lidingm/SpatialEvo-3B
SpatialEvo-3B by lidingm is a 3 billion parameter vision-language model fine-tuned from Qwen2.5-VL-3B-Instruct, specifically designed for advanced spatial reasoning tasks. It utilizes a novel self-evolving framework with a Deterministic Geometric Environment (DGE) to generate exact ground truth from 3D point clouds, enabling zero-noise online reinforcement learning across 16 spatial reasoning categories. This architecture allows for a shared-parameter policy to co-evolve as both a Questioner and Solver, making it highly effective for complex spatial intelligence applications without manual annotation.
Loading preview...
SpatialEvo-3B Overview
SpatialEvo-3B is a 3 billion parameter vision-language model developed by lidingm, fine-tuned from Qwen2.5-VL-3B-Instruct. Its core innovation lies in the SpatialEvo self-evolving framework, which addresses the challenge of acquiring high-quality spatial reasoning data.
Key Capabilities
- Self-Evolving Spatial Intelligence: Leverages a Deterministic Geometric Environment (DGE) to programmatically compute exact ground truth from 3D point clouds and camera poses, eliminating the need for manual annotations or pseudo-labels.
- Zero-Noise Reinforcement Learning: Enables online reinforcement learning across 16 diverse spatial reasoning task categories with precise ground truth.
- Adaptive Curriculum Learning: Features a lightweight Task Scheduler that drives adaptive curriculum learning based on historical accuracy, without requiring manual stage design.
- Co-Evolving Policy: A single shared-parameter policy functions as both a Questioner and a Solver, optimized via GRPO.
Good For
- Advanced Spatial Reasoning: Excels in tasks requiring deep understanding of 3D environments and spatial relationships.
- Research in Self-Supervised Learning: Ideal for exploring novel methods of data generation and model evolution in complex domains.
- Applications Requiring High Precision: Suitable for use cases where exact ground truth and robust spatial intelligence are critical.
Performance Highlights
SpatialEvo-3B demonstrates strong performance across various spatial benchmarks, achieving the highest average score of 51.1% compared to baselines like SpatialLadder, SpaceR, and SpatialSSRL, all evaluated on Qwen2.5-VL-3B. It notably achieves the best results on RealWorldQA (66.5%), EmbSpatial (61.2%), STARE (36.9%), and ViewSpatial (42.3%).