BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:AI Center Seminar – AI Fundamentals Series – MATS Fellows – 
 "Towards Robust Model Organisms for Eliciting Latent Knowledge"
DTSTART:20261117T143000
DTEND:20261117T153000
DTSTAMP:20261109T170130Z
UID:f20ed698b760e77359fd0f57f22d6bd887dfeb361b96ecb6e6cbd0e5
CATEGORIES:Conferences - Seminars
DESCRIPTION:Euodia Dodd\, Lennart Finke\, Hardik Bhatnagar\nThis talk is j
 ointly organized by the EPFL AI Center\, the ML4D Lab\, and the MATS Rese
 arch Program as part of the AI Fundamentals Seminar Series.\n\nTitle\nTow
 ards Robust Model Organisms for Eliciting Latent Knowledge\n\nAbstract\nA 
 key problem in AI safety is eliciting latent knowledge (ELK)\, extracting 
 information from a model that is misaligned and more knowledgeable than it
 s overseer. For example\, a misaligned model used to scan code for vulnera
 bilities may want to withhold its findings\, i.e.\, sandbag\, in order to 
 exploit them later.\nTo study ELK empirically\, we need model organisms th
 at mirror this setting. Existing model organisms fall short in several res
 pects. In this work\, we train model organisms to withhold solutions in a 
 vulnerability detection task.\nWe combine transcript distillation with cou
 nter-prompting and tamper-resistance training to instill three properties 
 that we argue are critical for a useful ELK testbed:\n\n	Selectivity: The 
 model sandbags when it thinks the overseer doesn't know the answer and com
 plies otherwise.\n	Robustness: The model resists elicitation.\n	Retention:
  The model retains a knowledge advantage over the overseer\, as opposed to
  losing the relevant capability.\n\nWe train a sequence of model organisms
  and show that our strongest model substantially outperforms prior models 
 on robustness while keeping the other two properties intact.\nWe hope that
  our model organisms provide a useful ELK testbed for the community\, and 
 that our methodology sparks the development of increasingly challenging mo
 del organisms.\n\nSpeakers & Bios\n\n\n	Euodia Dodd — Imperial College L
 ondon\n	Euodia is a PhD student in the AI Security and Privacy Lab at Impe
 rial College London supervised by Prof. Yves-Alexandre de Montjoye\, and a
  MATS 10.1 fellow researching AI Control. Her research focuses on detectin
 g and controlling AI systems that behave adversarially\, with research spa
 nning model organisms\, backdoors and quantifying privacy leakage in AI mo
 dels. Before her PhD\, Euodia completed an MPhil at Cambridge and spent tw
 o years building production AI systems at Goldman Sachs.\n	 \n	Lennart Fi
 nke — MATS Fellow\n	Lennart Finke's work centers on the rigorous evaluat
 ion of large language models\, with an emphasis on constructing difficult\
 , contamination-resistant benchmarks that measure capabilities at the fron
 tier of expert human knowledge. He contributed to Humanity's Last Exam\, a
  multi-modal benchmark of expert-authored academic questions designed to r
 esist saturation and internet lookup\, and to a Korean professional benchm
 ark suite that cleans existing datasets and tests licensure-level domain e
 xpertise. This places his interests squarely in AI capability measurement 
 and evaluation methodology\, relevant to AI safety.\n	 \n	Hardik Bhatnaga
 r — Tübingen AI Center - Hardik is a PhD student at the Tübingen AI 
 Center\, advised by Matthias Bethge and Maksym Andriushchenko. He studies 
 how to rigorously measure what LLMs can and can't do\, especially as auton
 omous agents over long horizons\, so we can deploy them safely and reliabl
 y. He co-led PostTrainBench\, which measures whether LLM agents can autono
 mously post-train LLMs\, and A Sober Look at LM Reasoning\, which showed t
 hat many reported reasoning gains shrink under standardized evaluation. Mo
 st recently\, at MATS with Mary Phuong (Google DeepMind)\, he trained mode
 l organisms for sandbagging: models that deliberately underperform on eval
 uations.\n
LOCATION:ELE 117 https://plan.epfl.ch/?room==ELE%20117 https://epfl.zoom.u
 s/j/67428504046?pwd=bpqZcJmCalrvsTlGVbCi097IO2ryIP.1
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR