What if an LLM designed the perfect neural architecture for traffic AI?

Imagine an AI managing traffic in a mixed city where autonomous and human-driven cars coexist. To decide, it must "read" everything at once: sensor data, camera video feeds, historical vehicle time-series, and even text instructions. Fusing these diverse sources into a single useful representation (state encoding) is a puzzle that previously took months of manual trial-and-error.

Now, researchers have automated the search for the ideal neural architecture (NAS) using a Large Language Model (LLM) itself. Instead of randomly trying thousands of combinations, the LLM leverages its prior knowledge of code and model architectures to suggest the most promising structures. The extra trick? The system "listens" to the quality of intermediate representations produced by each module (sensors, images, text) during the search, guiding choices toward more effective fusions.

Tested on a mixed-autonomy traffic control simulator, this approach discovered better architectures than classic NAS baselines and even the GENIUS framework, evaluating far fewer candidates. This means smarter AIs, trained faster with fewer resources. A step forward for robotics, autonomous driving, and any system that must make sense of a multimodal world.

The paper formalizes multi-source state encoder design as a composite Neural Architecture Search (NAS) problem. The search space includes source-specific modules (e.g., CNNs for images, RNNs/Transformers for time-series, text encoders) and a fusion module. The main novelty is using an LLM as the search controller (LLM-driven NAS).

Unlike traditional RL controllers (e.g., REINFORCE) or gradient-based methods (DARTS), the LLM generates PyTorch code for candidate architectures leveraging the "prior" learned during pre-training on vast code and scientific literature corpora. This drastically reduces the effective search space.

The key innovation for sample efficiency is the use of "intermediate output signals." During partial evaluation of a candidate architecture, the system extracts metrics on the latent representation quality of each branch (e.g., intrinsic dimensionality, mutual information with reward, orthogonality). These signals are fed into the LLM prompt for the next iteration, enabling "reasoning" on why a specific fusion works or fails.

Experiments on Mixed-Autonomy Traffic Control (Flow/SUMO benchmark) show the method outperforms Random Search, REINFORCE, DARTS, and GENIUS (SOTA LLM-NAS) in final reward and samples required. Ablation studies confirm both the LLM prior and intermediate signals contribute significantly. This paves the way for automated feature engineering for complex multimodal RL agents.