sim2real.world · gap index · about · arXiv:2611.03937
FlowBench: Counterfactual Imagination Training for Spatial Intelligence
- task
- Counterfactual spatial imagination for navigation and manipulation
- hardware
- not stated
- simulator
- AI2-THOR and ProcTHOR
- transfer
- fine-tuned
- sim result
- 29.5 percentage points
- real result
- 16.9 on real images
The paper does not report a comparable success rate on both sides, so the gap cannot be computed.
What the paper says
fine-tuning Qwen3-VL-8B on FlowBench-Sim-Train raises its accuracy by 29.5 percentage points in simulation and 16.9 on real images, a strong sim-to-real transfer.
fine-tuning Qwen3-VL-8B on FlowBench-Sim-Train raises its accuracy by 29.5 percentage points in simulation and 16.9 on real images, a strong sim-to-real transfer.
What it does not report
- Absolute post-fine-tune accuracy of Qwen3-VL-8B on FlowBench-Sim-Test
- Absolute post-fine-tune accuracy of Qwen3-VL-8B on FlowBench-Real-Test
- Pre-fine-tune accuracy of Qwen3-VL-8B on FlowBench-Real-Test
Extraction is automated; every quote above was checked as a verbatim substring of the paper text. If a row is wrong, say so and the correction is recorded and credited: [email protected].