BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//wp-events-plugin.com//7.2.3//EN
TZID:America/Los_Angeles
X-WR-TIMEZONE:America/Los_Angeles
BEGIN:VEVENT
UID:0-1673@hmc.edu
DTSTART;TZID=America/Los_Angeles:20250905T110000
DTEND;TZID=America/Los_Angeles:20250905T121500
DTSTAMP:20250904T005820Z
URL:https://www.hmc.edu/calendar/events/cs-colloquium-how-do-llms-reason-s
 ubbarao-kambhampati/
SUMMARY:CS Colloquium: “(How) Do LLMs Reason?” Subbarao Kambhampati
DESCRIPTION:“(How) Do LLMs Reason?”\nLarge Language Models\, auto-regre
 ssively trained on the digital footprints of humanity\, have shown impress
 ive abilities in generating coherent text completions for a vast variety o
 f prompts. While they excelled from the beginning in producing completions
  in appropriate style\, factuality and reasoning/planning abilities remain
 ed their Achilles heel (premature claims notwithstanding). More recently a
  breed of approaches dubbed “reasoning models” (LRMs). These approache
 s leverage two broad and largely independent ideas: (i) test-time inferenc
 e—which involves getting the base LLMs do more work than simply providin
 g the most likely completion\, including using them in generate and test a
 pproaches such as LLM-Modulo (that pair LLM generation with a bank of veri
 fiers) and (ii) post-training methods—which go beyond simple auto-regres
 sive training on web corpora by collecting\, filtering and training on der
 ivational traces (that are often anthropomorphically referred to as “cha
 ins of thought” and “reasoning traces”)\, and modifying the base LLM
  with it using supervised finetuning or reinforcement learning methods. Th
 eir success on benchmarks notwithstanding\, there are significant question
 s and misunderstandings about these methods–including whether they can p
 rovide correctness guarantees\, whether they do adaptive computation\, whe
 ther the intermediate tokens they generate can be viewed as reasoning trac
 es in any meaningful sense\, and whether they are costly Rube Goldberg rea
 soning machines that incrementally compile verifier signal into the genera
 tor or truly the start of a golden era of general purpose System 1+2 AI sy
 stems. Drawing from our ongoing work in planning\, I will present a broad 
 perspective on these approaches and their promise and limitations.\nSpeake
 r\nSubbarao Kambhampati is a professor of computer science at Arizona Stat
 e University. Kambhampati studies fundamental problems in planning and dec
 ision making\, motivated in particular by the challenges of human-aware AI
  systems. He is a fellow of Association for the Advancement of Artificial 
 Intelligence\, American Association for the Advancement of Science\, and A
 ssociation for Computing machinery\, and a recent recipient of the AAAI Pa
 trick H. Winston Outstanding Educator award. He served as the president of
  the Association for the Advancement of Artificial Intelligence\, a truste
 e of the International Joint Conference on Artificial Intelligence\, the c
 hair of AAAS Section T (Information\, Communication and Computation)\, and
  a founding board member of Partnership on AI. Kambhampati’s research as
  well as his views on the progress and societal impacts of AI have been fe
 atured in multiple national and international media outlets. He can be fol
 lowed on Twitter @rao2z.
ATTACH;FMTTYPE=image/jpeg:https://www.hmc.edu/calendar/wp-content/uploads/
 sites/39/2025/08/CS-colloq-speaker-Sept.jpg
CATEGORIES:General Feed,TCCS Feed: Symposium
LOCATION:Shanahan Center\, 320 E. Foothill Blvd.\, Claremont\, CA\, 91711\,
  United States
X-APPLE-STRUCTURED-LOCATION;VALUE=URI;X-ADDRESS=320 E. Foothill Blvd.\, Cla
 remont\, CA\, 91711\, United States;X-APPLE-RADIUS=100;X-TITLE=Shanahan Ce
 nter:geo:0,0
END:VEVENT
BEGIN:VTIMEZONE
TZID:America/Los_Angeles
X-LIC-LOCATION:America/Los_Angeles
BEGIN:DAYLIGHT
DTSTART:20250309T030000
TZOFFSETFROM:-0800
TZOFFSETTO:-0700
TZNAME:PDT
END:DAYLIGHT
END:VTIMEZONE
END:VCALENDAR