Apple's interest in LLMs and their applications in space computing shows no signs of slowing down, even as some claim that the Apple Vision Pro is dead. In April 2026, it was claimed that the Apple Vision Pro was a complete failure and that, as a result, we would never see a successor product. This rumor, while always seemingly absurd, has now been called into question.
See also: Google discovered zero-day attacks created with the help of AI
While the company’s Vision Product Group may have seen some changes, there’s still hope for a new generation of the Apple Vision Pro. Apple’s research into artificial intelligence suggests the company hasn’t abandoned its space-related projects. New papers published on the Apple Machine Learning explore the use of large language models (LLMs) in sign language analysis, 3D head modeling, and more.
Apple researchers also developed a new assessment system to assess the spatial-functional intelligence of LLMs.
The paper , titled “From Where Things Are to What They Do: Assessing Spatial-Operational Intelligence for Multimodal LLMs,” describes a new testing and scoring system for multimodal LLMs. Apple researchers developed an assessment framework that tests the spatial reasoning abilities of these models.
As the study explains, to mimic human understanding of a space and its objects, AI models rely on two distinct structures: a spatial representation that captures the layouts of objects and the relational structure, and a functional representation that encodes capabilities, purposes, and usage depending on the context. In other words, a multimodal LLM must understand the geometry of a given space, along with the purpose and location of objects within it.
Apple researchers say that existing evaluation methods, such as VSI-Bench, only test the first aspect, largely ignoring the second.
See also: Canonical: AI features in Ubuntu Linux distro
To address this issue, they developed the Spatial-Functional Intelligence Benchmark, abbreviated as SFI-Bench. It is described as a video-based benchmark with 1,555 expert-annotated questions derived from 134 internal video scans.
SFI-Bench tests whether models understand what objects in the scene are for, how they work, and how failures can be diagnosed. In other words, the benchmark tests whether artificial intelligence models understand what an object is, where it is, how it is used, what it is used for, and how it can be repaired.
If this sounds familiar, it's because Google has had tools with this kind of spatial awareness since at least 2024. At the I/O conference that same year, Google's AI model correctly identified an object in front of it as a turntable and even suggested how to repair the device.
In practice, SFI-Bench will serve to test similar and more advanced AI models. Some of the tests mentioned include asking an LLM to identify the largest subset of the same brand of bottles in a cupboard, cancel the current programme on a washing machine and what a TV remote control is used for.
Apple researchers tested several open-source and proprietary AI models with the SFI-Bench framework. As expected, Google Gemini 3.1 Pro achieved the best overall result, while Gemini-3.1-Flash-Lite came in third. GPT-5.4-High came in second.
See also: Artificial Intelligence is Transforming DevSecOps
However, the study notes that “Across all models, the global counting condition emerges as a key obstacle, revealing persistent limitations in synthetic and logical reasoning.” In other words, most current multimodal LLMs struggle with spatial memory, integrating functional knowledge, and linking perception to external knowledge.
