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

AI summary

Overview

Research area: Natural Language Processing / large language model evaluation, specifically closed-book factual question answering over long-tail knowledge and cross-source factual disagreement.

Technical level: Intermediate. The paper is readable without deep mathematical background, though familiarity with LLM benchmarking, parametric memory, and knowledge-conflict evaluation helps.

Scope: The paper introduces ElephantBench, a 1,094-question closed-book benchmark built from a low-exposure web corpus, and uses it to measure whether 32 language models can recall all independently verified accounts of a disputed long-tail fact rather than only the dominant one.

What This Paper Is About

Factual QA benchmarks usually assume one canonical answer per question, which hides cases where reputable sources report incompatible accounts of the same fact. The authors mine these naturally occurring disagreements out of the low-exposure portion of a web corpus, verify each side against the original documents and independent public sources, and turn them into a closed-book probe. The goal is to test whether a model's parametric memory retains all verified accounts of a long-tail fact, or whether it exhibits what the authors call "epistemic myopia" by holding only the dominant account.

Key Contributions

1. A diagnostic probe of parametric memory. Instead of collapsing performance into a single accuracy number, ElephantBench reports separate metrics for whether a model recalls a low-exposure fact at all (complete, partial, and failed recall, plus conditional completeness K = C/(C+P)) and whether it recovers every verified account of that fact.
2. A reusable two-stage graph-based construction pipeline. The pipeline uses knowledge-point clustering (SuperGPQA taxonomy) and named-entity/event retrieval (T-NER) to narrow the candidate pair space, then uses an LLM edge classifier to label pairs as none, support, or conflict. Each conflict edge seeds a local subgraph from which matched named-entity and clue-based questions are generated. The authors report this reduces the pairwise candidate space to at most 6.6% of exhaustive comparison, or at least a 15.2x reduction in pairwise LLM judgments.
3. A corpus-grounded analysis of memory completeness. By counting documents supporting each side of a disagreement, the paper connects exposure asymmetry in the corpus to a specific memory failure mode, linking the probe to data curation rather than only model ranking.
4. A verified, traceable benchmark release. All 1,094 items are traceable to source documents, externally verified by a web agent, and reviewed by human annotators; the benchmark, code, and data are released publicly.

Main Findings

• Even the strongest model is incomplete. Across 32 models, Kimi-K3 achieves the highest complete recall at 52.38%, followed by Gemini-3.1-Pro at 50.37% and GPT-5.5 at 50.18%. The three strongest models have failed recall rates of only 2.19% to 2.65%, so outright failure to recall the fact is rare. The dominant failure is partial recall, which ranges from 45.25% to 47.44% for those same models.
• Scale improves recall but not completeness. Across open-weight models, average complete recall is 5.48% for models under 10B parameters, while the best model above 1T parameters reaches 52.38%. Within Qwen3.5, scaling from 2B to 397B raises complete recall from 1.65% to 32.27% and reduces failed recall from 81.35% to 8.50%, but partial recall rises from 17.00% to 59.23%.
• Reasoning helps some models, hurts others. Enabling reasoning raises complete recall by 13.99 percentage points for GPT-5.6-Sol and 12.89 percentage points for GPT-OSS-120B. For Qwen3.5 it adds 0.73 to 4.39 percentage points from 9B to 397B, but reduces complete recall by 0.64 and 0.37 percentage points at 2B and 4B, where the model can treat one account as consensus and omit the other.
• Exposure asymmetry drives the failure mode. A one-standard-deviation increase in exposure to the more frequently reported account is associated with a 14.18-percentage-point increase in partial recall and a 10.17-percentage-point decrease in failed recall. The same increase for the less frequently reported account is associated with a 15.13-percentage-point increase in complete recall and a 15.41-percentage-point decrease in partial recall.
• Performance varies sharply by domain. Complete recall is highest for People, Organizations, and Events (38.7%) and Digital Content and Entertainment (32.1%), drops to 22.7% for News Events and Public Information, and is lowest for Government and Civic Affairs (12.0%) and Consumer Products and Services (6.2%). In these two lowest domains, partial recall dominates at 55.6% and 62.1%.
• An oracle over all 32 models still cannot fully close the gap. Greedily adding models from a start of Kimi-K3 raises complete recall from 52.4% to 81.2% and lowers failed recall from 2.4% to 0, saturating after 23 configurations. The remaining 18.8% (206 questions) stay partial: no configuration recovers all verified accounts.
• Question wording is not the driver. Comparing named-entity and clue-based formulations across eight frontier models, the mean difference in complete recall is 0.65 percentage points (95% CI [-1.05, 2.31]; p = 0.470), with four models favoring each formulation.
• Perplexity tracks graded outcomes. For Qwen3.6-27B, higher conditional answer perplexity bins contain fewer complete responses and more failures, suggesting PPL can serve as a cheaper proxy for generation-based evaluation.
• The automatic judge aligns with humans. Across four frontier models, human annotators and the GPT-5.6-Sol judge show 90.13% to 93.36% exact agreement, with Cohen's kappa from 0.815 to 0.877, and complete recall differing by 1.44 percentage points on average.

