BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Département de mathématiques et applications - ECPv6.2.2//NONSGML v1.0//EN
CALSCALE:GREGORIAN
METHOD:PUBLISH
X-WR-CALNAME:Département de mathématiques et applications
X-ORIGINAL-URL:https://www.math.ens.psl.eu
X-WR-CALDESC:évènements pour Département de mathématiques et applications
REFRESH-INTERVAL;VALUE=DURATION:PT1H
X-Robots-Tag:noindex
X-PUBLISHED-TTL:PT1H
BEGIN:VTIMEZONE
TZID:Europe/Paris
BEGIN:DAYLIGHT
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
DTSTART:20260329T010000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
TZNAME:CET
DTSTART:20261025T010000
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTART;TZID=Europe/Paris:20260108T120000
DTEND;TZID=Europe/Paris:20260108T130000
DTSTAMP:20260404T074437
CREATED:20260105T105704Z
LAST-MODIFIED:20260105T105751Z
UID:20657-1767873600-1767877200@www.math.ens.psl.eu
SUMMARY:Nathan Srebro: Learning to Answer from Correct Demonstrations
DESCRIPTION:Generative AI is increasingly presented as a potential substitute for humans\, including as human research subjects in various disciplines. Yet there is no scientific consensus on how closely these in-silico clones could represent their human counterparts. While some defend the use of these “synthetic users\,” others point towards the biases in the responses provided by the LLMs. Through an experiment using survey questionnaires\, we demonstrate that these latter critics are right to be wary of using generative AI to emulate respondents\, but probably not for the right reason. Our results i) confirm that to date\, models cannot replace research subjects for opinion or attitudinal research; ii) that they display a strong bias on each question (reaching only a small region of social space); and iii) that this bias varies randomly from one question to the other (reaching a different region every time). Besides the two existing competing theses (“representativity” and “social bias”)\, we propose a third one\, which we call call “machine bias”. We detail this term and explore its consequences\, for LLM research but also for studies on social biases. \n\n\n\nWe study the problem of learning to generate an answer (or completion) to a question (or prompt)\, where there could be multiple correct answers\, any one of which is acceptable at test time. Learning is based on demonstrations of some correct answer to each trainingquestion\, as in Supervised Fine Tuning (SFT).  Current standard practice focuses on maximum likelihood (ie log loss minimization) approaches\, but we argue that likelihood-maximization methods can fail even in simple settings.  Instead\, we view the problem as apprenticeship learning (i.e.\, imitation learning) in contextual bandits\, with offline demonstrations from some expert (optimal\, or very good) policy\, and suggest alternative simple approaches with strong guarantees. \n\n\n\nJoint work with Nirmit Joshi\, Gene Li\, Siddharth Bhandari\, Shiva Kasiviswanathan\, and Cong Ma \n\n\n\n  \n\n\n\nThese seminars are being made possible through the support of the CFM-ENS Chair « Modèles et Sciences des Données ». \n\n\n\nThe organizers: Giulio Biroli\, Alex Cayco Gajic\, Bruno Loureiro\, Stéphane Mallat\, Gabriel Peyré.
URL:https://www.math.ens.psl.eu/evenement/nathan-srebro-learning-to-answer-from-correct-demonstrations/
LOCATION:Amphi Jean Jaurès\, 45 rue d'Ulm\, PARIS\, 75005\, France
CATEGORIES:Séminaire Data de l’ENS
END:VEVENT
END:VCALENDAR