Spatial Atlas: Compute-Grounded Reasoning for Spatial-Aware Research Agent Benchmarks

Arun Sharma

cs.AIApr 13, 2026

Spatial Atlas: Compute-Grounded Reasoning for Spatial-Aware Research Agent Benchmarks

Authors: Arun Sharma

Organizations: University of Minnesota, Twin Cities

Abstract

We describe compute-grounded reasoning (CGR), a design pattern in which code computes selected sub-problems from explicit intermediate representations before a language model answers. Spatial Atlas implements CGR as an Agent2Agent (A2A) server with a spatial question-answering handler and a machine-learning engineering handler. The spatial handler asks a language model to extract a scene graph, and code then fills in missing distances and checks the extracted safety rules. A separate benchmark driver can also run a strict metric bridge. It computes the gap for horizontal-gap questions from segmentation masks and a reconstructed point map, and it passes that gap to the answering model as a fact. The bridge returns a fixed unavailable answer when an evidence check fails, and it never falls back to model-estimated coordinates. The ML-engineering handler generates pipeline code, parses validation scores, and caps the number of repair and refinement passes. Its code execution is off by default. The repository also provides four run modes that can write label-free journals, a shuffled-image control mapping, and journal validators that reject label-bearing fields. We report one private label-free operational run in which four paths each wrote eight prediction rows with zero retries. Labels stayed sealed, and no score was computed, so this run establishes operational integrity only. We report no FieldWorkArena result because the benchmark data were not accessible. We also omit every performance, latency, and resource-use number that lacks a reproducible run artifact.

Figures & tables

Panel of Figure 1 — Figure 1 : Spatial Atlas places two handlers behind one A2A server. The server passes each task to one handler through four fixed dispatch rules, and no model call takes part in dispatch. On the default spatial path, the Strong tier extracts a typed scene graph from the file descriptions, so entity positions are Strong-tier estimates from text. Code computes only the distances that the model did not supply, and the Strong tier answers from the resulting fact sheet. A Fast-tier self-grade below 0.6 triggers one Strong-tier refinement. The strict metric bridge runs only under the frozen benchmark driver, which supplies the raw question and the protocol identifier. It calls SpatialClaw’s tools [ 7 ] for SAM 3 masks [ 3 ] and Depth Anything 3 point maps [ 23 ] . The bridge fails closed and never falls back to the scene graph. The information-gain selector and the tier router exist in code, but no code calls them, so the figure omits them.

Tier	Model	Current use
Fast	GPT-4.1-mini	Self-grade, and entity phrases for the generic metric engine
Standard	GPT-4.1	MLE competition analysis
Strong	GPT-4.1	Scene-graph extraction, spatial answers, refinement, code generation, and code repair
Vision	GPT-4.1	Descriptions of images and video frames

Table 1 : This table lists the configured model tiers and their uses in the current code. It is a design table, and it reports no performance or resource-use result.

Panel of Figure 2 — Figure 2 : The strict metric bridge computes a horizontal gap from object masks and a point map. Each candidate pair of object masks passes through the steps from left to right, first along the top row and then along the bottom row. Here $\operatorname{pct}_{5}$ is the fifth percentile with linear interpolation, as in Eq. 12 . The erosion radius $r_{e}$ is the ceiling of the larger resolution ratio between a mask and the point map. Each mask is first resized to the point-map grid by nearest-neighbor sampling. The overlap coefficient is $\rho=\lvert E_{a}\cap E_{b}\rvert/\min(\lvert E_{a}\rvert,\lvert E_{b}\rvert)$ for the eroded masks $E_{a}$ and $E_{b}$ . The bridge returns QSPATIAL_UNAVAILABLE at once when the parser marks the question unsupported or when segmentation returns too few masks. A parse or manifest failure raises an error before the steps in this figure begin. Every dashed exit drops only the current pair, and the bridge returns QSPATIAL_UNAVAILABLE when no pair stays valid. The confidence filter has no exit of its own, because a mask that loses its points fails the voxel count. A segmentation or reconstruction service failure raises an error, and the bridge never falls back to the scene graph. A valid gap enters the fact sheet, and the Strong tier writes the final answer from it.

