BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//University of Iceland//AI Centre//EN
CALSCALE:GREGORIAN
METHOD:PUBLISH
BEGIN:VEVENT
UID:da955f6e-8885-4e43-a1b2-1806f5936c9e@ai.hi.is
DTSTAMP:20261011T062921Z
DTSTART:20260414T090000Z
DTEND:20260414T100000Z
SUMMARY:Transducing Language Models
DESCRIPTION:Modern language models define distributions over strings\, but 
 downstream tasks often require different output formats. For instance\, a 
 model that generates byte-pair strings does not directly produce word-leve
 l predictions\, and a DNA model does not directly produce amino-acid seque
 nces. In such cases\, a deterministic string-to-string transformation can 
 convert the model's output to the desired form. This is a familiar pattern
  in probability theory: applying a function f  to a random variable X ~ p 
 yields a transformed random variable f(X) with an induced distribution. Wh
 ile such transformations are occasionally used in language modeling\, prio
 r work does not treat them as yielding new\, fully functional language mod
 els. We formalize this perspective and introduce a general framework for l
 anguage models derived from deterministic string-to-string transformations
 . We focus on transformations representable as finite-state transducers---
 a commonly used state-machine abstraction for efficient string-to-string m
 appings. We develop algorithms that compose a language model with an FST t
 o marginalize over source strings mapping to a given target\, propagating 
 probabilities through the transducer without altering model parameters and
  enabling conditioning on transformed outputs. We present an exact algorit
 hm\, an efficient approximation\, and a theoretical analysis. We conduct e
 xperiments in three domains: converting language models from tokens to byt
 es\, from tokens to words\, and from DNA to amino acids. These experiments
  demonstrate inference-time adaptation of pretrained language models to ma
 tch application-specific output requirements.\n\nhttps://ai.hi.is/is/event
 s/colloquium-in-statistics-and-ai-18/
LOCATION:Íslensk erfðagreining\, Tjarnarsalur
URL:https://ai.hi.is/is/events/colloquium-in-statistics-and-ai-18/
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR