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

Short summary

This arXiv paper introduces GILP (Grounded Iterative Language Planning), combining trained world models with LLM-based reasoning to reduce hallucination in language agents. On graph-planning benchmarks, GILP cuts the hallucinated-state rate from 17.6% to 3.5% while raising task success to 84% with only 22% more API calls. A consistency gate flags disagreements between the LLM's proposed actions and predictions, triggering revision when needed.

• •Introduces GILP method combining parameterized world models with LLM reasoning
• •Reduces hallucination rate 5x (17.6% → 3.5%) on planning benchmarks
• •Achieves 84% task success rate with minimal overhead (22% more LLM calls)

Generated with AI, which can make mistakes.