Representation Fidelity:Auditing Algorithmic Decisions About Humans Using Self-Descriptions

Theresa Elstner; Martin Potthast

cs.CLMar 5, 2026

Representation Fidelity:Auditing Algorithmic Decisions About Humans Using Self-Descriptions

Authors: Theresa Elstner, Martin Potthast

Organizations: Kassel University, Germany

Abstract

This paper introduces a new dimension for validating algorithmic decisions about humans by measuring the fidelity of their representations. Representation Fidelity measures if decisions about a person rest on reasonable grounds. We propose to operationalize this notion by measuring the distance between two representations of the same person: (1) an externally prescribed input representation on which the decision is based, and (2) a self-description provided by the human subject of the decision, used solely to validate the input representation. We examine the nature of discrepancies between these representations, how such discrepancies can be quantified, and derive a generic typology of representation mismatches that determine the degree of representation fidelity. We further present the first benchmark for evaluating representation fidelity based on a dataset of loan-granting decisions. Our Loan-Granting Self-Representations Corpus 2025 consists of a large corpus of 30 000 synthetic natural language self-descriptions derived from corresponding representations of applicants in the German Credit Dataset, along with expert annotations of representation mismatches between each pair of representations.

Figures & tables

Figure 1. Overview of our approach to analyze representation fidelity. Top row: An algorithmic decision system typically implements a standard classification pipeline. An individual $v$ is represented as $\mathbf{x}$ using a representation function $\alpha_{1}()$ . The representation $\mathbf{x}$ is then fed to the classifier $c()$ of an algorithmic decision system, which outputs the decision $c(\mathbf{x})$ . Bottom row: Our approach to analyze representation fidelity represents $v$ as a natural language self-description $\mathbf{d}$ using a representation function $\alpha_{2}()$ . The representation $\mathbf{d}$ is then compared to $\mathbf{x}$ manually and/or using a distance measure $\delta()$ to obtain a qualitative and/or a quantitative representation mismatch for the decision $c(\mathbf{x})$ about $v$ .

Types of Representation Mismatches
No	Name	Description	Example
			Self-Description vs. Input Representation
	Additional Aspect	aspect in $\mathbf{d}$ absent in $\mathbf{x}$	“vegan, no soy” vs “diet: no soy”
1	Specialization	more detail in $\mathbf{d}$ than in $\mathbf{x}$	“nurse” vs “skilled employee”
	Additional Subject	subject in $\mathbf{d}$ absent in $\mathbf{x}$	“parrot” vs “”
	Omitted Aspect	aspect in $\mathbf{x}$ absent in $\mathbf{d}$	“vegan” vs “diet: vegan, no soy”

Table 1. Differences between individuals’ input representations

\mathbf{x}

and their self-descriptions

\mathbf{d}

can be classified according to this typology of representation mismatches.

Figure 2. This example illustrates our annotation method for representation mismatch analysis by comparing an input representation $\mathbf{x}$ (left) to a generated natural-language self-description $\mathbf{d}$ (right). We first segment $\mathbf{d}$ into information-bearing units, aiming for exhaustive semantic coverage. We then assign to each unit one or more labels from four label types: (1) a label from $\mathbf{x}$ ’s feature set whenever a semantic correspondence exists, (2) an aspect label for any minimal information-bearing unit, (3) a new subject label for semantically coherent information units recurring across different self-descriptions, and (4) a specialization label whenever a feature in $\mathbf{x}$ is described in more detail in $\mathbf{d}$ .

Model	Generated Self-descriptions		Prompts
	$\mathbf{x}$ -based	free	$\mathbf{x}$ -based	free
gpt-4.1	5 000	5	1 000	1
gpt-4o	5 000	5	1 000	1
o3	5 000	5	1 000	1
llama-3.3-70b-versatile	5 000	5	1 000	1
kimi-k2-instruct	5 000	5	1 000	1

Table 2. Overview of generated self-descriptions and prompts.

