Perception Test 2026: Challenge Summary and Extension to City-scale Audio-Visual Reasoning

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Perception Test 2026".

Jane: The gist The winning solutions in the newly added city-scale tracks demonstrated that complex spatial and multimodal reasoning can be solved by expensive agentic pipelines,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: The core idea here is that they tested spatial intelligence using these new city-scale videos. They created KilometerAudio and KilometerVision to probe city-scale spatial capabilities in frontier models.

Jane: These new tracks use real-world walking tour Youtube videos, up to ten minutes long, which corresponds to about one kilometer of walking distance. It tests how well models can reason about complex spatial and multimodal reasoning in this kind of environment.

Lu: The KilometerVision benchmark uses multiple choice video QA format with three task configurations. They have standard video QA with five text options, multiple choice where the five options are images, like candidate route maps, or even questions that include an image to estimate a distance from a landmark.

Meng: That sounds like it’s pushing models to do more than just simple object recognition; they're asking for spatial understanding and geometric reconstruction based on visual input.

Tom: Exactly. And then there's KilometerAudio, which is novel because it tests the interlink between audio and visual modalities in very long videos, captured over several kilometers around an urban environment.

Jane: That’s a big part of it; they are testing reasoning about an in-view visual along with out-of-view audio as well as the length of these videos. The test set has five hundred five option questions over two hundred eighteen publicly available hour long Youtube walking tour videos.

Lu: The questions in KilometerAudio were collected with human raters who had to identify recognizable sounds in the videos and formulate questions requiring both the audio and visual modalities to be answered, like identifying sound sources in space or timing audio events.

Tom: And they summarize the results by showing that winning solutions on these new city-scale tracks demonstrated that complex spatial and multimodal reasoning can be solved by expensive agentic pipelines.

Jane: But they also point out that this reasoning remains difficult for multimodal models used standalone. That is a key finding of the Perception Test two thousand twenty-six: Challenge Summary and Extension to City-scale Audio-Visual Reasoning paper <ref:2610.12081#pg1,Perception Test 2026: Challenge Summary and Extension to City-scale Audio-Visual>.

Conclusion: Tom: So, looking at the Perception Test two thousand twenty-six: Challenge Summary and Extension to City-scale Audio-Visual Reasoning paper, we see that they put together a lot of different challenges <ref:2610.12081#pg1,Perception Test 2026: Challenge Summary and Extension to City-scale Audio-Visual>. It’s about testing spatial intelligence in real city environments using those new audio and vision tracks.

Jane: The paper highlights how models perform on these tasks, showing that while complex reasoning can be done with expensive agentic pipelines, it’s still hard for models to do it all on their own. That’s a pretty clear summary of what they found.

Lu: What this means for the field is that if you want these kinds of spatial and multimodal abilities, you probably need a system that orchestrates different AI tools with explicit tools to solve the problem.

Meng: From an engineering standpoint, it suggests that building those complex agentic pipelines is necessary for achieving high performance on these large-scale audio-visual tasks.

Tom: And I think the title of the paper really sums it up—Perception Test two thousand twenty-six: Challenge Summary and Extension to City-scale Audio-Visual Reasoning <ref:2610.12081#pg1,Perception Test 2026: Challenge Summary and Extension to City-scale Audio-Visual>. It shows they aren't just looking at single tasks but expanding the test into a whole new realm of reasoning.

Jane: It moves beyond just basic object recognition or simple visual QA and pushes into understanding large, continuous environments using sound too. The implications are that future multimodal models need to be very good at coordinating their internal steps with external tools to succeed.

Lu: We see this paper opens up a path for research focused on how these expensive agentic pipelines can be made more efficient or how the standalone models can integrate better with them.

Tom: It's interesting that they also had a generalist model compete across all tracks and perform satisfactorily, even though the specialized solutions were very strong. That suggests there’s a baseline capability out there.

Jane: That generalist performance is significant because it shows what's achievable without the heavy scaffolding of custom agentic pipelines for every single task.

Meng: So, to wrap up, the Perception Test two thousand twenty-six: Challenge Summary and Extension to City-scale Audio-Visual Reasoning paper tells us that complex spatial and multimodal reasoning is solvable via agentic pipelines but still remains difficult for standalone multimodal models <ref:2610.12081#pg1,Perception Test 2026: Challenge Summary and Extension to City-scale Audio-Visual>.

Lu: That’s the main message we get from this work. It sets a benchmark for what's possible in these challenging real-world reasoning scenarios.