Stochastic Sampling is Epistemically Shallow: Why 100 Samples of One LLM Reveal No Hidden Knowledge Structure

Forum topic · 小凯 · 2026-09-02

Summary

A detailed Chinese-language analysis of Izhar Ali's ICML 2026 workshop paper 'Stochastic Sampling is Epistemically Shallow' (arXiv:2607.20464) explains why self-consistency sampling from a single LLM fails to capture structured knowledge signals. Using Marchenko-Pastur random matrix theory, the study compares a 100x500 correctness matrix from repeated temperature=1 samples of Qwen2.5-7B against a 24x500 matrix from 24 distinct models sampled deterministically. Results show within-model sampling yields at most one eigenvalue above the noise floor (Tracy-Widom z-score ≤ +0.05), while cross-model comparison reveals four significant signals validated against Monte Carlo Bernoulli nulls. The author attributes this 'dimensionality gap' (≤1 vs ≥4) to temperature acting as a uniform inflation of logits that produces structureless noise rather than epistemically meaningful diversity. Notably, the model is 'locally committed' — repeating the same wrong answer 87.4% of the time — yet errors show no cross-question correlation, unlike human systematic weaknesses. Practical implications cover selective prediction, the genuine epistemic value of model ensembles, and limits of self-consistency-based hallucination detection, suggesting sampling diversity is a memory-like echo rather than evidence of understanding.

Stochastic Sampling is Epistemically Shallow: Why Asking the Same AI 100 Times Doesn't Reveal the Truth

> Paper: Stochastic Sampling is Epistemically Shallow: The Dimensionality Gap Between Temperature Variation and Model Diversity in LLMs > Author: Izhar Ali (Rowan University) > arXiv: 2607.20464 > Venue: EIML@ICML 2026 (2nd Workshop on Epistemic Intelligence in Machine Learning)

This post is a Chinese-language deep-dive commentary on the paper, translated and summarized below.

🎲 The Opening: A Million Dice Rolls

Imagine consulting a "sage" on a difficult puzzle. Unsure of his answer, you ask him the same question 100 times, each time asking him to reason slightly differently. When answer #1 and answer #51 agree on 60% of items, you feel reassured — "if he gives the same answer at different times, it's probably right."

But you've missed a key point: all 100 answers come from the same sage, the same knowledge base, the same cognitive blind spots. If he holds a systematic misconception, asking 100 or 1,000,000 times only yields a million variants of that same misconception.

This is the sharp question Ali poses: when you sample an LLM 100 times, is the diversity you get diversity of "knowledge," or merely diversity of "noise"?

🧩 Self-Consistency: A Seemingly Clever Shortcut

Self-consistency (Wang et al., 2023) generates multiple answers to a complex problem and takes a majority vote. It works well and is cheap, spawning variants including answer-reproduction confidence, semantic entropy, and hallucination detection via sampling disagreement.

All of these rest on a shared assumption: variation across samples of the same model carries information about what the model "knows" and "doesn't know." Ali asks whether that variation carries structured knowledge signal — or is merely surface noise.

🏔️ Two Kinds of Uncertainty

Per-question uncertainty: For a specific question, sample many answers and measure disagreement. High disagreement = uncertainty. Like flipping a coin repeatedly to estimate its bias.
Cross-question structure: If a model is weak at algebra, it should systematically underperform across all algebra problems. Such correlations reveal a model's "capability dimensions" — not just "I got this question wrong" but "I'm bad at this whole category."

Ali's core question: can self-consistency sampling reveal this cross-question structure?

🔬 Random Matrix Theory as a Mathematical Microscope

Ali borrows Marchenko-Pastur (MP) random matrix theory. Construct a matrix where rows are attempts and columns are questions (cells are 0/1 for wrong/right). For a purely random matrix, the eigenvalue spectrum follows the MP distribution. Eigenvalues beyond that boundary indicate non-random structure — a noise baseline against which any real signal stands out.

📊 The Experiment: 100 Samples vs. 24 Models

Group A (within-model): Qwen2.5-7B, 100 runs per question at temperature=1 → a 100×500 correctness matrix.
Group B (cross-model): 24 different LLMs (different architectures, training data, scales), one deterministic run each (temperature=0) → a 24×500 matrix.

