Grounded Iterative Language Planning: How Parameterized World Models Reduce Hallucination Propagation in LLM Agents

2026-06-26

AI Summary

This work addresses the challenge of hallucination-induced state prediction errors in large language model (LLM) agents, which can propagate through planning and degrade task success rates. The authors propose the Grounded Iterative Language Planning (GILP) framework, which uniquely integrates a lightweight parametric world model with LLM-based reasoning. GILP employs a consistency-gating mechanism to dynamically verify the alignment between actions and state transitions, thereby effectively curbing hallucination propagation. The study introduces quantifiable metrics for operational hallucination—namely NodeMSE and validity accuracy—and demonstrates significant improvements: in real-world evaluations using GPT-4o-mini, the hallucinated state rate drops from 0.176 to 0.035, while simulated task success increases from 0.668 to 0.838, with only a modest ~22% increase in LLM invocation overhead.

Abstract

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.

Problem

• hallucination propagation
• language agents
• world models
• planning
• state hallucination

Innovation

• Grounded Iterative Language Planning
• parameterized world model
• hallucination reduction
• consistency gating
• LLM agents