Bold claim: AI still can’t tie its own knots, even in a 3D world. New research from Cornell shows that while current models excel at processing large blocks of text, they stumble when asked to reason spatially about physical tasks. In a study introducing KnotGym—a 3D reinforcement learning and generalization test—the team demonstrated that AI can untangle simple knots with surprising reliability but struggles to tie knots from loops or transform one knot into another, especially as the knot complexity grows.
KnotGym places AI agents inside a 3D simulator where they see straightforward loops and various knots and must perform actions to unknot, tie, or convert knots. The test includes a gradual difficulty ramp called a generalization ladder, enabling researchers to push models beyond their training regime and observe how well they generalize. According to lead author Zoë (Zizhao) Chen, this kind of spatial reasoning capability is largely missing from contemporary AI benchmarks, which tend to focus on text and 2D vision tasks.
In the experiments, untangling simple knots—up to four crossings, including the classic shoelace knot—was where AI performed best, achieving about 90% success. But tying knots with two crossings was about 83% successful, dropping sharply to 16% for knots with three crossings. For knots with more than three crossings, the models essentially failed, and conversion tasks mirrored this trend.
Chen argues that AI lacks the exploratory learning mindset that humans use when playing with puzzles or toys. “When kids play with a Rubik’s Cube, they experiment, reuse prior knowledge, and gradually discover sequences that reach a goal without breaking other constraints,” she explains. “That kind of exploratory, cumulative learning is what we want AI to develop—but it isn’t there yet.”
Looking ahead, the researchers plan to accelerate KnotGym’s evaluations by running it on GPUs, which can dramatically speed up simulations and model testing. They also intend to expand the environment to cover more complex manipulation tasks relevant to robotics and other real-world AI applications.
The work, partially funded by the National Science Foundation, Open Philanthropy, Nvidia, and NAIRR, appears in a NeurIPS presentation titled Knot So Simple: A Minimalistic Environment for Spatial Reasoning. This line of inquiry invites broader questions about how we design AI systems that can think and act in the physical world, not just reason about words and images. Do you agree that spatial reasoning should be a core focus for future AI development, or is text and image understanding still the top priority? And if you’re curious about the practical implications, what real-world tasks would you like to see AI tackle first with improved 3D reasoning?
