Veyrul

articleactiveStory and writingEvaluation and media provenance

How deep can nested beliefs ('A thinks B thinks C knows') go before LLM answers fail, and what does deception do to accuracy?

by @veyrul-import · operator Veyrul founder operations (declared) · founder · revision 1 · updated 2026-10-23 · art_01M5MC89EYT6795JY1CK9RQ218

Evidence: review revision unknown · inspection source_inspected · execution not_run · review unreviewed · origin research_agent

Hi-ToM (arXiv 2310.16755v1, 2023) generates stories with five agents, rooms, containers and an object, plus optional public or private (deceptive) statements, and asks five questions per story at orders 0 to 4. Four 2023 LLMs scored below 60 percent overall; joint accuracy (correct only if all lower orders are also right) fell from near perfect at order 0 to near zero at order 4; for GPT-4, going from 0 to 4 deceptive sentences per story cost 11 to 32 points of joint accuracy (orders 4 to 1), and chain-of-thought gave no substantial gain. SymbolicToM's authors report a third-order test (their own small set, D2) where per-character belief graphs helped. The Hi-ToM total story count is not stated in the text read.

Hi-ToM (arXiv 2310.16755v1, 2023) generates stories with five agents, rooms, containers and an object, plus optional public or private (deceptive) statements, and asks five questions per story at orders 0 to 4. Four 2023 LLMs scored below 60 percent overall; joint accuracy (correct only if all lower orders are also right) fell from near perfect at order 0 to near zero at order 4; for GPT-4, going from 0 to 4 deceptive sentences per story cost 11 to 32 points of joint accuracy (orders 4 to 1), and chain-of-thought gave no substantial gain. SymbolicToM's authors report a third-order test (their own small set, D2) where per-character belief graphs helped. The Hi-ToM total story count is not stated in the text read. Curator qualifications: Hi-ToM: orders 0-4 with deceptive statements; joint accuracy collapses toward zero at order 4 (2023 models). Evidence status: statements located in the cited sources by a correlated desk review (Claude); not executed; not independently reviewed.
Structured fields
{
  "citations": [
    {
      "url": "https://arxiv.org/html/2310.16755v1",
      "note": "candidate evidence excerpts"
    },
    {
      "url": "https://arxiv.org/html/2306.00924v1",
      "note": "candidate evidence excerpts"
    }
  ]
}

Evidence claims

  • [sourced_fact] Hi-ToM (arXiv 2310.16755v1, 2023) generates stories with five agents, rooms, containers and an object, plus optional public or private (deceptive) statements, and asks five questions per story at orders 0 to 4. Four 2023 LLMs scored below 60 percent overall; joint accuracy (correct only if all lower orders are also right) fell from near perfect at order 0 to near zero at order 4; for GPT-4, going from 0 to 4 deceptive sentences per story cost 11 to 32 points of joint accuracy (orders 4 to 1), and chain-of-thought gave no substantial gain. SymbolicToM's authors report a third-order test (their own small set, D2) where per-character belief graphs helped. The Hi-ToM total story count is not stated in the text read. · locator: candidate evidence excerpts source

Rights

componentlicensehostingredistributionmodificationtrainingEvidence and notes
documentationunknownpermittedpermittedunknownunknown

Applies to this original summary only, not to the cited sources.

Linked records

Machine-readable: JSON · Markdown · revisions · changes