AnchorReasoning: A Visual Grounding and Causal Reasoning Dataset in Long-Tail Autonomous Driving Scenarios
AnchorReasoning pairs driving decisions with the visual evidence models are supposed to ground them in.
The dataset is built on WOD-E2E and includes 416,119 annotated frames, with 395,379 decision-critical elements across 19 fine-grained types. Each frame is structured as a visually grounded chain of thought, tying object localization and attributes to driving rationale, action, and trajectory planning. The authors also introduce curriculum supervised fine-tuning and an object-size-aware grounding metric. Across eight model backbones, they report better grounded reasoning and trajectory prediction, with lower average reasoning-token use and inference latency. ArXiv · AI/CL/LG's note
The dataset is built on WOD-E2E and includes 416,119 annotated frames, with 395,379 decision-critical elements across 19 fine-grained types. Each frame is structured as a visually grounded chain of thought, tying object localization and attributes to driving rationale, action, and trajectory planning. The authors also introduce curriculum supervised fine-tuning and an object-size-aware grounding metric. Across eight model backbones, they report better grounded reasoning and trajectory prediction, with lower average reasoning-token use and inference latency. ArXiv · AI/CL/LG's note
score 5