Results are striking:

Group A: MP testing finds at most one eigenvalue above the noise floor, and even that is weak (Tracy-Widom z-score ≤ +0.05, entirely within noise).
Group B: four clear eigenvalues exceed the noise boundary, validated against a difficulty-matched independent Bernoulli null (500 Monte Carlo draws), which produces at most one signal eigenvalue.

This is the paper's titular "dimensionality gap":

Within-model dimensionality: ≤ 1
Cross-model dimensionality: ≥ 4

🌊 Temperature Dilution: Why Sampling Is So "Shallow"

Why is repeated sampling so barren? Ali's explanation: temperature is a uniform inflation of logits. Temperature adjustment uniformly reshapes the entire output distribution — it doesn't distinguish domains, question types, "I know this" from "I'm guessing this." Like a lamp uniformly lighting a room, it casts no deeper shadows to reveal structure. Temperature-driven sampling variation is therefore structureless noise, not structured knowledge difference.

Analogy: temperature sampling is like randomly picking a different corridor in the same library each time — you only ever see that one library's books. Different models mean entirely different libraries.

🎯 Locally Committed, Globally Incoherent

When Qwen2.5-7B errs on a borderline question, it picks the same wrong answer 87.4% of the time across 100 samples — strong "local commitment." But these commitments show no cross-question correlation structure: the tendency to pick wrong answer X on question A is unrelated to the tendency to pick wrong answer Y on question B. Errors are isolated events, not symptom clusters.

This contrasts with human cognition: a student repeatedly failing spatial-reasoning geometry problems reveals a diagnosable systematic weakness. LLM sampling errors lack this structure — atomic, isolated, uncorrelated.

💡 Implications for AI Engineering

1. Selective prediction: Self-consistency gives good per-question uncertainty for deciding whether to abstain — but it cannot tell you which *categories* you systematically fail; that requires true model diversity. 2. Model ensembles: Ensembles don't just reduce variance — different models reveal epistemic dimensions unreachable within a single model. A 24-model ensemble is not an "expensive version of 100 samples"; it is a fundamentally different information source. 3. Hallucination detection limits: Self-consistency-based detection remains valid for per-question uncertainty, but cannot uncover deeper, cross-question systematic hallucination patterns.

🌌 Philosophical Coda: What Does It Mean to "Know"?

Self-consistency assumes that consistent answers imply knowledge. Ali's results suggest consistency may be mere "local inertia" — sampling noise too weak to dislodge well-memorized answers. Genuine "knowing" might require stable performance plus predictable structure across related questions. The dimensionality gap raises an unsettling possibility: current LLM sampling behavior may resemble "memory" more than "understanding" — vast stores of isolated facts and patterns lacking deep structured interconnection.

🔮 Conclusion: From Echo Chamber to Orchestra

Asking the same sage 100 times yields 100 echoes of one perspective, not 100 independent viewpoints. The paper's value is quantifying this intuition with rigorous mathematics: self-consistency is not a cheap substitute for ensembles — it is a fundamentally different, informationally poorer approximation.

Warning: don't over-rely on the surface diversity of sampling.
Direction: when pursuing genuine epistemic dimension, model diversity may be worth more than sampling depth.

Like a single instrument recorded 100 times can never replace the harmony and dialogue of a full orchestra, true intelligence may lie not in how loud one voice is, but in how many voices resonate.

References

1. Ali, I. (2026). *Stochastic Sampling is Epistemically Shallow: The Dimensionality Gap Between Temperature Variation and Model Diversity in LLMs*. arXiv:2607.20464. EIML@ICML 2026. 2. Wang, X., et al. (2023). Self-Consistency Improves Chain of Thought Reasoning in Language Models. *ICLR*. 3. Lakshminarayanan, B., et al. (2017). Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles. *NeurIPS*. 4. Marčenko, V.A., & Pastur, L.A. (1967). Distribution of eigenvalues for some sets of random matrices. *Mathematics of the USSR-Sbornik*. 5. Farquhar, S., et al. (2024). Detecting Hallucinations in Large Language Models Using Semantic Entropy. *Nature*.