BEGIN:VCALENDAR VERSION:2.0 PRODID:IEEE vTools.Events//EN CALSCALE:GREGORIAN BEGIN:VTIMEZONE TZID:America/Los_Angeles BEGIN:DAYLIGHT DTSTART:20260308T030000 TZOFFSETFROM:-0800 TZOFFSETTO:-0700 RRULE:FREQ=YEARLY;BYDAY=2SU;BYMONTH=3 TZNAME:PDT END:DAYLIGHT BEGIN:STANDARD DTSTART:20261101T010000 TZOFFSETFROM:-0700 TZOFFSETTO:-0800 RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=11 TZNAME:PST END:STANDARD END:VTIMEZONE BEGIN:VEVENT DTSTAMP:20260618T062248Z UID:08F29111-1750-419D-AD81-45B3957D09BD DTSTART;TZID=America/Los_Angeles:20260626T120000 DTEND;TZID=America/Los_Angeles:20260626T130000 DESCRIPTION:Audio compression is often proposed to improve the efficiency o f multimodal large language models\, but its impact on downstream task per formance remains underexplored. This talk examines how semantic neural aud io codecs behave under token reduction constraints\, using cross-modal att ention as a signal to discard frames with low semantic content. On audio q uestion-answering benchmarks\, attention-guided frame selection removes 10 –30% of frames while matching baseline accuracy and answer consistency\, and identifies a critical compression threshold (keep ratio ~0.7) below w hich performance degrades sharply. The talk also discusses an "answer cons istency paradox" where models remain highly self-consistent (>98%) even as accuracy degrades and what this decoupling of consistency from correctnes s means for evaluating compressed multimodal systems in low-resource deplo yments.\n\nSpeaker(s): Prerana\n\nVirtual: https://events.vtools.ieee.org/ m/563360 LOCATION:Virtual: https://events.vtools.ieee.org/m/563360 ORGANIZER:vicky.h.lu@gmail.com SEQUENCE:26 SUMMARY:Attention-Guided Audio Compression for Multimodal LLMs URL;VALUE=URI:https://events.vtools.ieee.org/m/563360 X-ALT-DESC:Description: <br /><p class="MsoNormal">Audio compression is oft en proposed to improve the efficiency of multimodal large language models\ , but its impact on downstream task performance remains underexplored. Thi s talk examines how semantic neural audio codecs behave under token reduct ion constraints\, using cross-modal attention as a signal to discard frame s with low semantic content. On audio question-answering benchmarks\, atte ntion-guided frame selection removes 10&ndash\;30% of frames while matchin g baseline accuracy and answer consistency\, and identifies a critical com pression threshold (keep ratio ~0.7) below which performance degrades shar ply. The talk also discusses an "answer consistency paradox" where models remain highly self-consistent (&gt\;98%) even as accuracy degrades and wha t this decoupling of consistency from correctness means for evaluating com pressed multimodal systems in low-resource deployments.</p> END:VEVENT END:VCALENDAR