# Blind Men and the Elephant: Probing the Epistemic Myopia of LLMs under Long-Tail Divergent Knowledge

arXiv: https://arxiv.org/abs/2608.28478 · HTML: https://mlsanity.com/paper/2608.28478/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 22 of its 39 references that are in the map.
- First posted on arXiv 28 Aug 2026.

## Builds on (most cited first, 22 of 22)

- [When Not to Trust Language Models: Investigating Effectiveness of Parametric and Non-Parametric Memories](https://mlsanity.com/paper/2212.10511/) (2022)
- [DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence](https://mlsanity.com/paper/2606.19348/) (2026)
- [Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research](https://mlsanity.com/paper/2402.00159/) (2024)
- [Large Language Models Struggle to Learn Long-Tail Knowledge](https://mlsanity.com/paper/2211.08411/) (2022)
- [DoReMi: Optimizing Data Mixtures Speeds Up Language Model Pretraining](https://mlsanity.com/paper/2305.10429/) (2023)
- [AmbigQA: Answering Ambiguous Open-domain Questions](https://mlsanity.com/paper/2004.10645/) (2020)
- [Knowledge Conflicts for LLMs: A Survey](https://mlsanity.com/paper/2403.08319/) (2024)
- [Rephrasing the Web: A Recipe for Compute and Data-Efficient Language Modeling](https://mlsanity.com/paper/2401.16380/) (2024)
- [SituatedQA: Incorporating Extra-Linguistic Contexts into QA](https://mlsanity.com/paper/2109.06157/) (2021)
- [Rich Knowledge Sources Bring Complex Knowledge Conflicts: Recalibrating Models to Reflect Conflicting Evidence](https://mlsanity.com/paper/2210.13701/) (2022)
- [ConflictBank: A Benchmark for Evaluating the Influence of Knowledge Conflicts in LLM](https://mlsanity.com/paper/2408.12076/) (2024)
- [Resolving Knowledge Conflicts in Large Language Models](https://mlsanity.com/paper/2310.00935/) (2023)
- [WikiContradict: A Benchmark for Evaluating LLMs on Real-World Knowledge Conflicts from Wikipedia](https://mlsanity.com/paper/2406.13805/) (2024)
- [Nemotron 3 Ultra: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning](https://mlsanity.com/paper/2606.15007/) (2026)
- [T-NER: An All-Round Python Library for Transformer-based Named Entity Recognition](https://mlsanity.com/paper/2209.12616/) (2022)
- [DYNAMICQA: Tracing Internal Knowledge Conflicts in Language Models](https://mlsanity.com/paper/2407.17023/) (2024)
- [Scaling Laws for Fact Memorization of Large Language Models](https://mlsanity.com/paper/2406.15720/) (2024)
- [Recycling the Web: A Method to Enhance Pre-training Data Quality and Quantity for Language Models](https://mlsanity.com/paper/2506.04689/) (2025)
- [Who's Who: Large Language Models Meet Knowledge Conflicts in Practice](https://mlsanity.com/paper/2410.15737/) (2024)
- [RePro: Training Language Models to Faithfully Recycle the Web for Pretraining](https://mlsanity.com/paper/2510.10681/) (2025)
- [Consensus or Conflict? Fine-Grained Evaluation of Conflicting Answers in Question-Answering](https://mlsanity.com/paper/2508.12355/) (2025)
- [Generating Pretraining Tokens from Organic Data for Data-Bound Scaling](https://mlsanity.com/paper/2605.17849/) (2026)

## Abstract (arXiv)

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.