Fri 13 Nov
LiveGet it free
ZOTPAPER

News without the noise

← All concepts

fine-grained spatial reasoning

The idea is to make AI systems understand not just what objects are in a scene, but exactly where they are, their orientations, distances, and how they touch or support one another. This means combining visual appearance with geometry such as depth, surfaces, and coordinates, so models can answer “is the cup resting on the table?”

Current models often fail at this. Benchmarks show strong object detection but near-chance accuracy on exact angles, distances, and route-level geography; models fall back on object names and text cues. The common remedy is explicit geometry: point clouds for placement, deterministic code for spatial question generation, external tools for metric depth and 3D detection, and localization of spatial defects in generated images.

Papers
69
Released code
3
First seen
Aug 2026
Latest
Nov 2026

66 papers in the last two months, against 3 in the two before.

Start here

Three papers to read first, in the order they came out
  1. 1Replacing language-style decoding cuts robot trajectory errors by 60%

    Replacing autoregressive token decoding with direct continuous regression from frozen VLM hidden states reduces trajectory error by over 60%.

    Oct 2026
  2. 2AI models reconstruct relative layouts better than real-world object positions

    Vision-language models recover a scene's relative layout from multiple photos far better than its true metric scale, and that geometric error, not planning, caps navigation.

    Nov 2026
  3. 3AI agents locate sounds better after training with written reasoning

    A spatial audio agent answers questions, holds dialogue and calls tools on four-channel ambisonic recordings, scoring 44.9 percent on a new 12,000-sample benchmark, 7.4 points above prior systems.

    Nov 2026

How fast it is moving

JunJulAugSeptOctNov so far

Often studied with

vision-language modelA model that handles tasks combining visual and textual information, such as editing images from conversation or reasoning across images and text.36 togetherbenchmarkA standard set of tasks or data used to evaluate and compare the performance of AI systems.18 togetherworld model for agentsA predictive environment model that agents use to simulate outcomes, enabling planning, reasoning, and long-horizon task execution.7 together

Who is working on it

NVIDIA 5Shanghai Jiao Tong University 4Zhejiang University 4Tencent 3Alibaba Group 3Tsinghua University 2

The papers

Most central to this idea first, not most recent.

Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.