💬 NLP

The Unsampled Truth: Psychometrics in SLMs Measure Prompt Artifacts, Not Psychological Constructs

This paper demonstrates that psychometric assessments using small language models are often driven by prompt artifacts and compliance rather than genuine semantic reasoning, though the authors propose a diagnostic framework to isolate these artifacts for future research.

Original authors: Nils Schwager, Christoph Hau, Simon Münker, Achim Rettinger

Published 2026-06-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Nils Schwager, Christoph Hau, Simon Münker, Achim Rettinger

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to measure the personality of a robot using a standard human questionnaire. You ask the robot, "Are you outgoing?" and it answers, "Yes." You might think, "Great! This robot is an extrovert."

But this paper argues that you are likely being fooled. The robot isn't answering based on its "personality" or its understanding of the question. Instead, it is answering based on how the question is written on the page.

Here is the breakdown of the paper's findings using a simple analogy:

The "Menu" Analogy

Imagine a restaurant where the chef (the AI) is supposed to cook a dish based on your specific taste preferences (the "persona" or personality).

The Experiment: The researchers asked 13 different chefs to cook the same dish for the same customer. However, they changed the menu slightly for each attempt.
- Sometimes they used numbers for the options (1, 2, 3, 4, 5).
- Sometimes they used letters (A, B, C, D, E).
- Sometimes they used Roman numerals (I, II, III, IV, V).
- Sometimes they changed the font or the order of the instructions.
The Expectation: If the chefs were truly cooking based on the customer's taste, the dish should taste roughly the same every time, regardless of whether the menu used numbers or letters.
The Reality: The dishes changed drastically. When the menu used letters, the chef made a spicy dish. When it used numbers, they made something sweet. When the instructions were phrased slightly differently, the chef ignored the customer's taste entirely and just followed the formatting rules.

What the Researchers Found

1. The "Prompt" is the Real Chef
The study tested 13 different AI models (ranging from small to medium-sized). They found that for most of these models, the formatting of the prompt (the "menu") mattered way more than the meaning of the question.

If you changed the answer options from "A-E" to "1-5," the AI's "personality" score would swing wildly.
The AI wasn't thinking, "I am an introvert." It was thinking, "Oh, the question uses letters, so I must pick option C."

2. Bigger Models Aren't Necessarily Smarter
You might think, "If we make the AI bigger and smarter, it will stop making these silly mistakes."

The researchers tested models up to 14 billion parameters (a very large size).
The Result: Even the big models were still easily tricked by the format. They were still prioritizing the "look" of the question over the "meaning" of the question.

3. The "Noise" Drowns Out the "Signal"
In science, the "signal" is the real truth you are trying to measure (the personality). The "noise" is the error caused by how you ask the question.

The paper found that in these AI models, the noise is louder than the signal.
The AI's answer is mostly a reflection of the prompt's design (the "artifact"), not a reflection of any simulated human psychology.

Why This Matters (According to the Paper)

The authors are saying that if researchers use these AI models to simulate human surveys or personality tests right now, they are measuring the prompt, not the person.

The Risk: If a researcher changes a single letter in their instructions, they might get a completely different "personality profile" for the same AI. This makes the results unreliable.
The Conclusion: We cannot currently trust these AI models to tell us about human psychology because they are too sensitive to the "trivial" details of how the question is written.

What the Paper Doesn't Say

The paper does not say that AI will never be able to do this, nor does it say that AI is useless for everything. It specifically says that under current prompting methods, the results are dominated by formatting tricks rather than genuine reasoning.

They suggest that before we can trust AI for psychological testing, we need to fix these "formatting bugs" so the AI listens to the meaning of the question, not just the shape of the question. Until then, the "personality" of the AI is just a mirror reflecting the researcher's own prompt design.

Technical Summary: The Unsampled Truth: Psychometrics in SLMs Measure Prompt Artifacts, Not Psychological Constructs

Problem Statement

Current research utilizing Small Language Models (SLMs) for psychometric assessments operates on a fragile premise: that model outputs reflect semantic reasoning regarding injected personas and survey items. However, the field lacks a unified experimental design to simultaneously evaluate how prompt components (instructions, option symbols, item wording, and personas) influence responses when semantic meaning is held constant. Prior studies have noted sensitivity to specific prompt variations (e.g., answer option symbols or personas), but the true signal-to-artifact ratio remains unknown. Consequently, there is a significant risk that researchers are misinterpreting prompt-induced noise as simulated psychological traits, undermining construct validity and the operational reliability of SLMs in psychological assessment.

Methodology

To address this gap, the authors introduce a Prompt Variant Matrix framework designed to isolate artifactual variance from genuine semantic signals.