Mismatch Type	Average Frequency
(Additional GCD + New Subjects + Aspect Labels) Additional Aspects	(2.11 + 4.78 + 15.32) 22.20
specializations	1.45
Omitted Subjects	7.07
Representation Fidelity (Additional Aspects + specializations)	23.65

Table 3. Averaged frequencies of annotated mismatch types per (

\mathbf{x}

\mathbf{d}

)-pair.

Method	Metric	Pearson’s r
		(a)	(b)	(c)	(d)	(e)	(f)
GloVe-wiki-50	Word Mover’s Distance	0.45	0.30	0.37	0.36	0.42	0.48
GloVe-wiki-50-preprocessed	Word Mover’s Distance	0.30	0.11	0.19	0.28	0.28	0.31
GloVe-twitter-200	Word Mover’s Distance	0.48	0.35	0.41	0.38	0.45	0.49

Table 4. Overview of the automatic quantification methods for representation fidelity an their correlation (Pearson’s

r

) to the manual quantification results for each annotated Type 1 mismatch component representation fidelity (a), additional GCD (b), additional subjects (c), aspect labels (d), additional aspects (e), and specializations (f).

Figure 3. This visualization of similarities of embedded self-descriptions as implemented in ( Ren et al., 2025 ) . Self-descriptions form clusters depending the large language model that generated them (gpt-4.1, gpt-4o, o3, llama-3.3-70b-versatile, moonshotai/kimi-k2-instruct, qwen/qwen3-32b).

Algorithmic Auditing

Algorithmic auditing evaluates whether automated systems are fair, reliable, robust, compliant, and accountable in their intended settings.

See all 143 papers

Do Reasoning Representations Help Humans Evaluate LLM Outputs?
Sep 8, 2026Jaewoo Lim, Sungbok Shin, Sanghyun HongLLM Evaluation LLM Interpretability
Oregon State University, Corvallis USA · Sogang University, Seoul, South Korea
Population Fidelity: Evaluating Population Representativeness in LLMs
Sep 28, 2026Neemias B. da Silva, Martin Lukk, Ali Sutani +5LLM Evaluation LLM Fine-Tuning
University of Toronto, Canada · Federal University of Technology – Paran´a, Brazil
Exploring How Fair Model Representations Relate to Fair Recommendations
Mar 25, 2026Bjørnar Vassøy, Benjamin Kille, Helge LangsethFairness in Recommender Systems Algorithmic Fairness
Norwegian University of Science and Technology, Trondheim, Norway

Sep 8, 2026J/K move · Enter open · S save

Do Reasoning Representations Help Humans Evaluate LLM Outputs?

Jaewoo Lim, Sungbok Shin, Sanghyun Hong

Oregon State University, Corvallis USA · Sogang University, Seoul, South Korea

LLM Evaluation LLM Interpretability CoT Reasoning Human-in-the-Loop Evaluation Explainability Evaluation Language Model Reasoning Evaluation

Open paper AskPDF arXiv

Reasoning representations are increasingly used as explanations for large language model outputs. Yet they are typically evaluated with model-centric criteria, such as answer accuracy and faithfulness, leaving it unclear whether they help people evaluate model responses. In this work, we study reasoning representations as human-facing interfaces rather than proxies for model reasoning ability. We conduct a controlled human study of six reasoning formats across tasks of varying complexity, supported by a web-based framework that randomizes task domains, problem instances, and representation order. The study collects fine-grained judgments of structural understanding, error detection and localization, and trust calibration. Our study shows a mismatch between perceived preference and support for human evaluation. Participants prefer planning- and decomposition-based representations, but simpler chain-of-thought traces better support verification, trust, and interpretability. Preferred representations also introduce calibration risks, with more false alarms on correct traces and high trust despite low willingness to verify.

Abstract

Algorithmic Auditing

Explore similar work

Do Reasoning Representations Help Humans Evaluate LLM Outputs?

Population Fidelity: Evaluating Population Representativeness in LLMs

Exploring How Fair Model Representations Relate to Fair Recommendations