VoxelSage forbids the language model to calculate the measurements itself, and divides by almost two simulated blood loss

On 256 new simulator scenes, the liver resection planning tool reduces simulated bleeding by 38.89 percent — a system integration result, not evidence of clinical effectiveness.

Evaluating a liver tumour before the operation requires segmentation, physical space measurements, visual evidence production and resection planning, all of which on the same volume of CT. Existing tools often deal with these steps separately, and language models cannot reliably calculate a physical measurement from a scanner. VoxelSage, described by Binghong Qian, Xuanhe Liu, Yifan Xing, Wenjie Deng and their co-authors, responds with a strict separation of roles.

The architecture is at two ports. Port A interprets the demand and chooses the skills to mobilize; port B applies them to scanner volumes and segmentation masks, and returns structured results. By limiting physical measurements to port B, the system prevents the language model from producing them itself, thereby reducing the risk of manufactured numbers. Eight competencies are integrated — quantitative analysis, visual evidence generation, three-dimensional reconstruction, segmentation refinement, sequential planning — and the user can add them. For the order of the resections, a neuronal binder learned by cloning behavior sorts the candidate targets, and a guard behind the simulator confronts them with predefined constraints.

Of 256 simulator scenes not seen in training, the average simulated duration increased from 34.274 to 33.388 minutes, a modest gain of 2.59%, while the average simulated blood loss fell from 300.847 to 183.852 millilitres, a decrease of 38.89 percent from a deterministic reference. The authors themselves set the limit: these figures show system integration and performance at simulator level, not clinical efficacy or safety. The code is published.