The latest version of an autonomous-driving system that uses large language models during development posted the strongest combined curve-steering performance reported in a CARLA simulation, according to an arXiv preprint. Across the final 200 episodes, V3 had a median, or middle, lane error of 0.0627—approximately 15–20% lower than V0, V1 and V2, whose medians were 0.0780, 0.0733 and 0.0770—and the highest median curve exposure, at 160 steps per episode.
That lead did not extend to every measure. V3’s median steering-change magnitude was 0.3119, and its exposure-weighted curve-oscillation rate was 45.27%. V0’s rate was lower, at 43.18%, while V1 and V2 recorded 49.18% and 47.61%. The comparison is descriptive; no inferential tests or confidence intervals are reported.
At episode 1,000, V3 again had the lowest median lane error, 0.0514, the lowest oscillation rate among the non-degenerate variants, 45.89%, and the highest median curve exposure, 193.5 steps. By comparison, V2 had a lane error of 0.0849, a steering-change magnitude of 0.3708, an oscillation rate of 47.99% and exposure of 187.5 steps; V1 had 101.5 curve steps and a terminal oscillation rate of 49.69%.
One result needs a careful reading. V0’s late reductions in steering changes and oscillations coincided with curve exposure collapsing to 21.5 steps. The analysis treats that as reduced driving opportunity, not genuine stabilization—a reminder that a smoother-looking trace can partly reflect how much of a curve the system actually traversed.
The language model stayed in the development loop
GPT-5.2 was used offline during development to analyze logged episodes, identify failure patterns and recommend changes to the reinforcement-learning reward—the score used to shape trial-and-error learning. Human validation came before retraining. The initial V0 model was trained without LLM guidance on data from over 1,000 CARLA episodes.
At runtime, camera and LiDAR data fed four specialized ASIL-rated agents into a priority arbitrator. Control used PID or PPO policies, and safety constraints could override learned behavior. Two independent PPO-based actor-critic networks were invoked independently at each control tick and remained subject to Safety Agent override.
The framework also generated common-sense driving rules. Unsupported or unobservable simulator actions were excluded, and approximately 100 executable rules were integrated into CARLA.
When stop recommendations conflicted, an ASIL-D or ASIL-C agent recommending stop above the confidence threshold could veto other opinions and force an immediate stop.
A simulator result with hard limits
All of the reported performance evidence comes from CARLA simulation. The scenarios were highly randomized and varied environmental and traffic conditions, but the study described no physical-vehicle or real-world road validation.
The architecture’s veto mechanism is a design feature, not a measured safety outcome. No numerical collision, off-road or other safety-event results are reported, so the study does not establish safer driving or an absolute reduction in risk.
The variants were not randomized, and independent runs, random seeds, between-run variability estimates, confidence intervals and p-values were not reported. The results therefore show how the versions compared in this simulation, not that LLM reward refinement caused the differences.
The manuscript is an arXiv version 1 preprint dated 20 August 2026. Whether the pattern survives independent seeds and fully specified scenario distributions, transfers to physical vehicles and real traffic, or holds up under external audit of its generated rules and proposed ISO/ASIL implementation remains open.
Paper data and sources
Original title: Multi-Agent Orchestration with the Common-Sense Reasoning Capabilities of LLMs for Autonomous Driving
Authors: Mehdi Azarafza, Faezeh Pasandideh, Ali Ehteshami Bejnordi et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text