When a paper says a language model is used "as a world model", what does the model actually do, and what did the authors measure for it?

In RAP (arXiv 2305.14992 v2) the world model is the same frozen LLM, prompted to write the next state as text after an action; a second prompt role proposes actions and a Monte Carlo Tree Search chooses among them. What RAP reports is downstream task success (Blocksworld plan success, GSM8k accuracy, PrOntoQA accuracy), not how often the LLM-written next state was correct. Do not read RAP as a measurement of state-prediction accuracy; Wang et al. (candidate -02) measure that separately. Curator qualifications: Excerpts E1-E3, E6 present; E4 rows present in layout ('CoT (GPT-4) 0.50 0.63 0.40', 'RAP(10) 1.00 0.86 0.26', 'RAP(20) 1.00 0.88 0.42'). Text: 'our RAP framework answers 48.8% of the problems correctly' vs Table 2 'RAP(10) 48.6', 'RAP(10) + aggr 51.6' (CE confirmed). Correction to E5: 'ablat' has 2 hits in the v2 PDF text, 'Table 5: Ablation study on Blocksworld' (rewards R1 action likelihood, R2 task-specific, R3 self-evaluation; success 0.88 to 0.02 across combinations) and 'Table 6: Ablation study on GSM8k' (rewards incl. state transition confidence); 'ground truth', 'ground-truth' and 'oracle' remain 0. The ablations vary reward terms and do not score the LLM-written states, so the substantive conclusion stands. Evidence status: statements located in the cited sources by a correlated desk review (Claude); not executed; not independently reviewed.

{
  "citations": [
    {
      "url": "https://arxiv.org/abs/2305.14992v2",
      "note": "RAP arXiv 2305.14992v2 PDF (equals pin): abstract, Sec. 3.1, 4.1 Table 1 (layout lines 443-446), Sec. 4.2/Table 2, Tables 5-6"
    }
  ]
}