DiSR framework splits 3D perception from reasoning for efficient AI

A new paper from researchers at HUST (Haoze Sun, Jiequan Cui, Qingshan Xu, Richang Hong) proposes DiSR, a framework that explicitly disentangles 3D perception from spatial reasoning. Rather than jointly learning implicit 3D understanding and reasoning through massive end-to-end training, DiSR leverages the complementary strengths of modern perception models—which excel at estimating continuous 3D geometry—and large language models, which are powerful at compositional and symbolic reasoning. The pipeline reconstructs the physical world into structured 3D evidence using off-the-shelf expert perception models, then fine-tunes an LLM with LoRA to perform reasoning solely over that explicit geometric evidence.

DiSR achieves competitive performance on popular spatial reasoning benchmarks without large-scale 3D VQA training or complex tool-use policies. The authors highlight several advantages over end-to-end approaches: improved interpretability (since reasoning operates on explicit, inspectable 3D evidence), modularity (components can be swapped independently), and computational efficiency (no need for massive multimodal training runs). This suggests that explicitly separating perception and reasoning may be a scalable, effective paradigm for spatial intelligence, challenging the assumption that end-to-end modeling is the only path forward.