The Problem: Hallucination Propagation in Agentic Planning
LLM-based agents often struggle with state tracking because they rely on flexible, natural language reasoning that lacks a formal grounding mechanism. When an agent hallucinates a state change, that error propagates through subsequent planning steps, leading to compounding failures. While parameterized world models (trained transition predictors) offer better measurability through metrics like NodeMSE and delta accuracy, they typically lack the planning flexibility of frontier LLMs.
The GILP Approach: Hybrid Grounding
Grounded Iterative Language Planning (GILP) bridges this gap by pairing a high-level LLM planner with a lightweight, trained parameterized backbone. The system operates as follows:
- Drafting: The LLM generates a proposed action and an imagined state delta.
- Verification: The parameterized backbone provides an independent prediction of the state delta, risk, and value based on the current environment state.
- Consistency Gate: A gating mechanism compares the LLM's output with the backbone's prediction. If the two disagree, the agent is forced to revise its plan.
Performance and Reliability
GILP demonstrates that adding a small, specialized backbone significantly improves agent reliability without requiring massive compute overhead. In experiments using GPT-4o-mini, GILP reduced the hallucinated-state rate from 0.176 to 0.035. Furthermore, in calibrated simulator tests, the architecture increased task success rates from 0.668 to 0.838. This performance gain comes at a cost of approximately 22% additional LLM calls, representing a favorable trade-off for production systems where reliability and hallucination reduction are critical.