138" style="text-decoration:none"><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/2506.10138</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">http://scholarship.aisafety.ar/</a></span></p></span></div><br><div class="gmail_quote gmail_quote_container"><div dir="ltr" class="gmail_attr">El lun, 22 sept 2025 a las 12:26, 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"><span id="m_1791568302436882870gmail-docs-internal-guid-cbf77756-7fff-6368-1f62-a585869894b3" style="color:rgb(0,0,0)"><p dir="ltr" style="line-height:1.38;margin-top:12pt;margin-bottom:12pt"><span style="font-family:Arial,sans-serif;font-size:11pt;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><br></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"> Miércoles 24 de septiembre, 13:00 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"> Adrià Garriga-Alonso – Research Scientist, </span><a href="https://www.far.ai/" 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">FAR AI</span><span style="font-size:11pt;font-family:Arial,sans-serif;color:rgb(0,0,0);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></a><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">Among Us: A Sandbox for Measuring and Detecting Agentic Deception</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/XNDf9uskcRoZ6koW6" 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/XNDf9uskcRoZ6koW6</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-variant-ligatures:normal;font-variant-alternates:normal;font-variant-numeric:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">Prior studies on deception in language-based AI agents typically assess whether the agent produces a false statement about a topic, or makes a binary choice prompted by a goal, rather than allowing open-ended deceptive behavior to emerge in pursuit of a longer-term goal. To fix this, we introduce Among Us, a sandbox social deception game where LLM-agents exhibit long-term, open-ended deception as a consequence of the game objectives. While most benchmarks saturate quickly, Among Us can be expected to last much longer, because it is a multi-player game far from equilibrium. Using the sandbox, we evaluate 18 proprietary and open-weight LLMs and uncover a general trend: models trained with RL are comparatively much better at producing deception than detecting it. We evaluate the effectiveness of methods to detect lying and deception: logistic regression on the activations and sparse autoencoders (SAEs). We find that probes trained on a dataset of ``pretend you're a dishonest model: …'' generalize extremely well out-of-distribution, consistently obtaining AUROCs over 95% even when evaluated just on the deceptive statement, without the chain of thought. We also find two SAE features that work well at deception detection but are unable to steer the model to lie less. We hope our open-sourced sandbox, game logs, and probes serve to anticipate and mitigate deceptive behavior and capabilities in language-based agents</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><br></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/2504.04072" 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/2504.04072</span></a></p><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><a href="http://scholarship.aisafety.ar/?utm_source=chatgpt.com" 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">http://scholarship.aisafety.ar/</span></a></span><br></div>
</blockquote></div>