Prompted models have demonstrated impressive few-shot learning abilities.
Repeated interactions at test-time with a single model, or the composition of
multiple models together, further expands capabilities. These compositions are
probabilistic models, and may be expressed in the language of graphical models
with random variables whose values are complex data types such as strings.
Cases with control flow and dynamic structure require techniques from
probabilistic programming, which allow implementing disparate model structures
and inference strategies in a unified language. We formalize several existing
techniques from this perspective, including scratchpads / chain of thought,
verifiers, STaR, selection-inference, and tool use. We refer to the resulting
programs as language model cascades.
%0 Journal Article
%1 dohan2022language
%A Dohan, David
%A Xu, Winnie
%A Lewkowycz, Aitor
%A Austin, Jacob
%A Bieber, David
%A Lopes, Raphael Gontijo
%A Wu, Yuhuai
%A Michalewski, Henryk
%A Saurous, Rif A.
%A Sohl dickstein, Jascha
%A Murphy, Kevin
%A Sutton, Charles
%D 2022
%K language
%T Language Model Cascades
%U http://arxiv.org/abs/2207.10342
%X Prompted models have demonstrated impressive few-shot learning abilities.
Repeated interactions at test-time with a single model, or the composition of
multiple models together, further expands capabilities. These compositions are
probabilistic models, and may be expressed in the language of graphical models
with random variables whose values are complex data types such as strings.
Cases with control flow and dynamic structure require techniques from
probabilistic programming, which allow implementing disparate model structures
and inference strategies in a unified language. We formalize several existing
techniques from this perspective, including scratchpads / chain of thought,
verifiers, STaR, selection-inference, and tool use. We refer to the resulting
programs as language model cascades.
@article{dohan2022language,
abstract = {Prompted models have demonstrated impressive few-shot learning abilities.
Repeated interactions at test-time with a single model, or the composition of
multiple models together, further expands capabilities. These compositions are
probabilistic models, and may be expressed in the language of graphical models
with random variables whose values are complex data types such as strings.
Cases with control flow and dynamic structure require techniques from
probabilistic programming, which allow implementing disparate model structures
and inference strategies in a unified language. We formalize several existing
techniques from this perspective, including scratchpads / chain of thought,
verifiers, STaR, selection-inference, and tool use. We refer to the resulting
programs as language model cascades.},
added-at = {2022-07-24T00:04:33.000+0200},
author = {Dohan, David and Xu, Winnie and Lewkowycz, Aitor and Austin, Jacob and Bieber, David and Lopes, Raphael Gontijo and Wu, Yuhuai and Michalewski, Henryk and Saurous, Rif A. and Sohl dickstein, Jascha and Murphy, Kevin and Sutton, Charles},
biburl = {https://www.bibsonomy.org/bibtex/2d9b0c4297f06a774bfec8248871b89a5/trgardos},
description = {Language Model Cascades},
interhash = {8310c907d2e4e5e4f7fe93b812257b41},
intrahash = {d9b0c4297f06a774bfec8248871b89a5},
keywords = {language},
note = {cite arxiv:2207.10342Comment: Presented as spotlight at the Beyond Bases workshop at ICML 2022 (https://beyond-bayes.github.io)},
timestamp = {2022-07-24T00:04:33.000+0200},
title = {Language Model Cascades},
url = {http://arxiv.org/abs/2207.10342},
year = 2022
}