Grounded Iterative Language Planning: How Parameterized World Models Reduce Hallucination Propagation in LLM Agents
06:00 · June 29, 2026 · arXiv cs.AI RSS
World models for language agents come in two useful forms. An agent-based world model calls an LLM API and reasons flexibly in language, but its errors appear as hallucinated state changes that are hard to score with ordinary regression losses. A parameterized world model is a trained transition predictor; its errors are easier to measure with quantities such as NodeMSE, delta accuracy, and validity accuracy, but it is usually weaker as a standalone planner. We compare these two families on four graph-structured planning benchmarks and introduce operational hallucination metrics for the agent-based case. The comparison motivates \textbf{Grounded Iterative Language Planning} (GILP), which trains only a small parameterized backbone and combines it with API-based agent reasoning. The backbone supplies valid actions, predicted state deltas, risk, and value; the LLM drafts an action and imagined delta; and a consistency gate asks for revision when the two disagree. On real GPT-4o-mini calls, GILP reduces hallucinated-state rate from 0.176 to 0.035. In calibrated simulator ablations, it raises success from 0.668 to 0.838 while adding only ~22% extra LLM calls.
Summary
Language agents built on large models generate both actions and implicit predictions of how the world state will evolve after each step. These imagined transitions support flexible reasoning across long horizons, yet they introduce a specific failure mode: semantic hallucinations that are written into context and reused in later decisions. In graph-structured planning domains the resulting errors compound, as an incorrect claim about task completion or entity status can trigger invalid follow-on actions that the agent then attempts to reconcile with further fabricated state updates.
The paper contrasts this agent-style world model with a conventional parametric transition model. The latter supplies measurable supervised errors on action validity, state deltas, and value estimates, but lacks the semantic planning capacity of an LLM. To exploit the complementary strengths of each, the authors introduce Hybrid World-Model Planning. A small trained backbone predicts validity, next-state changes, risk, and value for candidate actions. At every step the language model still proposes an action and an imagined state delta in structured form; a lightweight consistency gate then compares the two predictions and issues a targeted correction only when they diverge beyond a fixed threshold.
Evaluations on four graph planning benchmarks demonstrate the practical effect. With GPT-4o-mini the hallucinated-state rate falls from 0.176 to 0.035, while calibrated simulator runs show task success rising from 0.668 to 0.838 at modest additional inference cost. The work also supplies operational metrics—hallucinated-state rate, propagation depth, and long-horizon error growth—that quantify how agent-generated state errors spread, together with ablation results that isolate the contribution of the parametric grounding component.
Why it matters
Directly addresses reliability of LLM agents, a core concern for Dutch AI research groups and ethical-AI initiatives; the hybrid technique is immediately actionable by teams fine-tuning small transition models and gating LLM calls.
