tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
nual Meeting of the Association for Computational Linguistic s pages 40294039 Online July
2020 Association for Computational Linguistics doi 10 18653v12020aclmain371 URL
httpsaclanthologyorg2020aclmain371
Michael S Bernstein Greg Little Robert C Miller Bj orn Ha rtmann Mark S Ackerman David R
Karger David Crowell and Katrina Panovich Soylent a wor d processor with a crowd inside
InProceedings of the 23nd annual ACM symposium on User interfac e software and technology
pages 313322 2010
Jarossuppress law Bsuppress lasiok Parikshit Gopalan Lunjia Hu Adam Tau man Kalai and Preetum Nakkiran Loss
minimization yields multicalibration for large neural net works In 15th Innovations in Theoretical
Computer Science Conference ITCS 2024 2024
Jaroslaw Blasiok Parikshit Gopalan Lunjia Hu and Preetu m Nakkiran When does optimizing a
proper loss yield calibration Advances in Neural Information Processing Systems 36 2024
Benjamin Brooks Alexander Frankel and Emir Kamenica Inf ormation hierarchies Econometrica
90521872214 2022
Yiling Chen and FangYi Yu Optimal scoring rule design arXiv preprint arXiv210707420 2021
Paul Duetting Vahab Mirrokni Renato Paes Leme Haifeng Xu and Song Zuo Mechanism design
for large language models arXiv preprint arXiv231010826 2023
Rujun Gao Naveen Thomas and Arun Srinivasa Work in progre ss Large language model based
automatic grading study In 2023 IEEE Frontiers in Education Conference FIE pages 14
IEEE 2023
Tilmann Gneiting Making and evaluating point forecasts Journal of the American Statistical
Association 106494746762 2011
Tilmann Gneiting and Adrian E Raftery Strictly proper scor ing rules prediction and estimation
Journal of the American statistical Association 102477359378 2007
Jason D Hartline Liren Shan Yingkai Li and Yifan Wu Optim al scoring rules for multi
dimensional eﬀort In The Thirty Sixth Annual Conference on Learning Theory pages 26242650
PMLR 2023
Dhamma Kimpara Rafael Frongillo and Bo Waggoner Proper l osses for discrete generative models
InInternational Conference on Machine Learning pages 1701517040 PMLR 2023
Yuqing Kong and Grant Schoenebeck Eliciting expertise wit hout veriﬁcation In Proceedings of the
2018 ACM Conference on Economics and Computation EC 18 page 195212 New York NY
USA 2018 Association for Computing Machinery ISBN 97814 50358293 doi 1011453219166
3219172 URL httpsdoiorg10114532191663219172
20Tom Kwiatkowski Jennimaria Palomaki Olivia Redﬁeld Mic hael Collins Ankur Parikh Chris Al
berti Danielle Epstein Illia Polosukhin Matthew Kelcey Jacob Devlin Kenton Lee Kristina N
Toutanova Llion Jones MingWei Chang Andrew Dai Jakob U szkoreit Quoc Le and Slav
Petrov Natural questions a benchmark for question answer ing research Transactions of the
Association of Computational Linguistics 2019
Nicolas S Lambert Elicitation and evaluation of statistic al forecasts working paper 2011
Edith Law and Luis Von Ahn Human computation Morgan Claypool Publishers 2011
Yingkai Li Jason D Hartline Liren Shan and Yifan Wu Optim ization of scoring rules In
Proceedings of the 23rd ACM Conference on Economics and Computa tion pages 988989 2022
Yuxuan Lu Shengwei Xu Yichi Zhang Yuqing Kong and Grant S choenebeck Eliciting informative
text evaluations with large language models the 25th ACM Conference on Economics and
Computation 2024
John McCarthy Measures of the value of information Proceedings of the National Academy of
Sciences of the United States of America 429654 1956
Robert C Miller Greg Little Michael Bernstein Jeﬀrey P Big ham Lydia B Chilton Max Goldman
John J Horton and Rajeev Nayak Heads in the cloud XRDS Crossroads The ACM Magazine
for Students 1722731 2010
OpenAI Six strategies for getting better results httpsplatformopenaicomdocsguidespromptengi neer
2023 Accessed 20240211
Long Ouyang Jeﬀrey Wu Xu Jiang Diogo Almeida Carroll Wain wright Pamela Mishkin Chong
Zhang Sandhini Agarwal Katarina Slama Alex Ray et al Tr aining language models to follow
instructions with human feedback Advances in neural information processing systems 3527730
27744 2022
Maneesha Papireddygari and Bo Waggoner Contracts with inf ormation acquisition via scoring
rules In Proceedings of the 23rd ACM Conference on Economics and Computa tion pages 703
704 2022
Ethan Perez Sam Ringer Kamil e Lukoˇ si ut e Karina Nguy en Edwin Chen Scott Heiner Craig
Pettit Catherine Olsson Sandipan Kundu Saurav Kadavath et al Discovering language model
behaviors with modelwritten evaluations In 61st Annual Meeting of the Association for Com
putational Linguistics ACL 2023 pages 1338713434 Association for Computational Lingui stics
ACL 2023
Pranav Rajpurkar Jian Zhang Konstantin Lopyrev and Perc y Liang SQuAD 100000 ques
tions for machine comprehension of text In Jian Su Kevin Du h and Xavier Carreras edi
torsProceedings of the 2016 Conference on Empirical Methods in Na tural Language Processing
pages 23832392 Austin Texas November 2016 Associatio n for Computational Linguistics doi
1018653v1D161264 URL httpsaclanthologyorgD161264
Johannes Schneider Bernd Schenk Christina Niklaus and M ichaelis Vlachos Towards llmbased
autograding for short textual answers arXiv preprint arXiv230911508 2023
21Alexander Wei Nika Haghtalab and Jacob Steinhardt Jailb roken How does llm safety training
failAdvances in Neural Information Processing Systems 36 2024
22Textual ElicitationGPT
AFMV AFV
Algorithm Class 1
GPT4 057 065
GPT35turbo16k 029 018
Algorithm Class 2
GPT4 042 065
GPT35turbo16k 006 011
Table 5 The Spearmans correlation between the instructor score for textual review and
ElicitationGPTscores
MV AQ
Algorithm Class 1 044 084
Algorithm Class 2 049 072
Table 6 The Spearmans correlation between the instructor score for numerical review and
scoring rules for numerical review
A More Empirical Evaluations
A1 How does the power of language model aﬀect scoring GPT4 vs GPT35
We implement ElicitationGPTwith the same scoring rules via queries to diﬀerent versions o f Chat
GPT Querying GPT35 signiﬁcantly reduces the correlatio n between the instructor score and the
ElicitationGPT score We ﬁnd the bottleneck to the performa nce of GPT is the performance on
the summarization task We observe a low correlation of GPT 35 output with the instructor re
view when GPT35 tends to split indicator states into diﬀer ent categories Bad summarization
of topics induces the ﬁltering of kmost important topics to fail However since we use the sam e
prompt to query GPT4 and GPT35 it is possible that there a re better approaches to implement
ElicitationGPTwith GPT35
A2 Which numerical score is better AQ vs MV
In our paper we test two scoring rules for numerical grade el icitation

