A new research framework teaches AI models to compare 3D objects side by side, not just describe them one at a time.
Most 3D language models can describe a single object well enough. Ask them to compare two objects in detail, though, and they fall apart. The researchers built three things to fix that: an instruction dataset called MO3D that drills fine-grained, multi-object comparisons; a model called Multi-3DLLM that uses a lightweight add-on, a Patch-Interaction Transformer, to track how parts relate within and between objects while keeping their shapes intact; and two practical tests, Shape Mating (does this piece fit that one) and Change Captioning (what changed between two scans of an object). Existing 3D-LLMs and 2D vision-language models both stumble on these tasks, the team found, because none of them were built to compare things or reason about geometry.
That comparison gap matters anywhere spatial judgment beats simple labeling - a robot picking the matching part, a CAD tool flagging what changed in a design, an AR app checking whether pieces align. Trained on the new dataset, Multi-3DLLM beat every baseline on the MO3D benchmark and, as a bonus, got better at plain single-object classification too - a sign the comparison training sharpens geometric understanding generally rather than just teaching a narrow trick.
Whether that improvement holds outside a benchmark the same team designed is the harder question - and the one that decides if this is a lab result or a building block.