Research map / Multi-agent systems and games / Human-AI and goal-conditioned
InfGen: Scenario Generation as Next Token Group Prediction
Zhenghao Peng, Yuxin Liu, Bolei Zhou
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 |
|---|---|---|---|
| add autoregressive structured agent-state decoding (vs parallel independent heads) (ablation) | SceneStreamer · WOMD initial state MMD strict, Size | MMD (lower better): 0.1172 -> 0.0743 | Improved |
| add autoregressive agent-state decoding (ablation) | SceneStreamer · WOMD initial state MMD strict, Position | MMD (lower better): 0.1603 -> 0.1291 | Improved |
| add autoregressive agent-state decoding (ablation) | SceneStreamer · WOMD initial state MMD strict, Heading | MMD (lower better): 0.1646 -> 0.1270 | Improved |
| add autoregressive agent-state decoding (ablation) | SceneStreamer · WOMD initial state MMD strict, Velocity | MMD (lower better): 0.2114 -> 0.1970 | Improved |
| add autoregressive agent-state decoding (ablation) | SceneStreamer · WOMD initial state MMD relaxed (all agents), Position | MMD (lower better): 0.3237 -> 0.2198 | Improved |
| add autoregressive agent-state decoding (ablation) | SceneStreamer · WOMD initial state MMD relaxed (all agents), Velocity | MMD (lower better): 0.1183 -> 0.0730 | Improved |
| train planner on SceneStreamer-Motion traffic instead of log-replay | TD3 RL planner · WOMD val (100 scenarios) | Average episodic reward: 32.24 -> 37.98 | Improved |
| add adaptive SDC state-forcing (closed-loop training) (ablation) | SceneStreamer-Motion planner training · WOMD val (100 scenarios) | Average episodic reward: 37.98 -> 39.23 | Improved |
| add adaptive SDC state-forcing (ablation) | SceneStreamer-Full (No RS) · WOMD val (100 scenarios) | Average episodic reward: 38.18 -> 38.81 | No clear effect |
| add reject sampling for colliding agents (ablation) | SceneStreamer-Full (w/ Ada) · WOMD val (100 scenarios) | Success rate: 0.7385 -> 0.7620 | Improved |
| add reject sampling for colliding agents (ablation) | SceneStreamer-Full (w/ Ada) · WOMD val (100 scenarios) | Route completion: 0.7230 -> 0.7345 | Improved |
| fine-tune to also generate agent states (SceneStreamer-Full vs Motion) | SceneStreamer · WOMD motion prediction (all agents) | ADEmin: 0.8730 -> 0.9339 | Hurt |
Closest papers by meaning
- Scenario Dreamer: Vectorized Latent Diffusion for Generating Driving Simulation Environments (CVPR 2025 · similarity 0.90)
- Generating Traffic Scenarios via In-Context Learning to Learn Better Motion Planner (AAAI 2025 · similarity 0.89)
- BehaviorGPT: Smart Agent Simulation for Autonomous Driving with Next-Patch Prediction (NeurIPS 2024 · similarity 0.88)
- Trajeglish: Traffic Modeling as Next-Token Prediction (ICLR 2024 · similarity 0.88)
- SceneDiffuser++: City-Scale Traffic Simulation via a Generative World Model (CVPR 2025 · similarity 0.87)
- Advancing Multi-agent Traffic Simulation via R1-Style Reinforcement Fine-Tuning (ICLR 2026 · similarity 0.87)
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.