nual Meeting of the Association for Computational Linguistic s pages 40294039 Online July

2020 Association for Computational Linguistics doi 10 18653v12020aclmain371 URL

httpsaclanthologyorg2020aclmain371

Michael S Bernstein Greg Little Robert C Miller Bj orn Ha rtmann Mark S Ackerman David R

Karger David Crowell and Katrina Panovich Soylent a wor d processor with a crowd inside

InProceedings of the 23nd annual ACM symposium on User interfac e software and technology

pages 313322 2010

Jarossuppress law Bsuppress lasiok Parikshit Gopalan Lunjia Hu Adam Tau man Kalai and Preetum Nakkiran Loss

minimization yields multicalibration for large neural net works In 15th Innovations in Theoretical

Computer Science Conference ITCS 2024 2024

Jaroslaw Blasiok Parikshit Gopalan Lunjia Hu and Preetu m Nakkiran When does optimizing a

proper loss yield calibration Advances in Neural Information Processing Systems 36 2024

Benjamin Brooks Alexander Frankel and Emir Kamenica Inf ormation hierarchies Econometrica

90521872214 2022

Yiling Chen and FangYi Yu Optimal scoring rule design arXiv preprint arXiv210707420 2021

Paul Duetting Vahab Mirrokni Renato Paes Leme Haifeng Xu and Song Zuo Mechanism design

for large language models arXiv preprint arXiv231010826 2023

Rujun Gao Naveen Thomas and Arun Srinivasa Work in progre ss Large language model based

automatic grading study In 2023 IEEE Frontiers in Education Conference FIE pages 14

IEEE 2023

Tilmann Gneiting Making and evaluating point forecasts Journal of the American Statistical

Association 106494746762 2011

Tilmann Gneiting and Adrian E Raftery Strictly proper scor ing rules prediction and estimation

Journal of the American statistical Association 102477359378 2007

Jason D Hartline Liren Shan Yingkai Li and Yifan Wu Optim al scoring rules for multi

dimensional eﬀort In The Thirty Sixth Annual Conference on Learning Theory pages 26242650

PMLR 2023

Dhamma Kimpara Rafael Frongillo and Bo Waggoner Proper l osses for discrete generative models

InInternational Conference on Machine Learning pages 1701517040 PMLR 2023

Yuqing Kong and Grant Schoenebeck Eliciting expertise wit hout veriﬁcation In Proceedings of the

