text
stringlengths
0
89.3k
ferent αrsettings To the right of the sequence names are the
αrvalues corresponding to the results reported in the main text
possesses stability under random occlusion settings around
30 occlusion
Simulating occlusion by randomly adding masks is
friendly to the recovery of observation matrix whereas the
occlusion scenario in reality tends to be more complex We
tested on the real occlusion data provided by the NRSfM
Challenge Dataset and the results are shown in Fig 14 The
figure shows the reconstruction error comparison with the
CSF2 15 on two sequences Balloon and Stretch Our ap
proach is superior in terms of mean performance regardless
of whether the data is occluded or not In addition the oc
clusion rate is the main factor affecting the performance of
the algorithm The average occlusion rate under all camera
motion types for Balloon is 38 while for Stretch it is 13
a Lowrank Property of Shape Sequences in H3WB Dataset
b Smoothing Property of Shape Sequences in H3WB Dataset
Figure 13 Analyzing the lowrank and smoothing properties of
shape sequences in the H3WB dataset The black dotted line rep
resents the values of the metrics for the pickup sequence in the
Mocap dataset The comparison reveals that the sequences in the
H3WB dataset have a greater magnitude of motion and are more
difficult to recover using lowrank and smoothing constraints
ae3don Seq Balloon
be3don Seq Stretch
Figure 14 Reconstruction error on missing data in NRSfM Chal
lenge Dataset For a better comparison the figure shows the test
results under multiple camera motion types and compares the re
construction errors without missing data
As a result the methods accuracy degradation is more obvi
ous on Balloon Apart from the occlusion rate the accuracyof our method also relies on the results of matrix comple
tion Matrix completion based on the lowrank assumption
tends to fail for some special occlusion scenarios eg the
object is completely invisible at some moments such as Ar
ticulatedtricky in NRSfM Challenge Dataset the object is
completely occluded in the first 35 frames and excessive
occlusion such as Tearingtricky with the occlusion rate of
56 Searching for more robust matrixcompletion algo
rithms or updating the observation matrix in iterations 36
are potential solutions
94 Additional Qualitative Results
In this section we provide more visualizations of the 3D
reconstruction results of our method Fig 10 shows addi
tional qualitative results on the NRSfM Challenge dataset
and Fig 11 illustrates 3D reconstruction results compared
with GT on the Semidense and H3WB datasetsarXiv240609363v2 csAI 19 Jun 2024ElicitationGPT Text Elicitation Mechanisms via Language Models
Yifan Wu
Northwestern University
yifanwuunorthwesterneduJason Hartline
Northwestern University
hartlinenorthwesternedu
Abstract
Scoring rules evaluate probabilistic forecasts of an unknown state against the realized state
and are a fundamental building block in the incentivized elicitation of inf ormation and the train
ing of machine learning models This paper develops mechanisms for sc oring elicited text against
ground truth text using domainknowledgefree queries to a large language model specifically
ChatGPT and empirically evaluates their alignment with human prefer ences The empiri
cal evaluation is conducted on peer reviews from a peergrading da taset and in comparison to
manual instructor scores for the peer reviews
1 Introduction
This paper constructs proper scoring rules for text and meas ures their alignment with human
evaluators Scoring rules evaluate elicited responses aga inst ground truth responses A scoring
rule is defined as aligned with humans if its ranking of responses by score is similar to the human
ranking of responses A scoring rule is defined as proper if truthful reporting of beliefs optimizes
expected score relative to those beliefs A main applicatio n of proper scoring rules is in training
machine learning models where they are known as proper loss f unctions
The theory of proper scoring rules and loss functions is well established for numerical predictions
eg the expected value of a number the probability that a l abel is correct or the distribution of an
unknown state In machine learning training with proper sc oring rules yields desirable properties
First nonproper scoring rules can misrank the Bayesian op timal predictor relative to nonoptimal
predictors Gneiting and Raftery 2007 while proper scor ing rules do not Second training with
a proper scoring rule leads to calibrated predictors even t hough the predictor may not be globally
optimal Blasiok et al 2024 Bsuppress lasiok et al 2024 A pred iction is calibrated if it is conditionally
correct Calibrated predictions are desirable because the y can be subsequently interpreted as true
probabilities in any decision problem
Our paper designs proper scoring rules for text and evaluate s their alignment with human
preferences The standard supervised finetuning SFT met hod applies a log score to evaluating
predictions of the next word in the text SFT scores a distrib ution over sequences of words instead
of the semantic meaning of a particular piece of text and doe s not align with human preferences on
outofsample reports For example SFT is known to score re ports such as I dont know below
incorrect answers since I dont know appears rarely in th e data Recently reinforcement learning
from human feedback RLHF Ouyang et al 2022 has resolve d this problem but RLHF is known
to be susceptible to manipulations eg Perez et al 2023 In contrast our proper scoring rules
for text may allow alignment improvements in SFT and avoid ma nipulations in RLHF though this
paper does not evaluate these possibilities directly
1Our approach is a framework for constructing scoring rules f or eliciting truthful textual reports
that evaluates text in its highdimensional semantic space This framework assumes that a data
set of prompts and ground truth responses analogous to feat ures and true labels in settings of
supervised learning The data is assumed to be partitioned i n advance into clusters The framework
treats each cluster separately
1 Identify the state space as a collection of summary points across the ground truth responses
in the cluster
2 Map each ground truth response into a state vector with coo rdinates corresponding to each
summary point with ternary values of agree disagree o r not applicable
3 Construct the prior distribution from the empirical dist ribution of the ground truth state
vectors
A response can then be scored by
4 Map the response text into a response vector analogously to the state vector construction
5 Score the response vector against the ground truth state v ector with a proper scoring rule