BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:America/Chicago
X-LIC-LOCATION:America/Chicago
BEGIN:DAYLIGHT
TZOFFSETFROM:-0600
TZOFFSETTO:-0500
TZNAME:CDT
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0500
TZOFFSETTO:-0600
TZNAME:CST
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260202T201804Z
LOCATION:241
DTSTART;TZID=America/Chicago:20251116T091000
DTEND;TZID=America/Chicago:20251116T094000
UID:submissions.supercomputing.org_SC25_sess209_ws_fgai121@linklings.com
SUMMARY:Evaluation of Test-Time Compute Constraints on Safety and Skill La
 rge Reasoning Models
DESCRIPTION:Adarsha Balaji, Le Chen, Rajeev Thakur, Franck Cappello, and S
 andeep Madireddy (Argonne National Laboratory (ANL))\n\nTest-time compute 
 scaling has demonstrated the ability to improve the performance of reasoni
 ng language models by generating longer chain-of-thought (CoT) sequences. 
 However, this increase in performance comes with a significant increase in
  computation cost. In this work, we investigate two compute constraint str
 ategies: (1) reasoning length constraint and (2) model quantization, and s
 tudy their impact on the safety performance of reasoning models. Specifica
 lly, we explore two approaches to apply compute constraints to reasoning m
 odels: (1) fine-tuning reasoning models using a length-controlled policy o
 ptimization (LCPO) based reinforcement learning method to satisfy a user-d
 efined CoT reasoning length, and (2) applying quantization to maximize the
  generation of CoT sequences within a user-defined compute constraint. Fur
 thermore, we study the trade-off between the computational efficiency and 
 the safety of the model.\n\nRecording: Livestreamed, Recorded\n\nRegistrat
 ion Category: Technical Program Reg Pass, Workshop Reg Pass\n\nSession Cha
 irs: Charlie Catlett (Argonne National Laboratory (ANL), University of Chi
 cago); Javier Aula-Blasco (Barcelona Supercomputing Center (BSC)); Valerie
  Taylor (Argonne National Laboratory (ANL), University of Chicago); Laura 
 Morselli (CINECA); and Rio Yokota (Institute of Science Tokyo)\n\n
END:VEVENT
END:VCALENDAR