Technical Summary: LISA – Cognitive Arbitration for Signal-Free Autonomous Intersection Management

1. Problem Statement

Urban intersection management has historically relied on infrastructure-centric paradigms, utilizing traffic signals and signs to regulate flow based on aggregate measurements rather than individual vehicle capabilities or intentions. While Connected and Autonomous Vehicles (CAVs) offer the potential for direct coordination, existing solutions face distinct structural limitations:

• Signal-based controllers (fixed-cycle, adaptive) remain vehicle-agnostic, optimizing phases based on aggregate flow rather than specific vehicle intents or priorities.
• Green Light Optimal Speed Advisory (GLOSA) improves kinematic efficiency but remains tethered to existing signal schedules, lacking mechanisms for multi-vehicle right-of-way resolution.
• Autonomous Intersection Management (AIM) eliminates signals via reservation slots but relies on deterministic scheduling of kinematic states, failing to reason about vehicle intent, priority classes, or contextual trade-offs like queue pressure.
• LLM-based approaches have emerged as reasoning tools but are typically used as auxiliary modules to select signal phases or correct Reinforcement Learning (RL) decisions, rather than acting as autonomous agents for direct, signal-free decision-making.

Furthermore, the high inference latency of Large Language Models (LLMs) poses a significant barrier to real-time control, as sub-second vehicle control cycles cannot wait for multi-second reasoning calls. The core challenge is whether LLM reasoning alone, without signal infrastructure or task-specific training, can effectively arbitrate crossing intents and deliver cooperative kinematic advisories in real time.

2. Methodology: The LISA Framework

The authors propose LISA (LLM-Based Intent-Driven Speed Advisory), a signal-free cognitive arbitration framework. LISA reframes intersection management as a cognitive arbitration problem where an LLM agent resolves competing right-of-way claims directly from vehicle intents, without signal phases or reservation schedules.

System Architecture

LISA employs a two-layer control architecture to separate high-level semantic reasoning from low-level kinematic execution:

1. High-Level Semantic Arbitration (LLM Agent): The LLM receives structured inputs regarding active approaches, declared vehicle intents (spatial, temporal, priority, energy), queue pressures, and conflict relations. It outputs symbolic right-of-way roles: CLEAR (proceed), YIELD (defer), or SHARE (concurrent reduced speed).
2. Low-Level Kinematic Execution: A deterministic executor converts these symbolic roles into specific, verifiable speed advisories ($v_{adv}$) for each vehicle based on its current distance to the stop line. This layer ensures temporal precision and safety, removing the LLM from the real-time control loop.

Key Mechanisms for Latency Mitigation

To address LLM inference latency, LISA integrates two complementary mechanisms:

• Memoized Arbitration Table (MAT): A cache stores previously computed right-of-way assignments keyed by a discrete "conflict signature" (encoding priority classes, urgency bands, queue depth, and energy preferences). This allows the system to serve recurring conflict states without invoking the LLM, achieving high cache hit rates.
• Anticipatory Arbitration: Queries are triggered when vehicles enter an "advisory horizon" (400m from the stop line), creating a latency buffer of approximately 14 seconds before vehicles enter the "near zone" (200m) where speed advisories are enforced. This ensures LLM responses arrive before they are needed.

Safety and Logic

• Safety Watchdog: A tile-based collision monitor runs independently of the LLM. If two vehicles with conflicting movements occupy the same tile, the later-arriving vehicle is overridden to a hold speed ($V_{hold} = 2.0$ m/s) to prevent collisions.
• Role Logic: The system enforces strict pre-emption for emergency vehicles. For saturated conflicting approaches with comparable queue pressure, the SHARE role allows both to proceed at reduced, pressure-weighted speeds to prevent starvation.

3. Key Contributions

1. Signal-Free Intent-Based Arbitration: LISA is the first framework to formulate intersection management as a cognitive arbitration problem where an LLM resolves right-of-way directly from approach-level intents without signal phases, reservation schedules, or fine-tuning.
2. Latency-Mitigated Two-Layer Architecture: The separation of semantic arbitration and deterministic execution, combined with the MAT and anticipatory querying, enables real-time performance. The MAT achieved hit rates up to 98.8% at medium demand, and anticipatory arbitration sustained zero LLM fallbacks across all conditions.
3. Comprehensive Empirical Evaluation: LISA was evaluated in the SUMO simulator against Fixed-cycle, SCATS (adaptive), AIM, and GLOSA across low, medium, and near-saturated traffic loads.
4. Multi-Dimensional Intent Satisfaction Metric: The authors introduce a structured evaluation axis decomposing performance into spatial, temporal, priority, and energy satisfaction components, providing a holistic view of vehicle-level intent fulfillment.

4. Experimental Results

Evaluated using the Gemini-2.5 Flash Lite model, LISA demonstrated significant improvements over baselines:

• Traffic Efficiency:

  • Mean Control Delay: LISA reduced mean control delay by up to 89.1% compared to fixed-cycle control at medium load.
  • Level of Service (LOS): At medium load ($v/c = 0.63$), LISA maintained LOS C, whereas all non-LLM baselines degraded to LOS F.
  • Waiting Time: Under near-saturated demand, LISA reduced mean waiting time by 93% relative to fixed-cycle control.
  • Peak Queue Length: LISA reduced peak queue length by 60.6% relative to fixed-cycle control at high load.

• Energy and Emissions:

  • LISA lowered fuel consumption by up to 48.8% relative to fixed-cycle control.
  • Kinetic energy loss was reduced by up to 26% at medium demand.

• Intent Satisfaction:

  • LISA achieved an overall intent satisfaction of 86.2% at medium demand, significantly outperforming the best non-LLM result (61.2% for GLOSA).
  • This gain was driven by superior performance in temporal delay compliance, priority service (especially for emergency vehicles), and energy-aware smoothness.

• LLM Performance:

  • The benchmarking phase identified Gemini-2.5 Flash Lite as the optimal backend, achieving a composite score of 93.3% with a mean latency of 1,592 ms and zero calls exceeding the 5-second threshold.

5. Significance and Claims

The paper claims that LISA demonstrates that LLM reasoning, when combined with lightweight latency mitigation and deterministic execution, can support real-time cooperative intersection arbitration without signal infrastructure.

The authors argue that LISA's primary significance lies in its ability to:

• Move beyond aggregate flow regulation to intent-aware coordination, resolving conflicts based on social priorities, route choices, and energy preferences.
• Operate in a zero-shot regime, relying on the LLM's general reasoning capabilities rather than task-specific training or fine-tuning.
• Provide a human-readable, interpretable conflict resolution process where the LLM acts as an explicit arbitration agent rather than a black-box optimizer.

The authors note that while LISA shows the strongest gains in the "transition regime" (medium load) where arbitration quality prevents the onset of persistent congestion, it remains effective even at high loads for managing congestion distribution. The paper concludes that this approach offers a viable path toward signal-free, cooperative autonomous driving, though future work is required to address multi-intersection coordination and robustness to prompt variations.