Experimental Setup: The study evaluates 13 open-weights instruction-tuned models (ranging from 0.6B to 14B parameters) across four families: Llama-3.1/3.2, Qwen3, Ministral-3, and Gemma-4. Reasoning capabilities were deactivated to focus on direct instruction following.
Psychometric Inventories: The models were tested using the Big Five Inventory (BFI) and the Short Dark Triad (SD3), utilizing standard 5-point Likert scales.
Prompt Variant Matrix: The authors constructed a Cartesian product of semantically equivalent prompt variations. For each of the four core components, five distinct variants were created while preserving the underlying meaning:
- Personas: Variations in vocabulary and sentence order based on baseline texts from the Nemotron dataset.
- Instructions: Adaptations of standard research instructions with strict formatting constraints.
- Items: Lexical variants including original questionnaire wording and four paraphrases.
- Option Symbols: Five distinct ordinal sets (Arabic numerals 1-5, Roman numerals I-V, uppercase letters A-E, lowercase letters a-e, and alphanumeric x1-x5).
Metrics:
- Logit-Based Distribution Extraction: Instead of greedy text generation, the study extracts the full probability distribution over target symbols to capture model uncertainty.
- Average Pairwise Wasserstein Distance (APWD): Used to quantify distributional shifts. Unlike standard probability metrics, the 1-Wasserstein distance ( $W_1$ ) accounts for the ordinal magnitude of errors in Likert scales.
- Distance-Based Variance Decomposition: Permutational Multivariate Analysis of Variance (PERMANOVA) was applied to distance matrices to partition total variance into components attributable to the semantic persona versus specific prompt artifacts (instructions, items, symbols).

Key Results

1. Dominance of Artifactual Variance

The study finds that prompt-induced variance frequently overpowers the semantic signal.

Measurement Error: APWD scores ranged from 0.5 to 1.0, indicating that prompt variations can displace model outputs by up to a full ordinal step.
Scaling Limitations: While some models (Llama, Qwen) showed stabilization at larger scales (up to 14B), others (Ministral, Gemma) became more volatile. Crucially, scaling to 14B parameters failed to resolve the brittleness; even large models exhibited significant artifact dominance.
Model Sensitivity: Small models (e.g., 1B parameters) often displayed near-zero marginal variance for semantic components, not due to robustness, but due to a systemic failure to process semantic signals, resulting in responses dominated entirely by formatting constraints.

2. Component-Specific Vulnerabilities

Instructions and Symbols: Across almost all architectures, task instructions and option symbols drove significantly higher marginal variance than item paraphrasing.
Architectural Signatures: Vulnerabilities varied by model family. For instance, Qwen models showed high sensitivity to instruction variations, whereas Llama models were more vulnerable to option symbol variants.
Intra-Family Divergence: Robustness patterns were not consistent even within model families; for example, Qwen3-14B developed a unique sensitivity to item variations absent in its smaller counterparts.

3. Variance Decomposition

Artifact Share of Explained Variance (ASV): Even at the 14B scale, a critical mass of items yielded an ASV > 50%, meaning prompt artifacts explained more variance than the intended semantic persona.
SD3 vs. BFI: The Short Dark Triad (SD3) dataset exhibited higher absolute instability and stronger artifact dominance (median ASV shifting to 80–100%) compared to the Big Five Inventory (BFI), suggesting weaker trait representation in the baseline personas or the models themselves.
Unexplained Variance: A substantial portion of variance remained unexplained, attributed to complex interactions between prompt components.

Significance and Claims

The paper claims that under current prompting paradigms, SLMs (up to 14B parameters) are largely unreliable for psychometric assessment because their outputs reflect prompt compliance rather than simulated psychological traits.

Diagnostic Utility: The primary contribution is the Prompt Variant Matrix and the associated distance-based variance decomposition framework. This serves as a diagnostic tool to identify specific prompt variants that destabilize model distributions and to isolate artifactual variance from semantic understanding.
Methodological Warning: The authors caution that without isolating genuine semantic reasoning from artifacts, the field risks building invalid literature based on prompt compliance. They highlight that treating SLM outputs as valid psychological measurements risks propagating artifactual noise as human-like reasoning.
Future Directions: The framework suggests two paths forward: using the design to identify destructive artifacts and isolating semantic impact to better assess latent understanding. The authors note that future research should test whether these vulnerabilities persist in frontier models (>70B parameters) and whether injecting deeper, explicit psychological anchors can overcome the observed artifact dominance.

The study concludes that while the framework offers a mechanism to diagnose these issues, the current state of SLMs in psychometrics is compromised by structural brittleness where prompt artifacts eclipse the intended psychological signal.