tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
and the ground truth is equally likely to correspond to any of the ground truths within the cluster
eg any of the instructor reviews of submissions on the sa me assignment from which we calculate
the empirical prior we use the principle of deferred decis ions to evaluate agent report against one
of the ground truth at the end of this analysis
ElicitationGPTprocesses the text report of the agent perhaps imperfectly and in a way that is
adversarially manipulated to obtain a numerical report r Now draw uniformly at random from
the empirical prior pa stateθ Since the numerical scoring rule used by ElicitationGPTis proper
and the state is drawn from the prior the expected score of an y ﬁxed report is at most the score
from reporting the prior ie EθpSrθEθpSpθ
32 Implementing ElicitationGPTwith Imperfect Language Oracles
Our oracle model abstracts basic functionalities away from the engineering constraints The main
constraint we face in real implementation is the context len gth constraint which we address in
this section For example when LLM receives the question an swering query of the reported text
the conversation history includes all summarization queri es and question answering queries to the
ground truth texts usually leading to insuﬃciency of the co ntext length As examples GPT4 has
a maximum context length of 32 ktokens GPT35turbo16k has a maximum context length of
16k In practice the LLMs are not able to process the conversati on history of all queries Moreover
for each query the cost is linearly related to the length of t he conversation history Thus the total
monetary cost of each task has a quadratic dependency on the n umber of queries in the conversation
history Our implementation deploys tricks to save context length of queries and monetary cost
Our algorithm saves context length by reusing history We mo dify the abstractions of oracles
to present our implementation of ElicitationGPT The main modiﬁcation is an additional input of
conversation history to the oracle which shows how we pass t he conversation history and reduce
context length
A conversation history h QiAiiis a list of conversations Each conversation is a round
QA of queryanswer pairs
Summarization ˆOS
Input a list Iiiof text conversation history h
Output
a list t1t2tz as a summary of the content in Iii
12A list of evidence E1Ez where
eachEi jeij is a list of evidence text that text jmentions tiin a segment eij
current round conversation QA
Note that we modify the summarization oracle to be stronger t han the perfect summariza
tion oracle In addition to a summarization the oracle also returns evidence text This
modiﬁcation is consistent with the actual prompt we use
Question Answering ˆOA
Input text review Rand a list of text statements T1Tm history h
Output
a vector r 01mof summarization points on T1Tm On each dimension
ri 1 ifRhas a segment supporting a positive opinion on T 0 if a negative opinion or
ifRdoes not mention anything related to T
current round conversation QA
The implementation of ElicitationGPTis the following fourstage algorithm
Input
a list of ground truth text I1Is within the same cluster eg instructor reviews on
submissions to the same homework assignment
an agent report Pcorresponding to the kth ground truth eg a peer review on the same
submission as the kth instructor review
Output A score in 0 1 on the agent report
1extracting dimensions of summary points and topics Deﬁnition 10 extracts semantic
dimensions of summary points ﬁrst In implementation we re verse the order of summariza
tion3 We ﬁrst summarize ground truth texts into topics then summ arize into indicator
states
Extract topics ﬁrst
t1tzE1Ezh1ˆOSIiih
Then for each topic ti extract the indicators
For each i Ti1TiziEi1Eizih2iˆOSEih1
2calculating prior Process the ground truth texts
For each topic ti for each ground truth text j process the indicator states on text jin topic
i
tiIjtih3ijˆOAIjTi1Tizih1h2i
For each indicator state θixti count the frequency pθix of 1s from ground truth texts
3The LLM is queried twice on the summarization task The ﬁrst q uery is a coarse summarization task which
leads to topics and the second query is a ﬁnegrained summar ization task for each topic The procedure follows the
OpenAIs oﬃcial prompt engineering tricks OpenAI 2023 which recommends easy tasks to be processed ﬁrst
13type variation variants
algorithm GPT model GPT4 GPT35
algorithm scoring rule 4 scoring rules for text Section 222
evaluation metrics correlation with instructor score correlation with
overall student grades
evaluation benchmark instructor text review instructor numerical review
algorithm numerical score direct GPT score
data course Algorithm Class 1 Algorithm Class 2 Mechanism
Design
Table 1 Summary of variants of empirical evaluation Furth er details in Section 41 and Section 42
3mapping agent report to belief Extract the reported states from the report text P Note
that here we query on the history in the ﬁrst stage
For each topic ti
tirih4iˆOAPTi1Tizih1h2i
4knowitornot scoring rule Apply the scoring rule Sprθk
Note that when implementing with ChatGPT we combine the sta ges extracting semantic di
mensions and calculating prior in a single prompt
4 Empirical Evaluation
We test diﬀerent conﬁgurations of ElicitationGPTon several data sets and compare to several bench
marks These variations are summarized in Table 1 and will be described in detail subsequently in
Section 41 Section 42 displays our evaluation results de scribed in Section 41 and summarizes
observations from the results
In Section 43 we show direct GPT queries are manipulable c ontrasting to the theoretical
guarantee of ElicitationGPTin Section 31 The manipulations increase the output score from GPT
if it is directly queried to compare peer review with instruc tor review
41 Dataset and Metric
We use peer review data from three classes two instances of a n algorithms class an undergraduate
course and one mechanism design class a graduate course In each assignment of the classes a
subset of students homework submission is drawn for peer re view Six to eight peers are randomly
matched with each homework submission For each class our d ataset also includes the students
aggregated ﬁnal score in the class We removed data from peer s who did not submit reviews for all
assignments in a class
Algorithm Class 1and 2 The dataset for Algorithm Class 1 consists of 276 reviews fr om 23
peers on 89 homework submissions to 12 assignments For Alg orithm Class 2 the dataset has 240

