GILP hybrid model reduces LLM agent hallucinations by 80%

Current world models for LLM agents fall into two camps: agent-based models that call an LLM API for flexible reasoning but suffer from hallucinated state changes that are hard to score, and parameterized world models that are easier to measure with regression losses but weaker as standalone planners. This new research from arXiv authors Xinyuan Song and Zekun Cai introduces a third path—Grounded Iterative Language Planning (GILP). GILP trains only a small parameterized backbone to supply valid actions, state deltas, risk, and value, then lets the LLM draft actions and imagined deltas. A consistency gate compares the two and asks for revision when they disagree, effectively grounding the LLM's language-based reasoning in a structured model.

Tested on graph-structured planning benchmarks, GILP delivers remarkable results. On real calls to GPT-4o-mini, the hallucinated-state rate dropped from 0.176 to 0.035—an 80% reduction. In calibrated simulator ablations, success rates improved from 0.668 to 0.838 while requiring only about 22% more LLM invocations. This hybrid approach offers a practical way to make LLM agents far more reliable without massive new model training or prohibitive compute costs, opening the door to more trustworthy autonomous planning and decision-making systems.