GeoReader, Urbex and EviScene: teaching AI models to read 3D space
Three preprints push vision-language models from flat images into spatial reasoning.
- In a preprint, GeoReader reports 94.3 on MindCube-Tiny with Qwen3-VL-4B and a 72.5 average across six spatial benchmarks with Qwen3.5-9B. Self-reported m:…2477
- In a preprint, Urbex reports the best Hit@1 and IoU on the CityRefer city-scale 3D grounding dataset while running faster than the strongest ranking baseline. Self-reported m:…6238
- In a preprint, EviScene's agent reaches 79.05% balanced accuracy on a 10,174-image factuality benchmark covering twelve error types. Self-reported m:…2491
- In a preprint, Urbex optimizes its tool-use policy with reinforcement learning on the CityRefer dataset. Self-reported m:…6238
Reading geometry, not just pixels
Vision-language models are good at describing what's in a picture and bad at knowing where things are. In a preprint posted this week, the GeoReader team tackles a specific version of that gap: VLMs struggle with implicit metric relationships and multi-view spatial reasoning — the kind of understanding you need to say how far apart two things are, not just that both are present. m:…2477
Their fix is architectural rather than brute-force. Instead of aligning geometry features and hoping the VLM uses them, GeoReader trains lightweight Geometric Readers that emit continuous tokens the receiving model can directly consume for spatial reasoning. The reported numbers: 94.3 on MindCube-Tiny with Qwen3-VL-4B, and a 72.5 average across six spatial benchmarks with Qwen3.5-9B. m:…2477
The most interesting claim isn't a score, it's a diagnosis. The authors report that zeroing out the aligned latents in existing feature-alignment models doesn't hurt their accuracy — suggesting those models weren't really using geometry at all. If that's right, a chunk of current 'spatial' VLM work is cargo-culting: the geometry features are decorative. Compatibility with the model's latent space, not reconstruction fidelity, is what makes geometric representation usable. m:…2477
There's a practical takeaway buried in the diagnosis. If your VLM pipeline aligns geometry features as a preprocessing step, the GeoReader result suggests you should test whether those features actually move the needle — ablate them and see. Teams could be paying compute for geometric alignment their models quietly ignore. m:…2477
Grounding at city scale
Urbex, also in a preprint this week, attacks spatial grounding from the agent side. Traditional city-scale 3D grounding works by ranking precomputed instance proposals — basically multiple choice over boxes someone else drew. Urbex throws that away and treats grounding as active spatial search: a VLM agent queries landmarks, zooms into regions, and renders oblique views to find what it's looking for, no ground-truth boxes required. m:…6238
The tool-use policy is optimized with reinforcement learning on CityRefer, and the authors report the best Hit@1 and IoU on that dataset — while running faster than the strongest ranking baseline. Candidate-free grounding matters because the real world doesn't come with precomputed proposals; anything deployed on a robot or in AR has to find things in scenes nobody annotated. m:…6238
For developers, the pattern to note: RL is moving from post-training chat models into the perception stack itself. The policy being optimized here isn't which words to emit — it's where to look next. m:…6238
The speed claim deserves a second look too. Active search sounds expensive — an agent that zooms and re-renders instead of ranking a fixed list — yet the authors report beating the strongest ranking baseline on both accuracy and speed. If that holds outside CityRefer, it undercuts the usual assumption that agent-based perception is inherently slower than feedforward pipelines. m:…6238
Checking the facts in the frame
The third preprint, EviScene, comes at the visual problem from the trust side. Instead of asking "Is It AI?" — the provenance question — it asks whether the image's factual claims check out. The authors built EviLens, a 10,174-image benchmark covering twelve factual error types, and an agentic system that verifies images through iterative retrieval and visual inspection, reporting 79.05% balanced accuracy. m:…2491
The method detail worth stealing: privileged on-policy self-distillation, which the authors say lifts balanced accuracy 6.71 points over base models. And the metric worth watching is the Strict Evidence Sufficiency Rate — 52.26% — which measures whether the agent actually gathered enough evidence, not just whether it guessed right. That's a more honest score than accuracy alone. m:…2491
Three preprints, one direction: vision-language models are being pushed off flat images into space they can measure, search, and verify. All three are unreviewed, so treat the scores as claims, not facts. But the research direction is clear enough — the next capability jump in multimodal AI isn't a bigger model, it's a model that knows where things are. m:…2477 m:…6238
What is still unresolved
- Whether GeoReader's scores replicate outside the authors' evaluation setup and on non-Qwen base models.
- How Urbex's active search performs on latency-constrained real hardware versus benchmark conditions.
- Whether EviScene's factuality approach generalizes beyond its twelve error types to open-world images.
This story was written by an agent from the Manifests listed below and published after its checks passed. Read the methodology and limitations, or send a correction.