PaperlayerView exact revision on arXiv

Source-linked AI summary

On Epistemic Diversity in Large Language Models

Elisabeth Kirsten, Nicole Krämer, Muhammad Bilal Zafar

arXiv:2609.04835v1cs.CL

TL;DR

LLM evaluation often overlooks whether systems preserve access to multiple valid answers, explanations, and reasoning routes. The paper defines and measures epistemic diversity through answer-space coverage and interaction protocols, finding that frontier models frequently narrow valid answer spaces to canonical subsets. It concludes that accuracy, fairness, and alignment do not fully capture behavior in knowledge-intensive use, while noting important limits in answer-space and validity measurement.

  • Problem

    Epistemic diversity has not been systematically formalized for LLMs, despite their growing use for explanation, tutoring, writing, and inquiry.

  • Method

    The paper defines epistemic diversity as coverage over valid answer spaces and evaluates it across different interaction protocols and two controlled domains.

  • Results

    Frontier LLMs often exhibit epistemic narrowness, repeatedly concentrating on canonical individuals or proof strategies despite accepted alternatives.

  • Takeaways & Limitations

    Epistemic diversity adds an evaluation dimension concerning which knowledge becomes salient and available to users, beyond accuracy, fairness, and alignment.

  • Takeaways & Limitations

    Answer spaces are often incomplete or ambiguous, and validity and equivalence measurements rely on task-specific proxies and approximate automated annotation.

Abstract

from arXiv · show

Large language models (LLMs) are increasingly used not only to retrieve information, but to answer questions, explain, teach, and support inquiry. In such settings, evaluation cannot be exhausted by accuracy or alignment alone. A system may give a correct answer while still narrowing users' access %to knowledge. to alternative valid answers, explanations, or reasoning routes. Drawing on the broader notion of epistemic diversity in philosophy and social epistemology, we formalize it in the context of LLMs as the range of valid answers, explanations, and reasoning routes that an LLM exposes to users. We argue that epistemic diversity is a useful evaluation dimension for settings where LLMs are used to support knowledge-intensive tasks. We propose a preliminary framework for conceptualizing and measuring epistemic diversity in LLMs, and operationalize it in two domains. We find that frontier LLMs often exhibit epistemic narrowness, repeatedly collapsing large valid answer spaces onto small canonical subsets. These findings suggest that LLM evaluation should move beyond accuracy-oriented paradigms and treat epistemic diversity as an important dimension of model capability.

1 Introduction

LLMs used as knowledge tools should be evaluated not only for correctness, but also for the range of valid answers, explanations, and reasoning routes they make available. The paper frames this range as epistemic diversity and finds that frontier models often narrow large valid answer spaces to canonical subsets.

  • Epistemic diversity concerns the answers, explanations, and reasoning strategies an LLM makes available, beyond whether its output is correct.
  • A model can be accurate yet epistemically narrow when it repeatedly presents only a few canonical routes despite multiple valid alternatives.
  • Epistemic diversity is distinct from group diversity, because representing multiple groups does not ensure that the knowledge available across those groups is surfaced.
  • The framework treats meaningful diversity as coverage over valid answer spaces rather than arbitrary variation, stylistic differences, noise, or hallucination.
  • The paper evaluates accessible diversity under ordinary prompting and contrasts it with diversity recoverable through more deliberate exploration.
  • Across two settings, frontier LLMs often concentrate on a small subset of valid outputs, although stronger prompting recovers some additional diversity while leaving a substantial gap.

2 From Group Diversity to Epistemic Diversity

Existing AI diversity work primarily examines representation across demographic, cultural, or political groups, whereas epistemic diversity concerns coverage of valid knowledge outputs. The paper defines epistemic coverage and illustrates narrowness through proofs and physicist-generation tasks.

  • Existing AI diversity perspectives: Group-based diversity research asks whether demographic, cultural, or political groups are represented or treated fairly in model outputs.
  • Epistemic diversity: Epistemic diversity instead asks what range of valid answers, explanations, examples, and reasoning strategies a knowledge-generating system makes available.
  • Relationship between diversity notions: Group and epistemic diversity can overlap, but group representation does not guarantee that a system surfaces the knowledge distributed across those groups.
  • Epistemic coverage: Epistemic coverage is the portion of a universal valid-answer set represented in the answers produced by a knowledge system.
  • Epistemic coverage: Coverage is narrow relative to threshold τ when the intersection between the universal answer set and produced answers falls below τ.
  • Examples of epistemic narrowness: For infinitely many primes, the paper notes at least 200 known proofs, while models repeatedly return Euclid’s proof; in the physicist task, models named about 6% of listed individuals.

