Blind Men and the Elephant: Probing the Epistemic Myopia of LLMs under Long-Tail Divergent Knowledge
ElephantBench constructs a graph-based knowledge probe to diagnose LLMs' ability to recall multiple divergent accounts of long-tail facts, revealing biases towards dominant sources.
Key Findings
Methodology
This work develops ElephantBench, a graph-structured probe built from low-exposure web documents containing factual conflicts. It employs knowledge point clustering and named entity recognition to identify candidate document pairs. An LLM-based edge classifier determines support or conflict relations, forming a document graph. From conflict edges, local subgraphs are sampled, and the LLM generates multi-account QA pairs grounded in source evidence. Source verification via web search and human review ensures answer authenticity. In evaluation, models attempt to recall all verified accounts without source access, with metrics including complete recall (C), partial recall (P), failure (F), and conditional completeness (K). The framework assesses whether models can remember diverse, conflicting accounts of long-tail facts.
Key Results
- The strongest models recover both accounts on only 52.4% of questions, indicating significant epistemic myopia. Scaling model size and reasoning improves recall but does not eliminate bias towards dominant accounts. Source exposure imbalance correlates with recall bias; models tend to recall the more prevalent account more reliably. Increasing exposure to minority accounts enhances recall completeness, yet biases persist across models and domains.
- Across 32 models, larger models and those with reasoning capabilities show improved performance, but incomplete recall remains prevalent. Notably, models perform better in domains like people and entertainment, whereas government and civic domains exhibit lower recall rates. These findings highlight the influence of source data distribution on model memory, emphasizing the need for balanced training data.
- This methodology offers a reproducible diagnostic tool for analyzing epistemic bias in parametric memory, guiding data curation and model development to better capture diverse knowledge sources.
Significance
This research advances the understanding of how LLMs encode and recall divergent accounts of facts, especially in the long tail of less common knowledge. By transforming naturally occurring web disagreements into a structured evaluation framework, it reveals the biases inherent in current models and datasets. The graph-based approach provides a scalable, source-traceable means to diagnose and mitigate epistemic myopia, addressing a core challenge in aligning LLMs with real-world knowledge diversity. Its implications extend to improving factual robustness, fairness, and transparency in AI systems, fostering more comprehensive and balanced knowledge representations.
Technical Contribution
The paper introduces a two-stage graph construction pipeline leveraging knowledge point clustering and entity recognition to efficiently identify candidate document pairs. An LLM-based edge classifier then labels support and conflict relations, enabling the automatic assembly of a source-grounded document graph. From conflict edges, local subgraphs are sampled, and LLMs generate multi-account QA pairs grounded in source evidence, which are validated through web search and human review. This end-to-end pipeline automates the detection of source disagreements and facilitates closed-book evaluation of models’ ability to recall multiple divergent accounts, a novel approach in long-tail knowledge assessment.
Novelty
This work is the first to systematically convert naturally occurring web-based factual disagreements into a closed-book, multi-account knowledge probe. Unlike prior benchmarks focusing on single-answer recall or open-book evidence reasoning, ElephantBench emphasizes the model's capacity to remember and reproduce multiple, conflicting verified accounts without access to sources. Its graph-based construction and source verification processes provide a scalable, source-traceable framework for long-tail knowledge evaluation, addressing a critical gap in understanding epistemic biases.
Limitations
- The approach relies heavily on the diversity and accuracy of web sources; if sources are biased or incomplete, the evaluation may not fully reflect the model's true knowledge capacity. Additionally, the current framework primarily targets factual conflicts and may not extend easily to more nuanced or subjective knowledge. The computational cost of source verification and human review limits scalability for larger datasets. Moreover, the method does not directly address how to mitigate biases, only diagnose them.
Future Work
Future research will focus on integrating multi-modal sources, expanding the scope to include temporal and contextual facts, and developing methods to actively balance source exposure. Enhancing model training with diverse, source-verified data could reduce biases. Additionally, extending the framework to evaluate models in multi-task, multi-modal settings will be crucial for real-world deployment. Investigating bias mitigation strategies based on source exposure analysis is also a promising direction.
AI Executive Summary
Large language models (LLMs) have revolutionized natural language understanding and generation, yet their knowledge often reflects biases towards dominant sources, especially in the long tail of less common facts. Traditional benchmarks evaluate whether models can recall a single canonical answer, but they overlook the complexity of real-world information, which frequently involves multiple conflicting accounts from diverse sources. To address this gap, the authors introduce ElephantBench, a novel graph-based knowledge probe that systematically captures and evaluates the model's ability to remember multiple, divergent accounts of long-tail facts.
ElephantBench constructs a document graph from web sources with low exposure, identifying factual conflicts through an edge classifier trained on support and conflict relations. From these conflicts, local subgraphs are sampled, and LLMs generate multi-account QA pairs grounded in source evidence. These pairs are rigorously validated via web search and human review, ensuring high-quality, source-traceable data. In the evaluation phase, models are tested in a closed-book setting, attempting to recall all verified accounts without access to source documents.
Experimental results across 32 models reveal a significant epistemic myopia: even the strongest models recall both accounts on only about half of the questions. While increasing model size and reasoning capabilities improves recall, biases towards the most prevalent sources persist. Analysis shows that source exposure imbalance heavily influences recall bias—more exposure to the majority account correlates with higher recall of that account, whereas minority accounts require more exposure to be fully remembered.
These findings underscore the importance of balanced data exposure and highlight the limitations of current models in capturing knowledge diversity. The graph-based pipeline offers a scalable, reproducible framework for diagnosing and understanding epistemic biases, guiding future efforts in data curation and model training. Overall, ElephantBench provides a critical step toward more truthful, fair, and comprehensive AI systems capable of representing the full spectrum of human knowledge.
Deep Dive
Abstract
Factual question answering (QA) typically assumes a single canonical answer, obscuring whether large language models (LLMs) retain divergent accounts of long-tail facts. To address this gap, we introduce ElephantBench, a closed-book knowledge probe comprising 1,094 questions generated through an auditable graph-based pipeline. The pipeline retrieves related documents from a low-exposure web corpus, identifies naturally occurring disagreements, and converts them into multi-account QA records. Each answer is verified against the originating documents and authoritative public web sources and is then reviewed by human annotators. Across 32 models, even the strongest model recovers both accounts on only 52.4% of questions, while on nearly all remaining questions it recalls one account but omits the other. Scaling model size and inference-time reasoning improve recall but do not eliminate this incompleteness. Corpus analysis further shows that exposure imbalance favors the dominant account, whereas greater minority-side exposure is associated with more complete recall. These findings establish ElephantBench as a reproducible knowledge probe for diagnosing epistemic myopia in parametric memory. More broadly, our graph-based benchmark construction pipeline provides an efficient and scalable way to turn long-tail corpora into source-traceable knowledge probes, supporting efforts to evaluate and advance the epistemic rigour of next-generation LLMs. Code is available at https://github.com/Tencent/ElephantBench.