BEGIN:VCALENDAR VERSION:2.0 PRODID:-//UC Irvine//CML Seminars//EN CALSCALE:GREGORIAN METHOD:PUBLISH X-WR-CALNAME:Foundations of Generative Discovery Beyond the Data with Flow and Diffusion Models X-WR-TIMEZONE:America/Los_Angeles BEGIN:VTIMEZONE TZID:America/Los_Angeles BEGIN:DAYLIGHT TZOFFSETFROM:-0800 TZOFFSETTO:-0700 TZNAME:PDT DTSTART:20070311T020000 RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU END:DAYLIGHT BEGIN:STANDARD TZOFFSETFROM:-0700 TZOFFSETTO:-0800 TZNAME:PST DTSTART:20071104T020000 RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU END:STANDARD END:VTIMEZONE BEGIN:VEVENT UID:2026-06-10-riccardo-de-santi@cml.ics.uci.edu DTSTAMP:20260610T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20260610T130000 DTEND;TZID=America/Los_Angeles:20260610T140000 SUMMARY:[CML Seminar] Riccardo De Santi: Foundations of Generative Discover y Beyond the Data with Flow and Diffusion Models LOCATION:Donald Bren Hall 4011 DESCRIPTION:Riccardo De Santi\, PhD Student\, ETH AI Center\n\nTitle: Found ations of Generative Discovery Beyond the Data with Flow and Diffusion Mod els\n\nAbstract: Recent progress in flow and diffusion models has made gen erative models powerful priors over complex scientific design spaces\, whi le reward-guided adaptation offers a practical way to steer them toward de sired properties. This talk asks what is needed to turn such steering into discovery. I will first discuss tail-aware reward guidance: rather than m aximizing average reward\, one may deliberately sacrifice expected reward to concentrate probability on rare\, high-value samples in the top tail of the reward distribution. I will then argue that discovery also requires d ebiasing the generative model itself. In the natural sciences\, available data are local\, limited\, and biased by prior discoveries and measurement processes\, so distribution matching can hide valid low-probability modes . I will present Flow Density Control (FDC) as a framework for distributio nal fine-tuning of flow and diffusion models\, allowing to perform tasks i ncluding entropy-driven mode discovery\, risk-sensitive adaptation\, and e xperimental design. Finally\, I will introduce mathematical foundations fo r out-of-distribution flow modeling through generable-set expansion rather than standard distribution matching. And present Active Flow Expansion (A ctFlow)\, a synthetic pre-training method that uses verifier feedback and active exploration in learned flow representations to expand coverage over valid molecular\, peptide\, and protein space\, while enjoying first-of-t heir-kind statistical guarantees for out-of-distribution generative modeli ng.\n\nhttps://cml.ics.uci.edu/seminars/2026-06-10-riccardo-de-santi X-ALT-DESC;FMTTYPE=text/html:Riccardo De Santi\, PhD Stu dent\, ETH AI Center

Title: Foundations of Generative Discov ery Beyond the Data with Flow and Diffusion Models

Abstract: Recent progress in flow and diffusion models has made generative models p owerful priors over complex scientific design spaces\, while reward-guided adaptation offers a practical way to steer them toward desired properties . This talk asks what is needed to turn such steering into discovery. I wi ll first discuss tail-aware reward guidance: rather than maximizing averag e reward\, one may deliberately sacrifice expected reward to concentrate p robability on rare\, high-value samples in the top tail of the reward dist ribution. I will then argue that discovery also requires debiasing the gen erative model itself. In the natural sciences\, available data are local\, limited\, and biased by prior discoveries and measurement processes\, so distribution matching can hide valid low-probability modes. I will present Flow Density Control (FDC) as a framework for distributional fine-tuning of flow and diffusion models\, allowing to perform tasks including entropy -driven mode discovery\, risk-sensitive adaptation\, and experimental desi gn. Finally\, I will introduce mathematical foundations for out-of-distrib ution flow modeling through generable-set expansion rather than standard d istribution matching. And present Active Flow Expansion (ActFlow)\, a synt hetic pre-training method that uses verifier feedback and active explorati on in learned flow representations to expand coverage over valid molecular \, peptide\, and protein space\, while enjoying first-of-their-kind statis tical guarantees for out-of-distribution generative modeling.

https:/ /cml.ics.uci.edu/seminars/2026-06-10-riccardo-de-santi URL:https://cml.ics.uci.edu/seminars/2026-06-10-riccardo-de-santi END:VEVENT END:VCALENDAR