How accurate was GPT-4 at predicting the next state of a text game, and which kinds of transitions (action-driven vs environment-driven, changing vs unchanged) failed?

Wang et al. (arXiv 2406.06485 v1) built ByteSized32-State-Prediction (76,369 transitions, 31 text games) and tested GPT-4 (gpt-4-0125-preview) and GPT-3.5 (gpt-3.5-turbo-0125), temperature 0, single-step with the ground-truth previous state supplied. Best GPT-4 accuracy on transitions where something changes was 59.9% for the whole transition; action-driven changes reached at most 77.1% and environment-driven changes at most 49.7%. Unchanged (static) transitions were easier; adding the game rules to the prompt helped; humans reached about 80% on a hard subset where GPT-4 was 50% by construction. Curator qualifications: Abstract/intro: 'model accuracy does not exceed 59.9% for transitions in which a non-trivial change in the world state occurs'; '76,369 virtual text environment state transitions'; 'after 10 steps, average simulation accuracy would reduce to 0.599^10, or less than 1%'. Full counts: '77.1' 2, '49.7' 2, '24.4' 1, '22.3' 1, '92.1' 2, '61.5' 2, '34.5' 2, '12.5' 1, 'gpt-4-0125-preview' 4, 'gpt-3.5-turbo-0125' 1. Curator arithmetic: 0.599^10 = 0.0059. 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/2406.06485v1",
      "note": "Wang et al. arXiv 2406.06485v1 PDF"
    }
  ]
}