Research map / Multi-agent systems and games / LLM agents
TrajAgent: An LLM-Agent Framework for Trajectory Modeling via Large-and-Small Model Collaboration
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 TrajAgent, an LLM-agent framework that automates trajectory modeling by planning and optimizing specialized trajectory models through a unified environment (UniEnv) and a collaborative LLM-agent/small-model learning loop.
Problem
Trajectory modeling requires task- and data-specific specialized models that are hard to adapt across heterogeneous trajectory data and diverse tasks, and existing unified approaches have limited performance, task coverage, and usability.
Approach
An LLM agent workflow (task understanding, planning, execution, summary) runs over a unified data/model environment, and an optimization agent tunes data augmentation operators and model hyperparameters for small specialized models using memory of scored past trials (reflexion-style), with joint DA-then-PO optimization.
Key results
- Across 11 model/subtask settings, TrajAgent improves original models by 2.38% to 69.91% (e.g., map-matching GraphMM Acc_m 0.2014 to 0.3422, +69.91%).
- On next-location prediction (FSQ), GETNext improves from 0.3720 to 0.4002 Acc@5 (+7.58%) and LLM-ZS from 0.3110 to 0.3350 (+7.72%).
- Ablation: removing memory drops Acc from 98% to 80% in model selection and joint optimization Acc@5 from 0.3724 to 0.1804 (w/o memory).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply joint optimization to GETNext | GETNext · FSQ next location prediction | Acc@5: 0.3720 -> 0.4002 | Helped |
| Apply joint optimization to LLM-ZS | LLM-ZS · FSQ next location prediction | Acc@5: 0.3110 -> 0.3350 | Helped |
| Apply joint optimization to MulT-TTE | MulT-TTE · Porto travel time estimation | MAE: 163.12 -> 128.57 | Helped |
| Apply joint optimization to DutyTTE | DutyTTE · Tencent travel time estimation | MAE: 190.82 -> 179.01 | Helped |
| Apply joint optimization to TrajBERT | TrajBERT · Porto trajectory recovery | MAE: 42.71 -> 27.78 | Helped |
| Apply joint optimization to GraphMM | GraphMM · Tencent map matching | Acc_m: 0.2014 -> 0.3422 | Helped |
| Apply joint optimization to S2TUL | S2TUL · FSQ trajectory user linking | Acc@5: 0.5755 -> 0.7802 | Helped |
| Apply joint optimization to LIMP | LIMP · Beijing mobility intention prediction | Acc_i: 0.745 -> 0.7627 | Helped |
| Apply joint optimization to GM-VSAE | GM-VSAE · Porto trajectory anomaly detection | AUC: 0.9892 -> 0.9899 | Helped |
| Apply joint optimization to DSTPP | DSTPP · Earthquake trajectory generation | MAE: 0.4611 -> 0.3584 | Helped |
| Apply joint optimization to CACSR | CACSR · FSQ trajectory representation | Acc@5: 0.31 -> 0.3472 | Helped |
| Remove reflection mechanism (ablation) | TrajAgent with DeepMove · FSQ next location prediction | Acc@5: 0.3724 -> 0.3212 | Hurt |
Surprising findings
- Removing performance scores from memory caused next-location Acc@5 to fluctuate and often fall below baseline, suggesting score-guided reflection is essential.
- Stronger reasoning LLMs (e.g., DeepSeek-v3) are less affected by memory size growth, while weaker models degrade significantly.
Limitations
- Optimization trap: agent sometimes ignores memory feedback, converging to ineffective configurations, especially with long memory.
- Memory saturation and noise accumulation degrade performance beyond about 10 memory entries.
- Weaker LLMs produce many failed plans and execution steps, reducing reliability.
Experimental setup
Datasets: Foursquare (FSQ) Brightkite (BGK) Porto Chengdu Tencent Beijing
Benchmarks: FSQ BGK Porto Chengdu Tencent Beijing
Metrics: Acc@5 Hit@5 MAE AUC Acc_m Acc_i
Baselines: Optuna UrbanLLM
Builds on: Reflexion HuggingGPT Optuna DeepMove
Closest papers by meaning
- Large Language Models as Urban Residents: An LLM Agent Framework for Personal Mobility Generation (NeurIPS 2024 · similarity 0.85)
- TrajMamba: An Efficient and Semantic-rich Vehicle Trajectory Pre-training Model (NeurIPS 2025 · similarity 0.85)
- STraj: Self-training for Bridging the Cross-Geography Gap in Trajectory Prediction (AAAI 2025 · similarity 0.85)
- DeMo: Decoupling Motion Forecasting into Directional Intentions and Dynamic States (NeurIPS 2024 · similarity 0.85)
- UniTraj: Learning a Universal Trajectory Foundation Model from Billion-Scale Worldwide Traces (NeurIPS 2025 · similarity 0.84)
- HPNet: Dynamic Trajectory Forecasting with Historical Prediction Attention (CVPR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.