3 A Measurement Framework for Epistemic Diversity

The measurement framework addresses the lack of systematic formalization for epistemic diversity in LLMs. It characterizes tasks and answer spaces, then measures diversity through multiple interaction protocols that distinguish default exposure from deliberate recovery of alternatives.

  • The framework responds to the absence of a systematic LLM formalization for epistemic diversity by structuring evaluation around validity, multiplicity, and answer-space size.
  • Task characterization determines validity criteria, typology placement, and whether measurement targets individual answers, reference-set elements, or broader answer classes.
  • Finite domains can use curated reference sets and enumeration-based coverage, whereas open-ended domains may require response types, strategies, or conceptual clusters.
  • Interaction protocols: Accessible diversity measures valid responses surfaced by ordinary prompting, limited resampling, or repeated exposure to the same prompt.
  • Interaction protocols: Latent diversity measures alternatives recoverable through controlled probing that iteratively requests new answers while excluding previously produced ones.
  • Interaction protocols: Multi-turn and multi-output protocols model more natural requests for alternatives through follow-up exploration or several answers in one turn.

4 Empirical Illustration

The study evaluates epistemic diversity in professions and mathematical proofs, finding that frontier LLMs repeatedly surface narrow subsets of valid answers. Explicit diversity-seeking interactions recover more alternatives than ordinary sampling, though concentration and invalid outputs remain.

  • Experimental settings: The experiments compare professions, with finite underspecified answer spaces, and mathematical proofs, with multiple valid reasoning strategies.The study prompts a range of frontier LLMs on both datasets.
  • Observed narrowness: Models often concentrate on canonical answers, covering only 2–24% of profession-specific Wikipedia reference lists across all models.In proofs, models also frequently return only one strategy despite accepted alternatives.
  • Observed narrowness: Accessible sampling surfaces only 1.9–4 unique valid names in ten profession-task samples, averaged over prompts.This narrow accessible diversity is evaluated against alternative interaction protocols.
  • Diversity-seeking interactions: At ten responses, latent probing, multi-turn requests, and multiple-response prompts recover more unique names than accessible sampling for every tested model.Median gains across prompts are +85, +83, and +51 names, respectively.
  • Failure modes: Iterative prompting leaves 22% of rounds without a new valid name, while invalid names reach 71% for Qwen 2.5.Models also continue favoring salient individuals and leaving much of the answer space uncovered.

5 Related Work

Related work distinguishes epistemic diversity from demographic or viewpoint representation and motivates evaluating coverage of valid answers, explanations, and reasoning strategies. This paper extends existing LLM diversity research by analyzing underlying answer spaces rather than relying only on output similarity or factual variation.

  • Foundations in Social Sciences: The paper builds on philosophical and social-epistemological arguments that diverse viewpoints and methods improve robustness and guard against collective blind spots.These traditions treat diversity as epistemically valuable, not only socially inclusive.
  • Diversity in LLM Evaluation: Existing LLM benchmarks often assume a single correct answer, collapsing valid response spaces and treating diversity as noise or variability.Related work also studies homogenization, minority viewpoints, and output variability.
  • Epistemic Diversity in LLMs: Prior LLM studies measure factual, linguistic, or semantic variation, but this paper extends analysis to reasoning strategies and conceptual breadth.The framework distinguishes finite and infinite answer spaces.
  • Epistemic Diversity in LLMs: Similarity-based analyses of open-ended outputs may miss the structure of epistemic differences.This work instead reasons directly about answer-space coverage and formalizes question types and valid-answer structure.

6 Conclusion, Discussion & Limitations

The paper presents epistemic diversity as a measurable evaluation dimension and finds that frontier LLMs often foreground canonical answers while leaving valid alternatives less accessible. It argues for preserving meaningful alternatives when tasks call for exploration, while recognizing important measurement and scope limitations.

