BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//NextToU//EN
CALSCALE:GREGORIAN
BEGIN:VEVENT
UID:0b55a097-e991-4960-a00a-d1adadb52ea7@nexttoyou
DTSTAMP:20261106T162922Z
DTSTART:20261113T163000Z
DTEND:20261113T173000Z
SUMMARY:CIS Seminar Series 2026 -Jingtong Hu Ph.D.-University of Pittsburg
 h
LOCATION:Kirkbride Lecture Hall 204\, Kirkbride Lecture Hall\, University 
 of Delaware\, Newark\, DE 19716\, USA
DESCRIPTION:How to Make LLMs Cool: Breaking the Thermal Wall through Therm
 ally Adaptive Algorithms and Human-in-the-Loop Pacing for On-Device Infer
 ence\n\nABSTRACT\n\nOn-device LLM inference is driven by critical needs f
 or privacy\, low-latency safety\, and off-grid availability. Yet sustaine
 d inference drives mobile GPU temperatures past safe thresholds within mi
 nutes\, forcing devices into aggressive throttling that destroys the very
  benefits motivating local execution. Current thermal management treats t
 his as a hardware-only problem — reactive frequency scaling that is blind
  to the workload it serves. In this talk\, I will present three systems t
 hat rethink thermal management as a cross-layer\, workload-aware design o
 bjective spanning model internals and human perception.\n\nFirst\, I will
  introduce GreenDynaMix\, a runtime that converts MoE expert precision in
 to a dynamic thermal control knob. A dual-PID control loop couples DVFS-d
 riven frequency reduction with sensitivity-aware precision demotion\, com
 pensating for throttling-induced throughput loss while maintaining output
  quality. Second\, I will present TRAIL\, which extends this idea to Visi
 on-Language Models by dynamically replacing thermally redundant transform
 er layers with pre-computed linear projections\, recovering over 50% of t
 hrottling-induced throughput loss while maintaining 95.9% accuracy on VLM
  benchmarks. Third\, I will describe GreenRead\, a perceptual throughput 
 pacing framework that uses gaze tracking to match token generation spe\n\
 nhttp://next-to-u.app/events/0b55a097-e991-4960-a00a-d1adadb52ea7
URL:http://next-to-u.app/events/0b55a097-e991-4960-a00a-d1adadb52ea7
END:VEVENT
END:VCALENDAR