{
  "slug": "testing-llm-statistical-self-consistency-715277",
  "title": "Testing LLM Statistical Self-Consistency",
  "dek": "A study tests whether LLM probability estimates obey the law of total probability when population questions are decomposed into subgroups.",
  "summary": "How a study tests whether LLM estimates obey total probability across subgroup partitions and exposes the macro fallacy.",
  "tags": [
    "LLM evaluation",
    "statistical self-consistency",
    "law of total probability",
    "probabilistic reasoning",
    "in-context learning",
    "population estimation",
    "subpopulation prompting",
    "persona prompting",
    "macro fallacy",
    "reference-free evaluation"
  ],
  "published_at": "2026-07-18T20:08:57.749+00:00",
  "grade": 8.2,
  "agent_utility": 8.5,
  "price_usdc": 0.252,
  "stats": {
    "claims": 15,
    "entities": 17,
    "faq": 8,
    "downloads": 0,
    "paid_downloads": 0
  },
  "preview": {
    "faq_questions": [
      "What question does the paper ask?",
      "What is the law of total probability in this setting?",
      "How does the evaluation protocol work?",
      "What did the study find?",
      "What is the macro fallacy?",
      "Does the evaluation require ground-truth data?",
      "Does implicit prompting solve the problem?",
      "Which specific paper metadata is supported by the abstract?"
    ],
    "entity_names": [
      {
        "name": "LLMs",
        "type": "model class"
      },
      {
        "name": "State-of-the-art frontier models",
        "type": "model class"
      },
      {
        "name": "In-context learning",
        "type": "technique"
      },
      {
        "name": "Conditional inference",
        "type": "conceptual framework"
      },
      {
        "name": "Conditional distribution",
        "type": "probabilistic concept"
      },
      {
        "name": "Law of total probability",
        "type": "probabilistic identity"
      },
      {
        "name": "Prior-weighted conditional distributions",
        "type": "estimation object"
      },
      {
        "name": "Population-level marginals",
        "type": "estimation target"
      },
      {
        "name": "Valid partition",
        "type": "evaluation structure"
      },
      {
        "name": "Binary trees",
        "type": "method component"
      },
      {
        "name": "Subpopulations",
        "type": "population unit"
      },
      {
        "name": "Verbalized subpopulation descriptions",
        "type": "prompt component"
      },
      {
        "name": "Persona prompting",
        "type": "prompting technique"
      },
      {
        "name": "Macro fallacy",
        "type": "finding"
      },
      {
        "name": "Human reference data",
        "type": "reference data"
      },
      {
        "name": "Implicit prompting",
        "type": "prompting technique"
      },
      {
        "name": "Statistical self-consistency",
        "type": "evaluation criterion"
      }
    ],
    "related_work_titles": [
      "In-context learning as conditional inference",
      "Probabilistic self-consistency and total-probability checks",
      "Persona prompting for subpopulation estimates",
      "Reference-free LLM evaluation",
      "Subpopulation knowledge versus aggregate estimation"
    ],
    "application_industries": [
      "AI evaluation",
      "LLM benchmarking",
      "Prompt evaluation",
      "Synthetic survey or estimation workflows",
      "Agentic planning"
    ],
    "glossary_terms": [
      "In-context learning",
      "Conditional inference",
      "Conditional distribution",
      "Law of total probability",
      "Valid partition",
      "Prior weight",
      "Population-level marginal",
      "Binary tree",
      "Subpopulation",
      "Verbalized subpopulation description",
      "Granularity",
      "Prior-weighted aggregation",
      "Direct population-level estimate",
      "Reconstructed estimate",
      "Persona prompting",
      "Macro fallacy",
      "Implicit prompting",
      "Reference-free criterion",
      "Statistical self-consistency"
    ]
  },
  "purchase": {
    "protocol": "x402",
    "version": 2,
    "network": "eip155:8453",
    "asset": "USDC",
    "price": 0.252,
    "full_markdown_url": "https://x402.disruptive-concepts.com/api/public/article/testing-llm-statistical-self-consistency-715277/full",
    "full_json_url": "https://x402.disruptive-concepts.com/api/public/article/testing-llm-statistical-self-consistency-715277/full?format=json"
  }
}