What exactly did arXiv 2610.06637 (v1) define and measure for long-horizon textual world modelling: environments, horizons, models, metrics, headline numbers, ablations, limits, release?

arXiv 2610.06637 (Wang, Gao, Yao, Dong, Walia, Das; UIC and Intuit AI Research; v1 submitted 2026-10-05 16:25:44 UTC; CC BY 4.0) fine-tunes Qwen3-4B-Instruct-2507 world models that keep a JSON "textual state" and predict a whole k-action sequence in one call as sparse set/remove edits that code replays (latent-SR). It evaluates horizons 1-10 (ScienceWorld, CEO-Bench) and 1-20 (Jericho) with a fixed-target protocol against raw-history recursive/direct baselines and GPT-5.2. latent-SR has the best averages on every metric (ScienceWorld Fact-F1 92.44 vs 88.94 for obs-direct; CEO-Bench SE-score 87.35 vs 81.85; Jericho BLEU 24.5 and Score MAE 1.728), with gains growing at long horizons, but obs-recursive is better on Jericho at h1 and h5. The abstract's "nearly 30%" Jericho error drop holds against obs-direct at h10 (28.9%) but not against obs-recursive (12.2%), which is the lower-error raw-history baseline there. No code, data or weights release is stated. Curator qualifications: Curator recompute: Jericho h10 error drop is 28.9% vs obs-direct but 12.2% vs obs-recursive, the stronger raw-history baseline at h10; the abstract's 'nearly 30%' therefore is not against the strongest baseline, unlike its ScienceWorld/CEO figures. At h1 obs-recursive has far lower error than latent-SR. Scope correction: report the Jericho gain as 'about 12% vs the strongest raw-history baseline (obs-recursive), about 29% vs obs-direct, at horizon 10'; latent-SR is worse at short horizons on Jericho. 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/2610.06637v1",
      "note": "arXiv 2610.06637v1 HTML abstract and Table 3 (Jericho fixed-target, MAE lower is better)"
    }
  ]
}