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

A new planning method called Grounded Iterative Language Planning (GILP) sharply reduces hallucination in large language model agents by pairing a small trained world model with API-based reasoning, according to research published on arXiv [1]. The approach addresses a core weakness in agent-based world models, where an LLM reasons flexibly but can produce hallucinated state changes that are difficult to measure with standard regression losses [2]. Parameterized world models, by contrast, are trained transition predictors whose errors are easier to quantify using metrics such as NodeMSE, delta accuracy, and validity accuracy, though they are typically weaker as standalone planners [2]. GILP combines the two: it trains only a small parameterized backbone and integrates it with API-based agent reasoning [2]. The backbone supplies valid actions, predicted state deltas, risk, and value, while the LLM drafts an action and imagined delta. A consistency gate then requests revision when the two disagree [2]. On real GPT-4o-mini calls, GILP reduced the hallucinated-state rate from 0.176 to 0.035 [2]. In calibrated simulator ablations, the method raised the success rate from 0.668 to 0.838 while adding approximately 22% extra LLM calls [2]. The researchers evaluated the two families of world models on four graph-structured planning benchmarks and introduced operational hallucination metrics for the agent-based case [2]. The work builds on a broader effort to make LLM-based agents more reliable in sequential decision-making tasks. Prior research has explored parameterized transition predictors for measuring planning errors, but those models have historically lacked the flexible reasoning capabilities of LLM APIs [2]. By gating the LLM’s output with a consistency check against the parameterized backbone, GILP retains the flexibility of language-based reasoning while curbing the propagation of hallucinated states [2]. The paper does not include quotes from the authors, and the research has not yet been peer-reviewed [1].