BEGIN:VCALENDAR
VERSION:2.0
CALSCALE:GREGORIAN
PRODID:UW-Madison-Physics-Events
BEGIN:VEVENT
SEQUENCE:0
UID:UW-Physics-Event-9665
DTSTART:20260428T140000Z
DTEND:20260428T160000Z
DTSTAMP:20260425T003807Z
LAST-MODIFIED:20260421T204018Z
LOCATION:3290 Chamberlin
SUMMARY:Large Language Models for Theoretical Physics and Cosmology: B
 enchmarking Reasoning\, Scaling Inference\, and Discovering Scientific
  Algorithms\, Preliminary Exam\, Tianyi Li\, Physics PhD Graduate Stud
 ent
DESCRIPTION:This presentation explores the application of Large Langua
 ge Models (LLMs) in theoretical physics and computational cosmology ac
 ross three key frontiers: evaluation\, inference optimization\, and al
 gorithmic discovery. First\, we introduce the Theoretical Physics Benc
 hmark (TPBench) to assess LLM reasoning\, demonstrating that while fou
 ndational models are advancing\, research-level physics remains a crit
 ical bottleneck. To address these reasoning limits\, we investigate te
 st-time scaling techniques. We propose a novel symbolic weak-verifier 
 that exploits the intrinsic mathematical structure of physics problems
 \, significantly outperforming standard scaling methods. Finally\, we 
 introduce MadEvolve\, an evolutionary optimization framework that tran
 sitions LLMs from solving established problems to discovering novel sc
 ientific methods. By autonomously generating and refining code\, MadEv
 olve yields improved algorithms for complex cosmological tasks such as
  initial condition reconstruction and 21cm foreground mitigation. Toge
 ther\, these works outline a concrete pathway for leveraging LLMs to a
 ccelerate autonomous discovery in physics.
URL:https://www.physics.wisc.edu/events/?id=9665
END:VEVENT
END:VCALENDAR