💻 computer science

LLM-Augmented Traffic Signal Control with LSTM-Based Traffic State Prediction and Safety-Constrained Decision Support

This paper proposes an LLM-augmented traffic signal control framework that integrates LSTM-based traffic state prediction with safety-constrained decision support to enhance adaptability and interpretability in dynamic traffic scenarios while ensuring zero operational violations.

Original authors: Jiazhao Shi

Published 2026-04-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiazhao Shi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a busy city intersection as a chaotic kitchen during the dinner rush. The traffic lights are the head chef, trying to decide which group of cooks (cars) gets to use the stove (the road) next.

For a long time, chefs have used two main strategies:

The Fixed Timer: The chef sets a timer to let the North-South cooks work for 30 seconds, then the East-West cooks for 30 seconds, no matter how many pots are actually boiling. This is simple, but if the North-South line is huge and the East-West line is empty, the North-South cooks get frustrated while the stove sits idle.
The Reactive Chef: This chef looks at the pots right now. If the North-South line is long, they keep the light green. But they are "blind" to what's coming down the hallway in the next minute.

This paper proposes a new, smarter kitchen manager: an LLM-Augmented Traffic Signal Control system. Here is how it works, broken down into simple parts:

1. The Crystal Ball (LSTM Prediction)

First, the system uses a tool called LSTM (a type of AI good at spotting patterns in time). Think of this as a Crystal Ball.
Instead of just looking at the cars currently at the stop line, the Crystal Ball looks at the last few minutes of traffic and predicts what will happen in the next minute. It guesses: "Oh, a huge wave of cars is coming from the North in 30 seconds, but the East side will clear up soon."

2. The Junior Sous-Chef (Predictive Controller)

Based on that Crystal Ball prediction, a standard computer program (the Predictive Controller) acts like a Junior Sous-Chef. It calculates the best move: "Since the North is getting crowded, let's keep the North-South light green a bit longer." It makes a quick, math-based decision.

3. The Wise Head Chef (The LLM)

Here is where the new idea comes in. The system doesn't just let the Sous-Chef make the final call. It passes the plan to a Large Language Model (LLM), which acts like a Wise Head Chef.

The Head Chef doesn't just look at numbers; it reads a "story" about the traffic. It gets a report that says:

"Current queue: 16 cars."
"Prediction: 24 cars in 60 seconds."
"Sous-Chef's plan: Keep the light green."

The Head Chef then uses its reasoning skills to say:

Diagnosis: "Yes, the North side is getting a massive surge. The Sous-Chef is right."
Recommendation: "Let's extend the green light, but not too long, or we'll block the pedestrians."
Explanation: "I'm extending the light because the Crystal Ball predicts a traffic jam forming in the next minute, and we need to clear the backlog before it spills over."

4. The Safety Inspector (Safety Filter)

This is the most critical part. Large Language Models can sometimes be chatty or make mistakes. If the Head Chef suggests something dangerous (like "Keep the light green for 10 minutes straight!"), the system has a Safety Inspector.

The Inspector checks the rules: "Wait, the maximum green time is 45 seconds. You can't do that."
If the Head Chef's idea breaks the rules, the Inspector fixes it or rejects it and goes back to the Sous-Chef's original safe plan. This ensures that no unsafe decisions ever happen, even if the AI gets confused.

What Did They Find?

The researchers tested this in a computer simulation (a virtual city) with three types of traffic:

Normal Traffic: Everyone is driving at a steady pace.
Rush Hour: One direction is packed; the other is empty.
Sudden Surge: A big event happens, and a massive crowd of cars appears out of nowhere.

The Results:

Better than the old ways: The new system was much better than the "Fixed Timer" and the "Reactive Chef."
The Head Chef shines in chaos: When traffic was normal, the Crystal Ball (LSTM) did most of the work. But when traffic got weird, sudden, or unbalanced (like the Rush Hour or Sudden Surge), the Wise Head Chef (LLM) made a big difference. It figured out complex situations that the math-only computer missed, reducing wait times and traffic jams significantly.
Zero Safety Violations: Thanks to the Safety Inspector, the system never broke the traffic rules (like running a red light or ignoring pedestrians).

The Big Takeaway

The paper argues that we shouldn't let a "Wise Head Chef" (LLM) run the kitchen alone because they might get distracted or make unsafe guesses. Instead, we should use them as a consultant.

Let the math (LSTM) do the heavy lifting of predicting the future, let the Head Chef (LLM) explain why a decision makes sense and catch tricky situations, and let the Safety Inspector make sure everyone follows the rules. This combination creates a traffic system that is not only efficient but also easy for humans to understand and trust.

