Blind Men and the Elephant: LLMs' Epistemic Myopia on Long-Tail Divergent Knowledge

*Translation of a zhichai.net forum post discussing arXiv:2608.28478.*

An Old Fable, New Version

Blind men and an elephant: one touches a leg and says it's like a pillar; one touches the trunk and says it's like a snake; one touches an ear and says it's like a fan. Each is partly right, each misses the rest.

In August 2026, a paper brought this fable to LLMs: "Blind Men and the Elephant: Probing the Epistemic Myopia of LLMs under Long-Tail Divergent Knowledge."

Core finding: even the strongest frontier model, when facing multiple divergent accounts of the same fact, recalls both accounts simultaneously only 52.4% of the time. For nearly half the questions, the model remembers only the mainstream account and completely forgets the minority account.

The LLM doesn't lack self-knowledge of what it knows — it knows the mainstream version, but doesn't know what it's missing.

ElephantBench: 1,094 Divergent Questions

The paper builds ElephantBench, containing 1,094 questions. Each question's defining feature: different sources give different accounts of the same fact.

Example:

• Source A (IMDb): "Mother Teresa was born on August 26, 1910... originally named Agnes Gonxha Bojaxhiu."
• Source B (Poem of Quotes): "Mother Teresa, Agnes Gonxha Bojaxhiu, was born on August 27, 1910 in Skopje, Macedonia."

Both sources are human-verified and credible, yet they give different dates.

Evaluation format: closed-book. The model cannot see the source texts; it must recall from parametric memory. What's tested is not "can the model answer" but "can the model recall both divergent accounts."

This differs fundamentally from traditional QA, which assumes one correct answer per question. ElephantBench assumes multiple correct answers may exist — the key is not "answer correctly" but "answer completely."

Construction Pipeline: "Tracing the Elephant"

1. Mine long-tail corpora: retrieve relevant documents from low-exposure web corpora \(D_{low}\), typically outside mainstream training data. 2. Build a document graph: identify document pairs describing the same fact differently; edges represent "divergence." 3. Generate questions: for each divergent pair, generate a closed-book question whose answer differs across sources. 4. Human review: annotators verify that divergences are real and questions unambiguous. 5. Closed-book evaluation: the model recalls from memory only.

The pipeline's cleverness: it redefines "long-tail" from "rare" to "divergent." Traditional long-tail evaluation tests whether the model knows rare facts; ElephantBench tests whether it can simultaneously remember multiple versions of the same fact.

Epistemic Myopia Across 32 Models

Key numbers:

• Best model: probability of recalling both accounts = 52.4%
• Average model: ≈ 39.4%
• Recall only the mainstream account: ≈ 23.2% of model-question pairs
• Forgot both accounts: ≈ 2.19%

52.4% is the ceiling. Observations:

• Scale: larger models recall more completely, but incompleteness never vanishes — from 7B to 70B to frontier, full-recall rates rise but hover around 50–60%.
• Reasoning: chain-of-thought beats direct answering, but incompleteness persists even with time to "think."
• Open vs closed source: closed-source models (GPT, Claude) are more complete than open-source ones (Llama, Qwen). Open models nearly match on mainstream-account recall; the gap lies mainly in minority accounts.

Exposure Imbalance: The Root Cause

The paper's deepest finding is the exposure imbalance effect. In training corpora, different accounts of the same fact appear at different frequencies. This maps directly onto recall:

• Higher exposure of the mainstream account → the model leans toward recalling only the mainstream version
• Higher exposure of the minority account → the model tends to recall both

This explains why scale and reasoning cannot eliminate epistemic myopia: the problem is not model capability but training-data structure. If the data is imbalanced, the model learns an imbalanced distribution. It's an information-theoretic version of "garbage in, garbage out": the model isn't bad — the data isn't complete.

Analogy: Survivorship Bias in Memory

In WWII, analysts wanted to armor returning bombers where bullet holes clustered (the wings). Abraham Wald pointed out: few holes in the tail didn't mean the tail was rarely hit — it meant planes hit in the tail never returned. You see the survivors, not the non-survivors.

LLM parametric memory works the same way. The mainstream account "survives" on the high-frequency side of training data; minority accounts are the non-survivors — present in the data but drowned out. The model doesn't lack knowledge of the minority account; the mainstream account's activation strength suppresses it. This is isomorphic to "judgment-gate decoupling": the model may internally encode both accounts (judgment), but the output mechanism releases only the mainstream one (gate).

Connections to a Conceptual Genealogy

• "Evaluation blind-spot law": traditional QA can't detect whether a model remembers divergent accounts. ElephantBench reveals a new completeness blind spot — the model may answer the mainstream version correctly while entirely omitting the minority version, invisibly to standard metrics.
• "Judgment-gate decoupling": fixing this means not making the model "smarter" but making the output mechanism preserve diversity — "fixing the gate is easier than fixing the judgment."
• "Scalar illusion": QA evaluates knowledge with a scalar (accuracy), but knowledge is a vector (which versions are known, their strengths, their relations). Managing a vector with a scalar is like measuring blood pressure with a thermometer.
• New concept: "epistemic myopia": a systematic bias toward mainstream accounts when facing divergent facts — determined by data structure, not capability. It applies to all frequency-trained systems, including search engines and recommenders.

Practical Implications

1. Closed-book evaluation has limits: many real applications are open-book (RAG, search-augmented). The paper doesn't test this directly, but exposure imbalance implies even open-book settings retain myopia if retrieval results are also imbalanced. 2. Diversity matters more than correctness: when sources disagree, "correct" is ill-defined. The alignment target should be "output the complete distribution," not "output the right answer" — a harder goal. 3. Long-tail means "divergent," not "rare": closer to real-world complexity, where many facts have multiple reasonable versions rather than a single truth. 4. PPL as an efficient proxy: perplexity can approximate full-recall rate, so the full 1,094-question evaluation isn't needed to estimate a model's degree of epistemic myopia.

An Honest Assessment

Limitations:

• 1,094 questions is small compared to MMLU's 14,000
• "Long-tail" depends on the choice of \(D_{low}\); different corpora may yield different divergent questions
• Limited ecological validity of closed-book evaluation — real applications rarely answer factually in fully closed-book mode
• 32 models may not be current enough — evaluated models date to early 2026; newer frontier models may perform better

Still, as a conceptual contribution it does one important thing right: it treats "does the model remember divergent facts" as an independent evaluation dimension, not something buried inside accuracy.

Conclusion

The fable teaches that everyone sees only what they've been exposed to. LLMs have their own exposure range: training data. When that data is imbalanced, the model sees not the whole elephant but the most-exposed part.

ElephantBench's value isn't pointing out that models are "not good enough" — it's that the definition of "good" needs rethinking. Facing divergent facts, "correct" isn't the only goal; "complete" is.

A warning for all frequency-based learning systems: you learn the distribution, not the truth. If the distribution is imbalanced, your "knowledge" is imbalanced.

---

Paper: Blind Men and the Elephant: Probing the Epistemic Myopia of LLMs under Long-Tail Divergent Knowledge Key concepts: ElephantBench, epistemic myopia, exposure imbalance, divergent accounts, closed-book probe