GeoWeaver: Grounding Visual Tokens with Geometric Evidence before Scene Reasoning

Researchers introduced GeoWeaver, a pre-reasoning framework for vision-language models that grounds visual tokens with geometry before language reasoning.

It builds a multi-level geometry bank and allocates geometry evidence to individual visual tokens to create a stronger grounded representation.

The approach improved performance on spatial and temporal reasoning benchmarks, showing better spatial intelligence can come from earlier geometry grounding.

The work suggests multimodal models may benefit more from geometry-aware visual representations than from adding geometry later in the pipeline.