Latent Dynamics · 2026-11-11

Evaluating LLM-Generated World Models for Neuro-Symbolic Reasoning

Souhaila Djaffal, Amna Jamal, Gagana Madhugiri Devaraja, Anidipta Pal, Jobst Heitzig

arXiv:2611.14419PDF

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

The authors evaluate whether five LLMs can generate executable world models as partially observable stochastic games (POSGs) from natural-language scenarios. They introduce a 44-scenario benchmark and a three-lane evaluation (mechanical validity, reference-relative structural agreement, and semantic checks). In the reference-backed game track, 134/135 pass a strict mechanical audit, but mean permutation-invariant structural F1 is only 0.366, and just 28/134 exactly match the authored reference.

Why it matters

The paper argues that “executability is a necessary but insufficient criterion”: downstream neuro-symbolic planning and reasoning rely on the generated world model’s structure and semantics, not only on whether it runs. The authors show that mechanically valid models can still differ substantially from references and that adding cultural context (e.g., naming a country) can restructure the generated model while leaving automated semantic scores only slightly changed.

Method

  • Prompted five LLMs to translate natural-language scenarios into POSGs encoded as structured JSON or executable Python, requiring interface compliance and internal consistency checks.
  • Evaluated generations with three lanes: Lane 1 mechanical verification (well-formedness/executability); Lane 2a reference-based structural agreement (coverage/precision/F1 under alignment and bounded unrolling); Lane 3 semantic diagnostics via predefined programmatic checks (human protocol specified but not run).
  • Tested three empirical tracks (reference-backed games, open-ended situational scenarios, and culturally contextualized JSON POSGs) over 410 first-pass generations and analyzed structural and semantic differences.

Limitation

In this paper, Lane 3 results come only from predefined programmatic checks. The human-annotation protocol (Appendix C) is specified but not yet run, so all Lane 3 and bias results reported here are diagnostic and preliminary.

Abstract (from arXiv)

Formal world models can support planning, verification, and safety reasoning, but constructing them from open-ended natural-language situations is difficult, particularly in multi-agent settings where several formalizations may be plausible. We study whether five LLMs can generate such world models as partially observable stochastic games and whether mechanical validity is sufficient to establish that the resulting model faithfully represents the intended decision problem. We introduce a benchmark of 44 scenarios spanning game, situational, and cultural settings and evaluate generated world models along three dimensions: mechanical validity, reference-relative structural agreement, and semantic criteria. Across three empirical tracks, we analyze 410 first-pass generations. In the reference-backed game track, 134/135 generated world models pass a strict mechanical audit, yet mean permutation-invariant structural F1 is only 0.366, with just 28/134 mechanically valid models exactly matching the authored reference. In the cultural track, explicitly naming the country changes the generated structural representation for most LLMs while producing only small changes in automated semantic scores. These results show that executability is a necessary but insufficient criterion for LLM-generated world models: models that are mechanically valid can still differ substantially in the structure used by downstream neuro-symbolic reasoning and planning.

Related papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-11-10 · Indexed by Pub-AI: 2026-11-11 UTC

preprint

Evaluating LLM-Generated World Models for Neuro-Symbolic Reasoning

Souhaila Djaffal, Amna Jamal, Gagana Madhugiri Devaraja, Anidipta Pal, Jobst Heitzig

Preprint posted: 2026-11-10 · Original record ↗

arxiv
2611.14419
Explicit source versions (1)
  • v1 · 2026-11-10
  1. Preprint available

    Source date: 2026-11-10 · First detected: 2026-11-11 UTC · Historical baseline

    Source evidence (2)
Abstract · arxiv

Formal world models can support planning, verification, and safety reasoning, but constructing them from open-ended natural-language situations is difficult, particularly in multi-agent settings where several formalizations may be plausible. We study whether five LLMs can generate such world models as partially observable stochastic games and whether mechanical validity is sufficient to establish that the resulting model faithfully represents the intended decision problem. We introduce a benchmark of 44 scenarios spanning game, situational, and cultural settings and evaluate generated world models along three dimensions: mechanical validity, reference-relative structural agreement, and semantic criteria. Across three empirical tracks, we analyze 410 first-pass generations. In the reference-backed game track, 134/135 generated world models pass a strict mechanical audit, yet mean permutation-invariant structural F1 is only 0.366, with just 28/134 mechanically valid models exactly matching the authored reference. In the cultural track, explicitly naming the country changes the generated structural representation for most LLMs while producing only small changes in automated semantic scores. These results show that executability is a necessary but insufficient criterion for LLM-generated world models: models that are mechanically valid can still differ substantially in the structure used by downstream neuro-symbolic reasoning and planning.