Research map / Multi-agent systems and games / Human-AI and goal-conditioned
MotionLM: Multi-Agent Motion Forecasting as Language Modeling
Ari Seff, Brian Cera, Dian Chen, Mason Ng, Aurick Zhou, Nigamaa Nayakanti, Khaled S. Refaat, Rami Al-Rfou, Benjamin Sapp
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 interactive attention across agents at 2 Hz instead of marginal rollouts (ablation) | MotionLM · WOMD interactive val | prediction overlap: 0.0404 -> 0.0292 | Improved |
| increase interactive attention frequency from 0.125 Hz to 2 Hz (ablation) | MotionLM · WOMD interactive val (ensemble) | mAP: 0.2007 -> 0.2150 | Improved |
| increase interactive attention frequency from 0.125 Hz to 2 Hz (ablation) | MotionLM · WOMD interactive val (single replica) | mAP: 0.1558 -> 0.1687 | Improved |
| increase number of rollouts from 1 to 512 per replica (ablation) | MotionLM · WOMD interactive val (ensemble) | mAP: 0.1524 -> 0.2150 | Improved |
| use 32 rollouts instead of 512 (ablation) | MotionLM · WOMD interactive val | mAP: 0.2065 vs 0.2150 | No clear effect |
| leak the query agent's future (acausal conditioning) instead of temporally causal conditioning (ablation) | MotionLM · WOMD interactive val (marginal metrics) | soft mAP: 0.4096 -> 0.4274 | Improved |
| condition on the query agent's trajectory (temporally causal) (ablation) | MotionLM · WOMD interactive val (marginal metrics) | soft mAP: 0.3951 -> 0.4096 | Improved |
| scale model from 9M to 27M parameters (ablation) | MotionLM · WOMD interactive val (single replica) | mAP: 0.1663 -> 0.1376 | Hurt |
| reduce model to 1M parameters from 9M (ablation) | MotionLM · WOMD interactive val (single replica) | mAP: 0.1663 -> 0.1713 | No clear effect |
| use joint autoregressive discrete-token model instead of best prior joint model JFP | JFP · WOMD interactive test | mAP: 0.2050 -> 0.2178 | Improved |
| use MotionLM for marginal prediction versus MTR-A | MTR-A · WOMD marginal test | soft mAP: 0.4594 -> 0.4507 | Hurt |
Closest papers by meaning
- Scene Transformer: A unified architecture for predicting future trajectories of multiple agents (ICLR 2022 · similarity 0.89)
- AutoTraces: Autoregressive Trajectory Forecasting via Multimodal Large Language Models (CVPR 2026 · similarity 0.89)
- DeMo: Decoupling Motion Forecasting into Directional Intentions and Dynamic States (NeurIPS 2024 · similarity 0.89)
- Large Scale Interactive Motion Forecasting for Autonomous Driving: The Waymo Open Motion Dataset (ICCV 2021 · similarity 0.89)
- Motion-Agent: A Conversational Framework for Human Motion Generation with LLMs (ICLR 2025 · similarity 0.87)
- Trajeglish: Traffic Modeling as Next-Token Prediction (ICLR 2024 · 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.
