BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//https://techplay.jp//JP
CALSCALE:GREGORIAN
METHOD:PUBLISH
X-WR-CALDESC:[104th TrustML Young Scientist Seminar] Talk by John Robertson
  (UT Austin) "Language Model Control and Reliability: Understanding Steer
 ing Vectors and Agentic Aging"
X-WR-CALNAME:[104th TrustML Young Scientist Seminar] Talk by John Robertson
  (UT Austin) "Language Model Control and Reliability: Understanding Steer
 ing Vectors and Agentic Aging"
X-WR-TIMEZONE:Asia/Tokyo
BEGIN:VTIMEZONE
TZID:Asia/Tokyo
BEGIN:STANDARD
DTSTART:19700101T000000
TZOFFSETFROM:+0900
TZOFFSETTO:+0900
TZNAME:JST
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
UID:997333@techplay.jp
SUMMARY:[104th TrustML Young Scientist Seminar] Talk by John Robertson (UT 
 Austin) "Language Model Control and Reliability: Understanding Steering V
 ectors and Agentic Aging"
DTSTART;TZID=Asia/Tokyo:20260623T140000
DTEND;TZID=Asia/Tokyo:20260623T150000
DTSTAMP:20260619T090957Z
CREATED:20260619T060139Z
DESCRIPTION:イベント詳細はこちら\nhttps://techplay.jp/event/99733
 3?utm_medium=referral&utm_source=ics&utm_campaign=ics\n\nDate and Time: J
 une 23\, 2026\, 14:00 -- 15:00 (JST)\nVenue: Online + Meeting RoomB at Ni
 honbashi/＜AIP専用＞会議室B\n*Meeting RoomB is available to AIP re
 searchers only\n\nTitle: Language Model Control and Reliability: Understa
 nding Steering Vectors and Agentic Aging\n\nSpeaker:\nJohn Robertson (UT 
 Austin)\n\nAbstract:\nReliable use of large language models requires both
  controlling their behavior and trusting that behavior over time\; this t
 alk will discuss approaches to each. Activation steering is an appealingl
 y lightweight way to control a model without retraining\, but its effecti
 veness varies sharply across concepts. Prior work often reads this as evi
 dence that some concepts cannot be captured by one direction. I argue tha
 t much of this variability is instead search difficulty: a useful rank-1 
 intervention usually exists\, but finding the right layer and coefficient
  is expensive. I show that the directional alignment of contrastive activ
 ations at the prompt boundary predicts where effective interventions emer
 ge\, turning steering into a budget-constrained search that geometry-guid
 ed optimization solves with roughly 40% fewer evaluations across three mo
 del families. I introduce concept granularity\, a measure of how much the
  locally agreed-upon steering direction rotates across input contexts. Gr
 anularity is computable from cached activations before any steering is ru
 n\, and it predicts both how hard a concept is to optimize and the steeri
 ng quality ultimately achievable. I close by turning from control to reli
 ability over time: deployed agents are still evaluated like freshly initi
 alized models\, even though a frozen-weight agent drifts as it compresses
  history\, retrieves from a growing memory store\, and revises facts. I b
 riefly present AgingBench\, a longitudinal benchmark that organizes this 
 degradation into four mechanisms (compression\, interference\, revision\,
  and maintenance) and localizes where in the memory pipeline reliability 
 breaks down.\n\nShort Bio:\nJohn T. Robertson is a PhD student in Electri
 cal and Computer Engineering at the University of Texas at Austin\, co-ad
 vised by Haris Vikalo and Atlas Wang. His research focuses on the reliabi
 lity and controllability of large language models\, spanning mechanistic 
 interpretability\, activation steering\, and the longitudinal evaluation 
 of deployed agents. He is currently first-authoring work on the geometry 
 of activation steering and contributing to a benchmark for long-lived age
 nt reliability\, both under review at NeurIPS 2026. John has a multi-disc
 iplinary history of work. During his undergraduate degree\, he developed 
 transformer-based methods for detecting tumor-causing viral DNA\, publish
 ed in the Journal of Computational Biology and PLOS Computational Biology
 . He additionally has several patents pending from his time at Texas Inst
 ruments' Kilby Labs implementing efficient vision networks on low-end dev
 ices\, and a second-place submission to the ICASSP SAND challenge for dia
 gnosing ALS severity from audio data. His work is supported by the Charle
 s W. and Margaret A. Tolbert Endowed Fellowship\, Amazon Web Services\, a
 nd the Texas Advanced Computing Center.
LOCATION:オンライン
URL:https://techplay.jp/event/997333?utm_medium=referral&utm_source=ics&utm
 _campaign=ics
END:VEVENT
END:VCALENDAR