Methodology in Plain English

The authors start from the discarded portion of a web corpus rather than the good part. Using the DCLM fastText quality classifier with a threshold, they split documents into a retained high-quality set and a filtered low-exposure set, then work only with the filtered set on the reasoning that rare facts are less likely to be reinforced during pretraining.

They build a document graph over those documents. Qwen3.6-27B assigns each document a knowledge-point label from the SuperGPQA taxonomy, producing within-cluster candidate pairs; T-NER retrieval adds cross-cluster pairs that share a normalized entity or event mention. An LLM then classifies each candidate pair's relation as none, support, or conflict, where conflict means the two documents address the same subject–attribute pair but report incompatible accounts.

Each conflict edge seeds a local subgraph containing the two endpoints and their support neighbors. From each subgraph, GPT-5.6-Sol generates a question about the disputed attribute plus the full set of distinct verified answers, in two matched forms: a named-entity question that names the target and a clue-based question that identifies it through at least two related attributes. Both forms share the same answer set.

Validation happens in three stages: an LLM checks that each answer is explicitly supported by at least one document in the subgraph; a web agent with WebSearch and WebFetch finds independent public evidence (prioritizing Wikipedia) for every answer on every item; and human reviewers inspect both the original documents and the retrieved evidence, rejecting entity mismatches, unsupported answers, answer leakage, and question–fact mismatches. The pipeline produced 8,254 candidate questions from 4,127 conflict subgraphs, leaving 1,094 QA pairs after validation and deduplication, distributed across 22 knowledge fields.

At evaluation time, models receive only the question, with no sources, graph structure, reference answers, or external tools. An LLM-as-a-judge assigns each response to complete, partial, or failed recall using a four-step rubric. The authors evaluate 26 open-weight models and 6 proprietary systems, with reasoning enabled by default and effort set to high where configurable, using an output budget of 32,768 tokens.

Why This Matters

Impact on research. The paper separates two questions that single-answer QA conflates: whether a model remembers a rare fact, and whether it remembers everything verified about that fact. That separation gives evaluation a new failure category to measure, and the construction pipeline offers a way to turn any low-exposure corpus into a source-traceable probe rather than requiring hand-authored questions.

Real-world applications:

• Retrieval-augmented and knowledge-grounded systems. Knowing that models default to the dominant account on disputed facts informs how much weight to give retrieved minority sources in pipelines for journalism, legal research, or fact-checking.
• Pretraining and fine-tuning data curation. The finding that minority-side exposure, more than majority-side exposure, tracks complete recall suggests a concrete lever for data mixtures rather than just model scaling.
• High-stakes factual domains. The weakest domains in the results — Government and Civic Affairs (12.0% complete recall) and Consumer Products and Services (6.2%) — are areas where giving a single confident answer can mislead users about vote counts, product prices, or regulatory details.
• Evaluation tooling. Conditional perplexity as a cheaper proxy, plus the reported judge-human agreement (kappa 0.815 to 0.877), makes rigorous completeness evaluation more practical to run repeatedly.

Industry relevance. The result that frontier models top out near 50% complete recall, and that an oracle over all 32 configurations still leaves 18.8% of questions partial, is a direct signal that scaling alone does not solve divergence handling. The reported scale-dependent reasoning effect — gains for large models, regressions at 2B and 4B — also cautions against assuming inference-time reasoning uniformly improves factual completeness.

Future Directions

• Broader disagreement coverage. The authors note the benchmark's 1,094 questions across 22 fields cannot exhaustively represent long-tail knowledge or the diversity of ways sources can disagree, and suggest expanding scale, domain coverage, and disagreement types.
• Mitigating minority-account omission. The specific regressions the authors observe, in which a reasoning-enabled model treats one account as consensus, point to targeted work on surfacing less salient accounts rather than generic capability scaling.
• Better exposure proxies. Because pretraining corpora are generally undisclosed, the paper's exposure findings are associations from one public corpus. A model-specific or otherwise stronger proxy for factual exposure would let these results be tested more directly.
• Cheaper, more reliable completeness scoring. Conditional perplexity aligns with graded outcomes for Qwen3.6-27B, but whether it holds across model families and scales remains open, as does evaluating models released after this study.

Target Audience

This paper is most useful to LLM evaluation and benchmark researchers, pretraining data-curation teams, and engineers building factual QA, retrieval-augmented, or fact-checking systems who need to know how models behave when reputable sources disagree. It is also relevant to readers tracking knowledge-conflict and long-tail knowledge work, since it positions closed-book completeness as a distinct axis from the open-book conflict detection that most prior benchmarks measure.

Authors’ 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.