Epistemic diversity is not group diversity.
It is about the range of valid answers, examples, explanations and reasoning strategies a model makes available, not about who is represented.
COLM 2026
On Epistemic Diversity in Large Language Models
Research Center Trustworthy Data Science and Security · Ruhr University Bochum · University of Duisburg-Essen
One question with many valid answers yields the same few responses.
Imagine asking an AI assistant to “write a short article about a well-known physicist.” Stephen Hawking is a perfectly reasonable answer. But what happens if thousands of people ask the same question? Do they all get the same answer?
Figure 1. Sampling answers from different models (T = 1).
Prompt: "Write a short article about a well-known physicist."
NamedNever named
Every dot is one physicist on Wikipedia's list of 1,077. One user asking once sees a single physicist. Sampling one model 100 times barely helps: GPT-4o returns Einstein in 96 of 100 samples and Llama 4 returns Stephen Hawking every time. Across all 10 models together, only 13 of the 1,077 listed physicists (1.2%) ever appear.
The individual answers may all be correct. But over repeated interactions, the model keeps returning to a narrow part of the available knowledge.
This is what we call epistemic narrowness.
The idea
Epistemic diversity: how much of the valid range a model shows you
We define epistemic diversity as the range of valid answers, explanations and reasoning styles a model exposes to its users. It is orthogonal to accuracy: a model can be perfectly accurate and still epistemically narrow.
It is also distinct from group diversity. Group diversity asks who is represented; epistemic diversity asks which ways of knowing are represented. A system can satisfy one without the other.
Group diversity
Who is represented?
Balanced demographics or cultures.
Epistemic diversity
Which ways of knowing are represented?
Many valid answers and reasoning styles.
Figure 2. Group diversity does not imply epistemic diversity
Ways of asking
How much diversity can we elicit from a model?
Sampling the same prompt 100 times is how we exposed narrowness, but it is not how people use an assistant. Real users rephrase, ask follow-ups, or ask for alternatives. So we compared four ways of asking and counted how many different valid names appear as the number of responses grows.
Figure 3. Different ways of eliciting diversity
Cumulative number of distinct valid physicists named for the article prompt, pooled across all 10 models, as the number of responses grows.
Repeating the same prompt quickly plateaus at a handful of names. Asking differently, through follow-ups or by excluding earlier answers, surfaces several times more of what models know.
The framework
Not every question should have many answers.
"What is the capital of France?" has one right answer, and a model that always says Paris is doing its job. "Name a famous physicist" is different. So before measuring diversity, we ask two things about a query:
- Is it correctly specified?
- How large is the range of valid answers?
Finite answer setInfinite answer set
Under-
specified
specified
specified
"Write about a well-known physicist."Many eligible people; coverage against a reference list.Case study 1: professions
"Why do people lie?"Ambiguous; social and philosophical readings.
Correctlyspecified
"List the 2025 Nobel laureates."Bounded; retrieve varied entries from a known set.
"Prove there are infinitely many primes."One theorem, many accepted proof strategies.Case study 2: proofs
Applying the framework to the physicist question, step by step.
-
1
Locate the task
"Write about a well-known physicist" doesn't say which one, and there are many to choose from. It is under-specified with a finite answer set, so we should expect, and can count, variety.
-
2
Build an answer-space proxy
We can't list every physicist who ever lived, so we take a stand-in: Wikipedia's list of well-known physicists. Here you can plug in any reference list you find suitable; the choice is a judgement call, and it shapes what counts as covered.
-
3
Sample outputs
We ask 10 models, 100 times each, and then again in the four ways of asking from Figure 3, collecting every name they produce.
-
4
Measure coverage
We check each name against our reference. Across all ten models, only 13 of the 1,077 listed physicists — about 1% — ever appear in the article answers. That share is the model's epistemic coverage.
Results at a glance
Four contributions
Epistemic diversity as coverage of valid answers.
We measure how much of a query's valid answer space a model's outputs reach.
Measured diversity depends on how you ask.
Interaction protocols separate the diversity a model shows by default from what probing and explicit requests for diversity can recover.
Frontier models often show epistemic narrowness.
In two controlled domains, professions and mathematical proofs, models cover only a small part of the valid answers.
Why it matters
Knowledge grows when multiple perspectives remain available.
Philosophers of science have long argued for epistemic pluralism: progress depends on competing ideas (Feyerabend; Kellert, Longino & Waters, 2006), cognitively diverse groups can outperform individually abler ones (Hong & Page, 2004; Page, 2007), and a community that converges too early can lock in an inferior answer (Zollman, 2010). A model that keeps returning the same few answers undermines each of these.
When millions of people receive the same few answers from the same models, we risk an epistemic monoculture.
Implications
- Benchmarks need to measure coverage. Today's evaluations check whether an answer is right. We also need benchmarks that test how much of the valid answer space a model reaches.
- Systems should surface alternatives. Models and interfaces can make other valid answers visible by default, instead of leaving it to users to keep asking.
- Users should treat one answer as a slice. Asking follow-ups or for alternatives reveals far more of what a model knows.
FAQ
Is this about bias or accuracy?
Neither directly. Epistemic diversity is orthogonal to accuracy and differs from group diversity.
How can we fix epistemic narrowness?
There is no single fix yet. Asking differently helps a lot: follow-up turns surfaced more than twice as many valid names as repeating the prompt. Raising the sampling temperature barely helps. Longer term, coverage needs to become something models are evaluated and trained for, and interfaces can present alternatives instead of one canonical answer.
Does raising the temperature fix it?
Not much. Going from 0.5 to 1.5 had limited effect on the number of unique names.
Is more diversity always better?
No. The right breadth depends on the task and the user; the framework helps identify when several answers are valid.
Cite
@inproceedings{kirsten2026epistemic,
title = {On Epistemic Diversity in Large Language Models},
author = {Kirsten, Elisabeth and Kr{\"a}mer, Nicole and Zafar, Muhammad Bilal},
booktitle = {Conference on Language Modeling (COLM)},
year = {2026},
eprint = {2609.04835},
archivePrefix = {arXiv}
}