Research map / Multi-agent systems and games / LLM agents
Long-term Traffic Simulation via Structured Autoregressive Modeling
Lingyu XIAO, Zexin Feng, Xintao Yan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use a frozen pretrained LLM backbone instead of the prior baseline (ablation) | RosettaSim · WOSAC val 2% | Composite score: 0.7631 -> 0.7638 | Improved |
| fine-tune only the last transformer layer versus fully frozen (ablation) | RosettaSim · WOSAC val 2% | Composite score: 0.7638 -> 0.7648 | Improved |
| fine-tune all layers instead of last layer only (ablation) | RosettaSim · WOSAC val 2% | Composite score: 0.7648 -> 0.7658 | No clear effect |
| tune last five layers instead of last one (ablation) | RosettaSim · WOSAC val 2% | Composite score: 0.7648 -> 0.7647 | No clear effect |
| use pretrained LLM weights instead of random initialization (ablation) | RosettaSim Qwen-0.5B · WOSAC val | Composite score: pretrained prior contributes the remaining 37%; architecture 63% | Improved |
| use modern LLM (Qwen2.5-0.5B) instead of GPT-2-Small (ablation) | RosettaSim · WOSAC val 2% (frozen) | Composite score: 0.5425 -> 0.7620 | Improved |
| use Qwen2.5-0.5B instead of Llama-3.2-1B (ablation) | RosettaSim · WOSAC val 2% (frozen) | Composite score: 0.7594 -> 0.7620 | Improved |
| use Qwen2.5-0.5B instead of TinyLlama-1.1B (ablation) | RosettaSim · WOSAC val 2% (frozen) | Composite score: 0.7397 -> 0.7620 | Improved |
| scale frozen Qwen2.5 from 1.5B to 3B (ablation) | RosettaSim · WOSAC val 2% (frozen) | Composite score: 0.7620 -> 0.7589 | Hurt |
| scale last-layer-tuned Qwen2.5 from 0.5B to 3B (ablation) | RosettaSim · WOSAC val 2% (last layer) | Composite score: 0.7637 -> 0.7641 | No clear effect |
| use Llama-3.2-1B instead of Qwen2.5-0.5B with last layer tuned (ablation) | RosettaSim · WOSAC val 2% (last layer) | Composite score: 0.7637 -> 0.7650 | No clear effect |
| replace Zipfian motion token distribution with uniform (coefficient 0) (ablation) | RosettaSim · WOSAC val | Composite score | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SMART — SMART: Simultaneous Multi-Agent Recurrent Trajectory Prediction (ECCV 2020) · better on 3 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Trajeglish: Traffic Modeling as Next-Token Prediction (ICLR 2024 · similarity 0.90)
- BehaviorGPT: Smart Agent Simulation for Autonomous Driving with Next-Patch Prediction (NeurIPS 2024 · similarity 0.88)
- Advancing Multi-agent Traffic Simulation via R1-Style Reinforcement Fine-Tuning (ICLR 2026 · similarity 0.87)
- AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications (ICML 2026 · similarity 0.87)
- DriveGPT: Scaling Autoregressive Behavior Models for Driving (ICML 2025 · similarity 0.86)
- DriveCombo: Benchmarking Compositional Traffic Rule Reasoning in Autonomous Driving (CVPR 2026 · similarity 0.86)
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.