1. Problem Statement

Traffic signal control is a critical component of Intelligent Transportation Systems (ITS), yet conventional methods face significant limitations:

Fixed-time and Rule-based methods: These struggle to adapt to dynamic, non-recurrent traffic demands (e.g., sudden surges, directional imbalances) and lack the ability to anticipate future congestion.
Reinforcement Learning (RL): While data-driven RL methods (e.g., IntelliLight, CoLight) show promise, they often suffer from poor generalization across unseen networks, lack interpretability (acting as "black boxes"), and are purely reactive rather than predictive.
Direct LLM Control: While Large Language Models (LLMs) offer reasoning and interpretability, using them as direct, low-level real-time controllers is unsafe due to potential latency, output inconsistency, and the inability to strictly guarantee safety constraints (e.g., minimum green time, pedestrian crossing).

Core Challenge: How to leverage the reasoning and interpretability of LLMs to enhance adaptive traffic control without compromising operational safety or real-time performance.

2. Methodology

The authors propose a hybrid framework that positions the LLM as a high-level decision-support module rather than a direct controller. The system integrates four main components:

A. Traffic State Representation & Prediction (LSTM)

Input: The system collects real-time intersection data including queue length ( $q$ ), accumulated waiting time ( $w$ ), vehicle count ( $n$ ), average speed ( $v$ ), lane occupancy ( $o$ ), and current signal phase ( $p$ ).
Model: A Long Short-Term Memory (LSTM) network is trained to forecast short-term future traffic states (horizon $H=60$ seconds) based on a historical window ( $K=120$ seconds).
Output: Predictions for future queue lengths, waiting times, and demand scores.

B. Predictive Signal Controller (Baseline)

A predictive controller uses the LSTM forecasts to calculate a "demand score" for each feasible signal phase.
It selects a candidate action (e.g., extend current green or switch phases) that minimizes predicted future congestion. This serves as the baseline for comparison.

C. LLM-Augmented Decision Support

Role: The LLM acts as a reasoning engine, not a direct actuator.
Input: It receives a structured JSON object containing:
- Current and predicted traffic states.
- The candidate action generated by the LSTM controller.
- Operational constraints (min/max green time, pedestrian requests, emergency vehicles).
Function: The LLM evaluates the candidate action, diagnoses congestion patterns (e.g., directional imbalance, spillback), and generates:
- A recommendation (accept, reject, or adjust duration).
- A natural-language explanation for human operators.
- A risk flag.
Trigger Mechanism: To reduce computational cost, the LLM is not called at every step. It is triggered only when traffic conditions are ambiguous or abnormal (e.g., significant queue growth or small differences in demand scores between phases).

D. Safety-Constrained Action Filtering

Critical Component: Before any action is executed, a deterministic safety filter validates the LLM's recommendation against strict traffic signal constraints ( $C$ ).
Logic:
- If the LLM recommendation violates constraints (e.g., exceeding max green time), the system either modifies it to the nearest feasible value or reverts to the original LSTM candidate action.
- This ensures zero constraint violations in the final execution.

3. Key Contributions

Novel Framework: Proposes an LLM-augmented architecture that combines LSTM temporal prediction with structured LLM reasoning and a safety filter, avoiding the pitfalls of using LLMs as direct low-level controllers.
Interpretability: Unlike black-box RL, the system provides natural-language explanations and congestion diagnoses, making automated decisions transparent to human traffic managers.
Safety Assurance: Introduces a rigorous safety filter that guarantees all executed actions adhere to operational constraints, addressing the safety-critical nature of traffic control.
Empirical Validation: Demonstrates that LLMs provide the most significant value in dynamic and non-recurrent scenarios (directional peaks, sudden surges) where purely numerical rules or standard predictive models struggle.

4. Experimental Results

The framework was evaluated in a SUMO microscopic simulation under three scenarios: Balanced Demand, Directional Peak, and Sudden Surge. It was compared against Fixed-Time, Rule-Based, and LSTM-Predictive baselines.

Key Findings:

Performance Gains: The LLM-augmented method outperformed all baselines.
- Under Sudden Surge conditions, it reduced average waiting time by 14.4% and queue length by 18.4% compared to the LSTM-only baseline.
- Improvements were modest in balanced traffic but became substantial as traffic conditions became more irregular.
Ablation Study:
- Removing the LLM module resulted in a clear performance drop, confirming its value in complex reasoning.
- Removing the safety filter led to a slight performance increase but introduced a 4.8% constraint violation rate, proving the filter is essential for safety.
Safety & Reliability: The final system achieved 0% constraint violations across all scenarios.
Explainability: The LLM explanation validity rate remained above 89%, indicating the generated reasoning was consistent with the traffic state and final actions.

5. Significance

This study provides a practical pathway for integrating Generative AI into safety-critical infrastructure. It demonstrates that:

LLMs are best suited as "co-pilots": They excel at high-level reasoning, diagnosis, and explanation but should not replace deterministic safety layers.
Hybrid approaches are superior: Combining the temporal forecasting of deep learning (LSTM) with the semantic reasoning of LLMs yields better adaptability than either method alone.
Human-in-the-loop readiness: By providing interpretable explanations and ensuring safety constraints, this framework bridges the gap between automated control and human trust, facilitating the deployment of next-generation intelligent traffic systems.