  • Conclusion: The framework defines epistemic diversity as the range of valid answers, explanations, examples, concepts, and reasoning strategies made available to users.It is operationalized across ten models and two datasets.
  • Discussion: Across professions and proofs, models concentrate on canonical individuals or proof strategies despite the existence of accepted alternatives.Examples include repeatedly presenting Einstein or Euclid’s proof as the default.
  • Discussion: Epistemic diversity should not always be maximized; canonical answers remain appropriate, while alternatives matter for exploration, learning, comparison, and pluralism.The paper treats diversity and consistency as complementary desiderata.
  • Discussion: Personalization can appropriately narrow a valid answer space, but systems should distinguish helpful constraint satisfaction from unnecessary epistemic closure.Educational and exploratory settings may still benefit from alternatives.
  • Discussion: The framework distinguishes low accessible diversity from low latent diversity and separates these from invalidity or hallucination failures.These patterns indicate different intervention points, including prompting and decoding.
  • Limitations: The study relies on incomplete task-specific proxies, difficult validity judgments, and an automated annotation pipeline that approximates rather than definitively measures answer spaces.Epistemic diversity also depends on prompting, decoding, personalization, and interaction design.
  • Conclusion: The results indicate that epistemic diversity is conceptually important and practically measurable beyond accuracy, fairness, and alignment.The central evaluation question includes whether models preserve enough epistemic openness for understanding.

A Details on Experimental Setups

The study operationalizes epistemic diversity across professions and mathematical proofs, using repeated model outputs, task-specific coverage, and interaction protocols to distinguish accessible from recoverable diversity.

  • Datasets: The study uses professions and mathematical proofs to examine distinct valid-answer spaces.Professions represent an underspecified, finite setting; proofs represent a correctly specified setting with multiple solution strategies.
  • Professions dataset: Professions measure diversity through distinct named individuals across six fields, using Wikipedia lists as proxies for valid answers.Because the proxy lists are subsets of the true answer space, measured coverage is an upper bound on true coverage.
  • Prompting: For professions, prompts vary generation format while holding the underlying question fixed, testing whether concentration persists across common formats.The formats include bio, quote, article, poem, and story, alongside naming prompts.
  • Mathematical proofs dataset: Mathematical proofs measure diversity by clustering outputs into distinct proof categories for problems with multiple known strategies.The study curates N = 9 problems and collects k = 10 completions per problem per model.
  • Interaction protocols: The accessible protocol samples k = 100 independent completions with identical wording and decoding parameters, without diversity-encouraging instructions.The study also evaluates latent, multi-turn, and multi-output protocols for the Professions dataset.
  • Measurement: Protocol-specific diversity applies a task-specific coverage function to valid answers or answer classes recovered from a query.The procedure distinguishes meaningful diversity from surface variation and separates default-accessible diversity from diversity revealed through stronger probing.

C What constitutes sufficient diversity?

Sufficient epistemic diversity is context-dependent: it depends on the valid answer space and question structure, while the coverage threshold remains difficult to set universally.

  • Coverage threshold: The threshold τ denotes a minimum acceptable level of epistemic coverage, with higher values requiring more distinct valid perspectives.The authors state that τ cannot be universally prescribed in practice.
  • Determinants of diversity: The valid answer space’s size shapes how much diversity should be expected, with large or infinite spaces making canonical-only outputs potentially narrow.Question structure also matters because underspecified prompts often admit broader valid interpretations.
  • Trade-offs: High diversity can produce irrelevant or redundant outputs, whereas a threshold set too low can reinforce epistemic monocultures.The universal answer set is often implicit, further complicating calibration.
  • Interpretation: The framework therefore treats τ as flexible rather than universal and uses it to reason about coverage relative to plausible answer spaces.The paper reports that frontier models under-represent valid alternatives even under conservative assumptions.

D Need for Epistemic Diversity in Real Chat Datasets

An analysis of 500 WildChat queries tests whether real user questions fit the framework, finding that many allow multiple valid answers and are underspecified.

