Visual puzzles force an artificial intelligence system to see what is in an image and then apply rules to get an answer. If its visual assumptions are baked into the solving code, either mistake is hard to find or fix.
Vis2State keeps those jobs separate. It first writes down an explicit description of the objects, rules, and goal, then writes a separate program to solve that description. Each part can be revised independently using feedback from re-inspecting the image or running the program.
This separated design beat direct step-by-step reasoning by up to 15.5 percentage points on standard benchmarks. That matters because separating a problem statement from its solver makes visual reasoning easier to debug and more reliable.