Research map / Multi-agent systems and games / LLM agents

CoLLMLight: Cooperative Large Language Model Agents for Network-Wide Traffic Signal Control

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes CoLLMLight, a cooperative LLM-agent framework for network-wide traffic signal control that uses spatiotemporal graph prompts, complexity-aware reasoning depth, and simulation-driven SFT plus environment-feedback refinement of a LoRA-tuned Llama 3.1 8B.

Problem

Existing LLM-based traffic signal control agents treat intersections independently, lacking cross-intersection coordination, and uniform long chain-of-thought reasoning is computationally costly for real-time control.

Approach

Combine a structured spatiotemporal observation prompt over neighboring intersections with a three-level complexity classifier (n_c = 0, 1, >1) that selects no-cooperation, simple, or full cooperative reasoning, and fine-tune the LLM on GPT-4o-structured reasoning chains from CityFlow simulations followed by environment-feedback refinement.

Key results

• CoLLMLight-8B achieves the lowest ATT on New York 1 (920.98 s) and New York 2 (1218.94 s) among compared methods, zero-shot from Syn-Train.
• On New York 1 it reduces AWT by 25.17% relative to the second-best method, and on New York 2 by 16.09%.
• Ablation shows removing policy refinement raises New York 1 ATT from 920.98 to 1085.88 and removing reasoning chain optimization (w/o both) raises it to 1153.74.

Surprising findings

• Performance of Deepseek R1-32B is comparable to CoLLMLight-8B on Jinan and Hangzhou, and removing reasoning-chain optimization on New York worsens results (w/o RO ATT 1167.37 vs w/o both 1153.74 for NY1 in some metrics).

Limitations

• Relies on GPT-4o to generate reasoning chains and pseudo labels, adding dependence on a proprietary model.
• Performance degrades with long historical windows (beyond 5 to 10 timesteps), indicating sensitivity to input length.
• Some large-LLM comparisons use partial settings (10% of intersections for R1-671B, 1200 s), limiting direct comparability.

Experimental setup

Datasets: Syn-Train New York 1 New York 2 Jinan 1 Jinan 2 Jinan 3 Hangzhou 1 Hangzhou 2

Benchmarks: New York 1 New York 2 Jinan 1 Jinan 2 Jinan 3 Hangzhou 1 Hangzhou 2

Metrics: average travel time average waiting time

Baselines: FixedTime MaxPressure MPLight AttendLight PressLight CoLight Efficient-CoLight Advanced-CoLight LLMLight

Builds on: LLMLight CoLight CityFlow chain-of-thought prompting

Closest papers by meaning

• VLMLight: Safety-Critical Traffic Signal Control via Vision-Language Meta-Control and Dual-Branch Reasoning Architecture (NeurIPS 2025 · similarity 0.88)
• π-Light: Programmatic Interpretable Reinforcement Learning for Resource-Limited Traffic Signal Control (AAAI 2024 · similarity 0.86)
• IntersectionZoo: Eco-driving for Benchmarking Multi-Agent Contextual Reinforcement Learning (ICLR 2025 · similarity 0.85)
• DiffLight: A Partial Rewards Conditioned Diffusion Model for Traffic Signal Control with Missing Data (NeurIPS 2024 · similarity 0.84)
• DriveCombo: Benchmarking Compositional Traffic Rule Reasoning in Autonomous Driving (CVPR 2026 · similarity 0.84)
• COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.83)