text
stringlengths
0
89.3k
reviews from 24 peers on 59 homework submissions to 10 assig nments On each submission both
the instructor and the peers leave a numerical review and a te xtual review on each of the three
14explicit rubric points AnswerAlgorithm ProofAnalysi s and Clarity Additional to reviews the
instructor grades the peers with two scores in 0 10 normalized to 0 1 in our analysis one for
the quality of the numerical review and the other one for the textual review
Mechanism Design Class The dataset for Mechanism Design Class has 112 reviews from 1 2
peers on 32 homework submissions to 8 assignments On each s ubmission both the instructor
and the peers leave a numerical review and a textual review on each of the three explicit rubric
points Answer Argument and Clarity Additional to revie ws the instructor grades the peer with
one score for the quality of the textual review For this data set the instructor does not grade the
numerical review
Metric As our metric we compute the correlation of algorithmic sco re with two ground truth
scores
theinstructor score of the review which is potentially noisy but reflects the ins tructors
preference over reviews
theoverall student grades a weighted average over homework scores peer grading scor es
exam scores etc reflecting the peers ground truth perfor mance
We compute the correlation of peers ranking with our two met rics since the algorithmic scores
are not on the same scale with instructor scores We calculat e the average score of each individual
peer throughout the quarter sort them by this average and u se Spearmans rank correlation co
efficient Spearmans correlation to compare two rankings The Spearmans correlation calculates
the average quadratic distance of each item in two rankings The average quadratic distance is nor
malized such that perfectly aligning rankings have correla tion 1 reversed rankings have correlation
1 and independent draws have expected correlation 0
For correlation with our two metrics instructor score and o verall student grades we test four
scoring rules for textual review as introduced in Section 2 2 To understand the performance of
ElicitationGPT we compare against the following benchmarks
The algorithmic scores for numerical review We evaluate wh ether numerical reviews or
textual reviews are more aligned with the metrics We presen t the average quadratic AQ
scoring rule as the benchmark numerical scoring rule and com pare it for alignment with
ElicitationGPTon textual reviews In Appendix A2 we show that among numer ical scoring
rules the AQ scoring rule for numerical review better align s with instructor preference on
numerical reviews
The instructor scores for both textual and numerical review We evaluate whether the in
structor scores or ElicitationGPTscores better align with overall student grades as their tru e
performance
The naive approach the direct GPT score by prompting GPT to compare reviews and output
a score directly In additional alignment comparison in Se ction 43 we include manipulation
strategies that tricks GPT to output a higher score in contr ast to the theoretical properness
of ElicitationGPT
We present the implementation details in Appendix B includ ing prompts
15Numerical Textual ElicitationGPT Textual
AQ AMV AV AFMV AFV GPT
Algorithm 1 074 058 069 057 065 084
Algorithm 2 074 033 032 042 065 073
MD 032 061 062 054 063 043
Table 2 The Spearmans correlation between the instructor score for textual review and the
following scores algorithmic score for numerical review Average Quadratic ElicitationGPT and
direct GPT queries
Numerical Textual Numerical Textual ElicitationGPT Textual
Instructor score AQ AMV AV AFMV AFV GPT
Algorithm 1 049 055 031 063 068 06 065 058
Algorithm 2 069 048 042 023 027 052 063 046
MD 047 019 077 066 074 059 043
Table 3 The Spearmans correlation between overall student grades and the following scores
instructor score for both numerical review and textual revi ew algorithmic score for numerical
review Average Quadratic ElicitationGPT and direct GPT queries
42 Results
ElicitationGPTis built on OpenAIs GPT4 The rank correlation with our two metrics are shown
in Table 2 and Table 3 respectively We summarize the observ ations from empirical results in
Section 421 Section 422 Section 423 and Section 4 24 To make the study more replicatable
we set max temperature of the model to 0 001
In Appendix A1 we present a comparison of different scoring r ules implemented with GPT4
and GPT35 In Appendix A2 we compare the alignment of diffe rent numerical scoring rules with
the instructor score of numerical reviews
421 ElicitationGPT which scoring rule is better
Filtering top 2 topics with the most summary points improves the robustness of scoring rules among
which the AFV scoring rule is consistently aligned with two m etrics both the instructor score and
the overall student grades In fact we observe that witho ut filtering topics ElicitationGPTis
affected more by low summarization quality When GPT fails at s ummarization topics with the
same semantic meaning repeatedly appear and get scored ind ucing noise in ElicitationGPTscores
422 Is textual score more informative than numerical score
On the correlation metric with the instructor score Table 2 the AQ score on numerical review
has the highest correlation with instructor score on textua l review in the two algorithm classes
but low correlation in mechanism design class We suggest tw o possible reasons First since the
instructor scores numerical reviews and textual reviews at the same time the two scores could be
strongly correlated Second the two algorithm classes are undergraduate classes where undergrad
uate students possibly read numerical scoring rubrics more carefully than graduate students In
16mechanism design we observe some peers leave highquality textual reviews but numerical reviews
far away from instructors
On the correlation metric with the overall student grades T able 3 ElicitationGPTon textual
review aligns better than the AQ score on numerical review w hich suggests that textual reviews
convey more information about students true performance t han numerical reviews
423 Can we show ElicitationGPTis more robust ElicitationGPTvs Instructor
We believe the scores generated by ElicitationGPTare less noisy than instructor scores We observe
that in most of the cases where ElicitationGPTranks a peer very differently from the instructor
the average instructor score is affected by outlier scores on o ne or more reviews
We validate the robustness of ElicitationGPTby setting the metric as overall student grades
which we view as the peers true performances of assessing ho mework submissions Table 3
ElicitationGPTscores with topic filtering all display a more robust correla tion with overall stu
dent grades than instructor scores on textual review
424 Is algorithmic score better than directly querying ElicitationGPTvs GPT
We directly ask GPT to compare the reviews and to output a scor e In Table 2 and Table 3
we present the Spearmans correlation of this GPT output in t he column textual GPT When the
metric is instructor scores direct GPT score aligns with in structor score well in the two Algorithm
classes but not in the Mechanism Design class When the metr ic is overall student grades the
direct GPT score has similar correlation level as the instru ctor score while ElicitationGPTaligns
better
We suggest two possible reasons why the above observations h appen First on Algorithm
classes direct GPT score aligns well with the instructor sc ore because the prompt provides the