Research map / Multi-agent systems and games / Human-AI and goal-conditioned
RefAV: Towards Planning-Centric Scenario Mining
Cainan Davidson, Deva Ramanan, Neehar Peri
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 |
|---|---|---|---|
| filter tracks by referred class instead of labelling all tracks referred | ground-truth tracks · RefAV scenario mining | HOTA-Temporal: 13.3 -> 21.4 | Improved |
| use program synthesis (RefProg) instead of LLM API as black box | Le3DE2E tracks · RefAV scenario mining | HOTA-Temporal: 37.2 -> 50.1 | Improved |
| use LLM black box instead of ReferGPT | Le3DE2E tracks · RefAV scenario mining | HOTA-Temporal: 20.2 -> 37.2 | Improved |
| use image-embedding similarity instead of class filtering on predicted tracks | Le3DE2E tracks · RefAV scenario mining | HOTA-Temporal: 19.2 -> 17.2 | Hurt |
| use image-embedding similarity instead of class filtering on oracle tracks | ground-truth tracks · RefAV scenario mining | HOTA-Temporal: 21.4 -> 24.6 | Improved |
| replace ground-truth tracks with predicted Le3DE2E tracks | RefProg · RefAV scenario mining | HOTA-Temporal: 64.8 -> 50.1 | Hurt |
| use Claude 3.7 Sonnet instead of Qwen-2.5-7B-Instruct for code generation (ablation) | RefProg · RefAV scenario mining (RefProg LLM ablation) | HOTA-Temporal: 31.6 -> 50.1 | Improved |
| use Claude 3.7 Sonnet instead of Gemini 2.0 Flash for code generation (ablation) | RefProg · RefAV scenario mining (RefProg LLM ablation) | HOTA-Temporal: 45.2 -> 50.1 | Improved |
| use Claude 3.5 Sonnet instead of Qwen-2.5-7B-Instruct for code generation (ablation) | RefProg · RefAV scenario mining (RefProg LLM ablation) | program failure rate: 18.1 -> 0.5 | Improved |
| swap tracker from Le3DE2E to ReVoxelDet | RefProg · RefAV scenario mining | HOTA-Temporal: 50.1 -> 43.9 | Hurt |
| subsample tracker inputs from 10 Hz to 2 Hz (ablation) | RefProg · RefAV scenario mining | HOTA-Temporal | No clear effect |
| use RefProg zero-shot instead of trained PromptTrack on nuPrompt | PromptTrack (PETR) · nuPrompt (zero-shot RefProg) | AMOTA: 0.259 -> 0.321 | Improved |
Closest papers by meaning
- SmartRefine: A Scenario-Adaptive Refinement Framework for Efficient Motion Prediction (CVPR 2024 · similarity 0.88)
- Future-Aware End-to-End Driving: Bidirectional Modeling of Trajectory Planning and Scene Evolution (NeurIPS 2025 · similarity 0.87)
- Motion Forecasting in Continuous Driving (NeurIPS 2024 · similarity 0.86)
- Hybrid-Driving: An Autonomous Driving Decision Framework Integrating Large Language Models, Knowledge Graphs and Driving Rules (AAAI 2025 · similarity 0.86)
- DriveSpatial: A Benchmark for Spatiotemporal Intelligence in VLMs for Autonomous Driving (NeurIPS 2026 · similarity 0.85)
- STSBench: A Spatio-temporal Scenario Benchmark for Multi-modal Large Language Models in Autonomous Driving (NeurIPS 2025 · similarity 0.85)
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.