2018 ACM Conference on Economics and Computation EC 18 page 195212 New York NY

USA 2018 Association for Computing Machinery ISBN 97814 50358293 doi 1011453219166

3219172 URL httpsdoiorg10114532191663219172

20Tom Kwiatkowski Jennimaria Palomaki Olivia Redﬁeld Mic hael Collins Ankur Parikh Chris Al

berti Danielle Epstein Illia Polosukhin Matthew Kelcey Jacob Devlin Kenton Lee Kristina N

Toutanova Llion Jones MingWei Chang Andrew Dai Jakob U szkoreit Quoc Le and Slav

Petrov Natural questions a benchmark for question answer ing research Transactions of the

Association of Computational Linguistics 2019

Nicolas S Lambert Elicitation and evaluation of statistic al forecasts working paper 2011

Edith Law and Luis Von Ahn Human computation Morgan Claypool Publishers 2011

Yingkai Li Jason D Hartline Liren Shan and Yifan Wu Optim ization of scoring rules In

Proceedings of the 23rd ACM Conference on Economics and Computa tion pages 988989 2022

Yuxuan Lu Shengwei Xu Yichi Zhang Yuqing Kong and Grant S choenebeck Eliciting informative

text evaluations with large language models the 25th ACM Conference on Economics and

Computation 2024

John McCarthy Measures of the value of information Proceedings of the National Academy of

Sciences of the United States of America 429654 1956

Robert C Miller Greg Little Michael Bernstein Jeﬀrey P Big ham Lydia B Chilton Max Goldman

John J Horton and Rajeev Nayak Heads in the cloud XRDS Crossroads The ACM Magazine

for Students 1722731 2010

OpenAI Six strategies for getting better results httpsplatformopenaicomdocsguidespromptengi neer

2023 Accessed 20240211

Long Ouyang Jeﬀrey Wu Xu Jiang Diogo Almeida Carroll Wain wright Pamela Mishkin Chong

Zhang Sandhini Agarwal Katarina Slama Alex Ray et al Tr aining language models to follow

instructions with human feedback Advances in neural information processing systems 3527730

27744 2022

Maneesha Papireddygari and Bo Waggoner Contracts with inf ormation acquisition via scoring

rules In Proceedings of the 23rd ACM Conference on Economics and Computa tion pages 703

704 2022

Ethan Perez Sam Ringer Kamil e Lukoˇ si ut e Karina Nguy en Edwin Chen Scott Heiner Craig

Pettit Catherine Olsson Sandipan Kundu Saurav Kadavath et al Discovering language model

behaviors with modelwritten evaluations In 61st Annual Meeting of the Association for Com

putational Linguistics ACL 2023 pages 1338713434 Association for Computational Lingui stics

ACL 2023

Pranav Rajpurkar Jian Zhang Konstantin Lopyrev and Perc y Liang SQuAD 100000 ques

tions for machine comprehension of text In Jian Su Kevin Du h and Xavier Carreras edi

torsProceedings of the 2016 Conference on Empirical Methods in Na tural Language Processing

pages 23832392 Austin Texas November 2016 Associatio n for Computational Linguistics doi

1018653v1D161264 URL httpsaclanthologyorgD161264

Johannes Schneider Bernd Schenk Christina Niklaus and M ichaelis Vlachos Towards llmbased

autograding for short textual answers arXiv preprint arXiv230911508 2023

21Alexander Wei Nika Haghtalab and Jacob Steinhardt Jailb roken How does llm safety training

failAdvances in Neural Information Processing Systems 36 2024

22Textual ElicitationGPT

AFMV AFV

Algorithm Class 1

GPT4 057 065

GPT35turbo16k 029 018

Algorithm Class 2

GPT4 042 065

GPT35turbo16k 006 011

Table 5 The Spearmans correlation between the instructor score for textual review and

ElicitationGPTscores

MV AQ

Algorithm Class 1 044 084

Algorithm Class 2 049 072

Table 6 The Spearmans correlation between the instructor score for numerical review and

scoring rules for numerical review

A More Empirical Evaluations

A1 How does the power of language model aﬀect scoring GPT4 vs GPT35

We implement ElicitationGPTwith the same scoring rules via queries to diﬀerent versions o f Chat

GPT Querying GPT35 signiﬁcantly reduces the correlatio n between the instructor score and the

ElicitationGPT score We ﬁnd the bottleneck to the performa nce of GPT is the performance on

the summarization task We observe a low correlation of GPT 35 output with the instructor re

view when GPT35 tends to split indicator states into diﬀer ent categories Bad summarization

of topics induces the ﬁltering of kmost important topics to fail However since we use the sam e

prompt to query GPT4 and GPT35 it is possible that there a re better approaches to implement

ElicitationGPTwith GPT35

A2 Which numerical score is better AQ vs MV

In our paper we test two scoring rules for numerical grade el icitation