REINS lets an LLM trajectory planner meet real-time limits by storing verified rules

Researchers led by Zhihong Cui have presented REINS, an agent harness that lets a large language model plan trajectories within real-time limits and improve over time, in a paper accepted at NeurIPS 2026.

Existing methods either keep the model in the decision loop, which is too slow for millisecond control, or move it offline, which leaves the planner unable to change at run time, the authors write. REINS ties the model's outputs to a learned dynamics model, checks them with calibrated predicates and forward rollouts, and stores verified results in an indexed rule memory. Familiar scenes are then handled from memory, and only new ones call the model.

On simulated and real-world benchmarks at the tactical decision layer, REINS reached more than 98% compliance, cut collisions by nearly half and had a latency of 1.9 to 4.6 ms, with the model called less often as the memory grew.