BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:America/Los_Angeles
X-LIC-LOCATION:America/Los_Angeles
BEGIN:DAYLIGHT
TZOFFSETFROM:-0800
TZOFFSETTO:-0700
TZNAME:PDT
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0700
TZOFFSETTO:-0800
TZNAME:PST
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260730T152640Z
LOCATION:Exhibit Hall
DTSTART;TZID=America/Los_Angeles:20260728T174700
DTEND;TZID=America/Los_Angeles:20260728T174700
UID:dac_DAC 2026_sess306_LBR080@linklings.com
SUMMARY:Late Breaking Results: Fisher-Guided Selective Error Reconstructio
 n for Quantized LLMs
DESCRIPTION:Seonha Ryu (DGIST), Il Hong Suh (coga-robotics), and Yeseong K
 im (DGIST)\n\nWe propose FOCUS, a hardware-aware PTQ recovery framework fo
 r low-bit LLM inference. FOCUS uses Fisher information to select approxima
 tely 1.5% structurally critical row-column intersections and optimizes the
 ir corrections via ridge regression to match output error, enabling a smal
 l set of parameters to compensate global quantization error. Unlike fixed-
 rank SVD updates that distribute correction capacity uniformly, FOCUS conc
 entrates precision on error-critical locations while remaining lightweight
 . It supports latency-hidden CPU-GPU co-execution, where sparse compensati
 on is offloaded to the CPU. Experiments show that FOCUS improves perplexit
 y and downstream accuracy under both 2-bit and 4-bit PTQ, achieving up to 
 29% higher token throughput on the Llama-3.2-3B model.\n\nTrack: Student\n
 \n
END:VEVENT
END:VCALENDAR