and the ground truth is equally likely to correspond to any of the ground truths within the cluster

eg any of the instructor reviews of submissions on the sa me assignment from which we calculate

the empirical prior we use the principle of deferred decis ions to evaluate agent report against one

of the ground truth at the end of this analysis

ElicitationGPTprocesses the text report of the agent perhaps imperfectly and in a way that is

adversarially manipulated to obtain a numerical report r Now draw uniformly at random from

the empirical prior pa stateθ Since the numerical scoring rule used by ElicitationGPTis proper

and the state is drawn from the prior the expected score of an y ﬁxed report is at most the score

from reporting the prior ie EθpSrθEθpSpθ

32 Implementing ElicitationGPTwith Imperfect Language Oracles

Our oracle model abstracts basic functionalities away from the engineering constraints The main

constraint we face in real implementation is the context len gth constraint which we address in

this section For example when LLM receives the question an swering query of the reported text

the conversation history includes all summarization queri es and question answering queries to the

ground truth texts usually leading to insuﬃciency of the co ntext length As examples GPT4 has

a maximum context length of 32 ktokens GPT35turbo16k has a maximum context length of

16k In practice the LLMs are not able to process the conversati on history of all queries Moreover

for each query the cost is linearly related to the length of t he conversation history Thus the total

monetary cost of each task has a quadratic dependency on the n umber of queries in the conversation

history Our implementation deploys tricks to save context length of queries and monetary cost

Our algorithm saves context length by reusing history We mo dify the abstractions of oracles

to present our implementation of ElicitationGPT The main modiﬁcation is an additional input of

conversation history to the oracle which shows how we pass t he conversation history and reduce

context length

A conversation history h QiAiiis a list of conversations Each conversation is a round

QA of queryanswer pairs

Summarization ˆOS

Input a list Iiiof text conversation history h

Output

a list t1t2tz as a summary of the content in Iii

12A list of evidence E1Ez where

eachEi jeij is a list of evidence text that text jmentions tiin a segment eij

current round conversation QA

Note that we modify the summarization oracle to be stronger t han the perfect summariza

tion oracle In addition to a summarization the oracle also returns evidence text This

modiﬁcation is consistent with the actual prompt we use

Question Answering ˆOA

Input text review Rand a list of text statements T1Tm history h

Output

a vector r 01mof summarization points on T1Tm On each dimension

ri 1 ifRhas a segment supporting a positive opinion on T 0 if a negative opinion or

ifRdoes not mention anything related to T

current round conversation QA

The implementation of ElicitationGPTis the following fourstage algorithm

Input

a list of ground truth text I1Is within the same cluster eg instructor reviews on

submissions to the same homework assignment

an agent report Pcorresponding to the kth ground truth eg a peer review on the same

submission as the kth instructor review

Output A score in 0 1 on the agent report

1extracting dimensions of summary points and topics Deﬁnition 10 extracts semantic

dimensions of summary points ﬁrst In implementation we re verse the order of summariza

tion3 We ﬁrst summarize ground truth texts into topics then summ arize into indicator

states

Extract topics ﬁrst

t1tzE1Ezh1ˆOSIiih

Then for each topic ti extract the indicators

For each i Ti1TiziEi1Eizih2iˆOSEih1

2calculating prior Process the ground truth texts

For each topic ti for each ground truth text j process the indicator states on text jin topic

i

tiIjtih3ijˆOAIjTi1Tizih1h2i

For each indicator state θixti count the frequency pθix of 1s from ground truth texts

3The LLM is queried twice on the summarization task The ﬁrst q uery is a coarse summarization task which

leads to topics and the second query is a ﬁnegrained summar ization task for each topic The procedure follows the

OpenAIs oﬃcial prompt engineering tricks OpenAI 2023 which recommends easy tasks to be processed ﬁrst

13type variation variants

algorithm GPT model GPT4 GPT35

algorithm scoring rule 4 scoring rules for text Section 222

evaluation metrics correlation with instructor score correlation with

overall student grades

evaluation benchmark instructor text review instructor numerical review

algorithm numerical score direct GPT score

data course Algorithm Class 1 Algorithm Class 2 Mechanism

Design

Table 1 Summary of variants of empirical evaluation Furth er details in Section 41 and Section 42

3mapping agent report to belief Extract the reported states from the report text P Note

that here we query on the history in the ﬁrst stage

For each topic ti

tirih4iˆOAPTi1Tizih1h2i

4knowitornot scoring rule Apply the scoring rule Sprθk

Note that when implementing with ChatGPT we combine the sta ges extracting semantic di

mensions and calculating prior in a single prompt

4 Empirical Evaluation

We test diﬀerent conﬁgurations of ElicitationGPTon several data sets and compare to several bench

marks These variations are summarized in Table 1 and will be described in detail subsequently in

Section 41 Section 42 displays our evaluation results de scribed in Section 41 and summarizes

observations from the results

In Section 43 we show direct GPT queries are manipulable c ontrasting to the theoretical

guarantee of ElicitationGPTin Section 31 The manipulations increase the output score from GPT

if it is directly queried to compare peer review with instruc tor review

41 Dataset and Metric

We use peer review data from three classes two instances of a n algorithms class an undergraduate

course and one mechanism design class a graduate course In each assignment of the classes a

subset of students homework submission is drawn for peer re view Six to eight peers are randomly

matched with each homework submission For each class our d ataset also includes the students

aggregated ﬁnal score in the class We removed data from peer s who did not submit reviews for all

assignments in a class

Algorithm Class 1and 2 The dataset for Algorithm Class 1 consists of 276 reviews fr om 23

peers on 89 homework submissions to 12 assignments For Alg orithm Class 2 the dataset has 240