AIBRIEF
A new training free method called ByDeWay V2 injects plain language spatial relations like "left of" and "touching" into vision language model prompts, with no model retraining, so a robot or safety monitor can check the answer instead of trusting
A vision model that misplaces a cup on a counter, or tells a robot the obstacle is on the wrong side, is the kind of error that decides what AI is allowed to do in the real world. A new patch called ByDeWay-V2 does not retrain the model. It gives the model a structured "where is it" trail written in plain predicates like "left of," "above," "inside," and "touching," so a human or a safety monitor can read the answer instead of trusting it.
An object detector, YOLO-World-L, computes pairwise geometric relations between the objects it spots in an image. Those relations are injected as human-readable spatial predicates into the prompt given to a multimodal large language model (a vision-language model that can also reason about text), alongside the depth cues the model already sees. The model itself does not change. The lightest configuration runs on a CPU inside a roughly 40-token context budget, framed as suitable for resource-constrained, real-time decision support.
On BLINK, a visual-reasoning benchmark, the patch posts a 46 percent relative F1 improvement over the depth-prompting baseline on BLINK's spatial subset for Qwen2.5-VL. On VSR, a visual-spatial-reasoning benchmark, BLIP-Base recovers from near-random spatial reasoning to an F1 of 0.53. Hallucination is checked via POPE, a polling-based object-existence probe. The numbers are paper-claim rather than deployment-claim.
The source is a single non-peer-reviewed arXiv preprint. All quantitative claims live inside benchmark subsets, and the framework is a research artifact rather than a shipped system. If independent reproduction or peer review shows the predicates do not generalize beyond VSR, BLINK, and POPE, the legibility claim collapses.