BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//wp-events-plugin.com//7.4.0.1//EN
TZID:Asia/Kolkata
X-WR-TIMEZONE:Asia/Kolkata
BEGIN:VEVENT
UID:218@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20260807T140000
DTEND;TZID=Asia/Kolkata:20260807T150000
DTSTAMP:20260803T144416Z
URL:https://cds.iisc.ac.in/events/m-tech-research-thesis-colloquium-102-cd
 s-07-august-2026-watermarking-open-weight-large-language-models/
SUMMARY:M.Tech Research Thesis {Colloquium}: 102: CDS: 07\, August 2026 “
 Watermarking Open-Weight Large Language Models”
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nM.Tech Research 
 Thesis {Colloquium}\n\n\n\nSpeaker: Mr. Miroojin Bakshi\nS.R. Number: 06-1
 8-01-10-12-24-1-25023\nTitle: “Watermarking Open-Weight Large Language M
 odels”\nResearch Supervisor: Dr. Danish Pruthi\nDate &amp\; Time : Augus
 t 07\, 2026 (Friday)\, 02:00 PM\nVenue : #102 CDS Seminar Hall\n\n\n\nABST
 RACT\nAs large language models (LLMs) become widely used\, it is increasin
 gly difficult to tell whether a passage was written by a person or produce
 d by a model. Watermarking addresses this by embedding a hidden statistica
 l signal in model outputs so that a verifier can later attribute the text.
  Most existing watermarking methods change how tokens are sampled at gener
 ation time. This approach works when a model provider controls inference t
 hrough an API\, but it fails for open-weight models: once the weights are 
 public\, a user can simply bypass any sampling-time watermark and still ge
 nerate high-quality text.\n\nThis thesis introduces OpenStamp\, a watermar
 k for open-weight language models. Instead of changing the sampling proces
 s\, OpenStamp edits the model’s final output layer (the unembedding laye
 r)\, so generation from the released checkpoint produces text with a detec
 table watermark. Detection compares how likely a candidate passage is unde
 r the released watermarked model versus a privately retained base model\, 
 using a length-normalized log-likelihood ratio.\n\nWe evaluate OpenStamp a
 gainst prior open-weight watermarking methods along four axes motivated by
  open-weight deployment: (i) detectability at low false-positive rates\, t
 ogether with the trade-off against text quality\; (ii) robustness to parap
 hrasing of generated text\; (iii) durability under post-release weight mod
 ifications such as fine-tuning and quantization\; and (iv) impact on downs
 tream task accuracy. Based on our evaluation\, OpenStamp achieves near-per
 fect detection at low false-positive rates and\, at comparable text qualit
 y\, matches or exceeds the detectability of prior open-weight methods. Aft
 er LLM-based paraphrasing and after post-hoc fine-tuning\, it remains more
  detectable than baseline approaches\, although detectability still degrad
 es under continued fine-tuning. On the downstream tasks we evaluate\, it s
 hows little to no accuracy degradation. Taken together\, these results sho
 w that OpenStamp is an effective and practical mechanism for verifying the
  provenance of text from open-weight LLMs.\n\n\n\nALL ARE WELCOME
CATEGORIES:Events,MTech Research Thesis Colloquium
END:VEVENT
BEGIN:VTIMEZONE
TZID:Asia/Kolkata
X-LIC-LOCATION:Asia/Kolkata
BEGIN:STANDARD
DTSTART:20250807T140000
TZOFFSETFROM:+0530
TZOFFSETTO:+0530
TZNAME:IST
END:STANDARD
END:VTIMEZONE
END:VCALENDAR