BEGIN:VCALENDAR
VERSION:2.0
PRODID:OpenCms 21.0.19
BEGIN:VTIMEZONE
TZID:Europe/Berlin
X-LIC-LOCATION:Europe/Berlin
BEGIN:DAYLIGHT
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
DTSTART:19700329T020000
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=3
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
TZNAME:CET
DTSTART:19701025T030000
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=10
END:STANDARD
END:VTIMEZONE				
BEGIN:VEVENT
DTSTAMP:20260626T104535
UID:6624e959-713b-11f1-a79a-000e0c3db68b
SUMMARY:Talk by David Holzmüller
DESCRIPTION:Dr. David Holzmüller will give a presentation in room 8.539, PWR 57 (math/physics building) to which you are warmly invited. Note that this room has cool air conditioning! The presentation will take approximately one hour.You can also attend online at https://unistuttgart.webex.com/meet/max.schoelpple\nTitle: TabICLv2: A better, faster, scalable, and open tabular foundation model\nAbstract:Tabular foundation models, such as TabPFNv2 and TabICL, have recently dethroned gradient-boosted trees at the top of predictive benchmarks, demonstrating the value of in-context learning for tabular data. We introduce TabICLv2, a new state-of-the-art foundation model for regression and classification built on three pillars: (1) a novel synthetic data generation engine designed for high pretraining diversity; (2) various architectural innovations, including a new scalable softmax in attention improving generalization to larger datasets without prohibitive long-sequence pretraining; and (3) optimized pretraining protocols, notably replacing AdamW with the Muon optimizer. On the TabArena and TALENT benchmarks, TabICLv2 without any tuning surpasses the performance of the current state of the art, RealTabPFN-2.5 (hyperparameter-tuned, ensembled, and fine-tuned on real data). With only moderate pretraining compute, TabICLv2 generalizes effectively to million-scale datasets under 50GB GPU memory while being markedly faster than RealTabPFN-2.5. We provide extensive ablation studies to quantify these contributions and commit to open research by first releasing inference code and model weights at https://github.com/soda-inria/tabicl , with synthetic data engine and pretraining code to follow. 
DTSTART;TZID=Europe/Berlin:20260626T113000
URL;VALUE=URI:https://www.isa.uni-stuttgart.de/aktuelles/veranstaltung/Talk-by-David-Holzmueller/
END:VEVENT
END:VCALENDAR