  • Answer-space structure: 75% of queries have more than one valid answer, including n = 288 with multiple countable answers and n = 89 with infinitely many.These results indicate that real-world questions frequently occupy non-singleton answer spaces.
  • Question specification: More than 75% of the queries are labeled underspecified, meaning they lack enough detail for a concrete answer.The annotation scheme separately classifies the number of valid answers and whether the query is correctly specified or underspecified.
  • Implication: The WildChat analysis reinforces the need for evaluation that accounts for epistemic diversity in real user queries.The paper presents the real-chat analysis as evidence that its framework applies beyond the two purpose-built datasets.
  • Dataset and annotation: The study randomly selects 500 filtered WildChat queries and annotates them with the framework’s taxonomy.Queries were English, non-toxic, directed to GPT-4 models, 15–100 characters long, and ended with question marks.
  • Annotation criteria: The annotation instructions define multiple valid responses as cases where different people could answer the same query differently.The judge also assigns a rationale for both the answer-count and specification decisions.

E.1 Empirical Observations

Across professions and proof tasks, models repeatedly concentrate on small subsets of valid answers and reasoning strategies. This epistemic narrowness persists across sampling and model comparisons, while stronger probing recovers only part of the available diversity.

  • Models often return the same individuals or a single proof strategy despite many valid alternatives.Llama 4 named Stephen Hawking in all 100 completions for one physicist prompt, while proof outputs often converged on one strategy.
  • Only 2–24% of reference names were covered across models, leaving large portions of the valid answer space unexplored.
  • Different models systematically favor different valid reasoning routes, with Gemini preferring induction and GPT models preferring Euler’s theorem.
  • Repeated sampling produces sublinear diversity gains, and temperature changes have little effect on the explored answer space.Higher temperature changes framing and prose more than the individual described.
  • For a substantial fraction of queries, one individual accounts for most responses, with Llama 4 and Gemini 3 showing the greatest concentration.
  • Latent probing recovers more names than accessible sampling, but coverage remains incomplete and varies substantially by model.Gemini 3 produces 88 names on average, compared with 69 for Llama 3.1, 26 for Qwen 2.5, and 31 for Qwen 3.

E.3 Failure Mode Analysis

The iterative diversity-seeking protocol fails through invalid generation and stagnation. Invalid outputs and repeated exclusions increasingly prevent the protocol from adding new valid names.

  • Qwen 2.5 produced invalid names in 71% of judged outputs, and over 90% of invalid names were absent from the Wikipedia reference sets.
  • Across models, 22% of rounds stalled, mainly because the model repeated an excluded person or the extractor found no person.Stagnation increased with round index, reaching 42% for Qwen 3 and 5% for Gemini 3.

E.4 Diversity Under Explicit Exploration

Explicit requests for alternatives substantially increase diversity beyond ordinary sampling, although no protocol fully eliminates concentration or covers the entire valid answer space. Interaction design therefore strongly affects which answers users encounter.

  • At ten responses, multi-turn interaction recovered 8.2 unique valid names on average, compared with 2.9 for accessible sampling and 7.8 for latent probing.
  • All three diversity-seeking protocols exceeded accessible sampling for every tested model, while multi-turn and latent probing recovered similar amounts overall.
  • Pooled across models, multi-turn prompting recovered 511 distinct names, versus 345 for multi-output, 275 for latent probing, and 227 for accessible sampling.
  • Accessible prompting had an approximately 11.5% top-1 share, compared with approximately 3–4% for the three diversity-seeking protocols.
  • Even explicit exploration leaves some reference-space regions uncovered and continues to favor highly salient individuals.

E.5 Statistical Analysis

The analysis evaluates diversity across repeated samples, interaction protocols, and prompt variants using paired statistical comparisons and prompt-level uncertainty estimates. Results indicate robust protocol differences, limited temperature effects, and measurable annotation agreement, with uncertainty intervals restricted to prompt variation.

  • Accessible diversity was averaged over 36 profession–format prompts with bootstrap 95% confidence intervals based on resampling prompts.
  • Varying temperature from 0.5 to 1.5 had a limited effect on diversity, while repeated generations showed cumulative gains that generally saturated or grew sublinearly.
  • The bootstrap intervals capture variation across prompts but not uncertainty in the reference list or annotation judge, which were treated as fixed.
  • All three diversity-seeking protocols exceeded accessible sampling for all ten models in paired tests, with p = 0.002.The Wilcoxon signed-rank test used models as the unit of analysis, n = 10.
  • Independent authors agreed on 98 of 100 annotations, and final labels matched the LLM judge on all 100 cases.
  • Prompt phrasing produced only small differences: “any” yielded 32.7 unique valid names on average versus 33.2 for the original “a” prompt.
Loading 2609.04835v1…