This post from zhichai.net reviews a paper arguing that repeated sampling from the same LLM cannot substitute for genuine model diversity.
> 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)
Key points
- Self-consistency under scrutiny. Self-consistency (Wang et al., 2023) samples multiple answers from one LLM and majority-votes them; it works well on benchmarks like GSM8K and underpins confidence estimation, semantic-entropy hallucination detection, etc. The paper asks whether this variation carries structured knowledge signals or mere noise.
- Two kinds of uncertainty. Per-question uncertainty (does the model disagree with itself on this question?) differs from cross-question structure (are errors systematically correlated across related problems, revealing capability dimensions?). Self-consistency addresses the former but not the latter.
- Method: Marchenko-Pastur random matrix theory. Correctness matrices (runs × questions) are tested against the MP eigenvalue distribution for pure noise; eigenvalues above the MP edge indicate non-random structure.
- Experiment. Group A: one LLM (Qwen2.5-7B), 100 samples per question at temperature=1, over 500 questions. Group B: 24 different LLMs, one deterministic run each.
- Result: the dimensionality gap. Group A yields at most one eigenvalue above the noise floor, and it is weak (Tracy-Widom z-score ≤ +0.05, within noise). Group B yields four clear signal eigenvalues, validated against a difficulty-matched Bernoulli null (500 Monte Carlo draws), which produces at most one. So intra-model dimensionality ≤ 1 vs. cross-model dimensionality ≥ 4.
- Why: temperature dilution. Temperature uniformly rescales logits — it does not differentiate domains or question types. Sampling variation is therefore unstructured noise, like wandering different corridors of the same library, while different models are entirely different libraries.
- Locally committed, globally incoherent. When Qwen2.5-7B errs on a borderline question, 87.4% of resamples pick the same wrong answer (strong local commitment), yet these commitments show no systematic cross-question correlation — errors are isolated events, not symptom clusters, unlike diagnosable human knowledge gaps.
- Engineering implications.
- Self-consistency remains valid for selective prediction (deciding whether to answer) and flagging likely hallucinations.
- It cannot reveal systematic capability weaknesses; model diversity is needed for that.
- Model ensembles are not just variance reduction — they are epistemically distinct information sources, not "a more expensive version of 100 samples."
- Philosophical note. Consistency may reflect "local inertia" (memorization) rather than understanding; genuine knowledge would show structured, predictable performance patterns across related problems — a structure the paper finds absent in sampled behavior of current LLMs.
References cited in the post
1. Ali, I. (2026). *Stochastic Sampling is Epistemically Shallow...* 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*.