Strategy name	Task type	Template contents
tabular	Tabular classification or regression	AutoGluon TabularPredictor with a 300-second limit, which falls back to LightGBM when AutoGluon is not installed
nlp	Text classification	TF-IDF unigram and bigram features (up to $50,000$ ) with logistic regression and a cross-validation check
vision	Image classification	Flattened $32\times 32$ pixel features with a tree-ensemble classifier
timeseries	Forecasting	Lag features (1, 7, 14, and 28 steps), rolling features, and a LightGBM regressor
general	Mixed or unknown	A gradient-boosting classifier or regressor

Table 2 : This table lists the strategy templates that the handler inserts into the code-generation prompt. The Strong tier writes the final script, so a template guides the code and is not the executed code.

Panel of Figure 3 — Figure 3 : Four operational paths ran together in the V37 run. The paths were the scene-graph path, question-only baseline, correct-image metric path, and schema-matched shuffled-image metric control. Each path produced eight label-free prescore prediction rows, so the run wrote 32 rows with zero retries. The execution, warning-scan, restoration, cleanup, and terminal-seal gates passed. In the public driver, the shuffled-image mode swaps each attached image through a control mapping $\phi$ . That mapping must hold 84 pairs, be a bijection, and have no fixed point, and the loader checks the digest of every source and control image. The terminal finalizer ran once and wrote the smoke seal and the terminal result. Protected labels and ground truth remained sealed, and no score or paired comparison was computed, so the run establishes operational integrity only. The run is private, and the public repository cannot reproduce it.

Explore similar work

Nov 10, 2025cs.CV

SpatialThinker: Reinforcing Scene Graph-Grounded Spatial Reasoning via Dense Rewards

Multimodal large language models (MLLMs) have achieved remarkable progress in vision-language tasks, but continue to struggle with spatial reasoning. Existing spatial MLLMs rely on large-scale datasets, explicit 3D inputs, architecture-specific modifications, or sparse Reinforcement Learning (RL) methods that provide insufficient guidance for spatially-grounded reasoning. We introduce SpatialThinker. To our knowledge, it is the first MLLM unifying Scene Graph Generation (SGG) and visual reasoning in a single pass via online RL. The model simulates human-like spatial perception by constructing a mental scene graph of task-relevant objects and relations, and reasoning toward an answer via dense spatial rewards. Our contributions are threefold: (1) SGG-grounded reasoning: integrating SGG directly within the reasoning chain rather than as a disjoint preprocessing step; (2) STVQA-7K: a high-quality spatial VQA training dataset via a scalable synthesis pipeline; and (3) a dense spatial reward design that enforces structured grounding during RL and generalizes to improve broad visual perception. SpatialThinker-7B achieves 3.6

\times

larger gains over SFT and

1.7\times

better in- and out-of-distribution generalization than sparse RL. Trained on only 7K samples, SpatialThinker-7B matches GPT-5 and outperforms GPT-4o, while SpatialThinker-30B surpasses both GPT-5 and Claude 4 Sonnet on average across 14 spatial and real-world benchmarks, demonstrating that structured spatial grounding with reward-aligned reasoning enables robust spatial understanding with limited data.

Hunar Batra, Haoqin Tu, Hardy Chen +3

University of Oxford · University of California, Santa Cruz

Spatial Reasoning Spatial Grounding Multimodal Large Language Models Geospatial Reasoning Scene Understanding

Ask PDF

Aug 4, 2026cs.CL

MultiGlobeQA: A Multilingual and Globally Diverse Benchmark for Geospatial Reasoning

Geospatial reasoning, i.e., computing distances, containment, and other spatial relations over real-world entities, is central to navigation and logistics, yet large language models (LLMs) struggle with the required geometric and topological computation despite storing considerable geographic knowledge. Existing benchmarks localize these failures only partially: they are synthetic or smallscale, largely monolingual, and offer limited control over geographic coverage. We introduce MultiGlobeQA, a multilingual benchmark of 46,060 question-answer pairs spanning 14 spatial-function families and 15 answer formats, with execution-based ground truth over three knowledge graphs. It covers 201 countries and territories via income- and density-stratified sampling, with parallel questions in English and 16 additional high- and low-resource languages. Across parametric, reasoning, and agentic settings, LLMs collapse on tasks requiring grid indexing and shape computation, while topological relations and directions fare best. Retrieval and tool use yield considerable gains, yet performance plateaus below two thirds even when gold facts are supplied, indicating that computation, not access to knowledge, is the bottleneck. Models also underperform on low-income regions, a gap that gold facts widen rather than close.

Martin Böckling, Elizaveta Nosova, Heiko Paulheim +1

Data and Web Science Group, University of Mannheim, Germany

Geospatial Reasoning Multilingual Benchmark

Ask PDF

Aug 3, 2026cs.CV

SpatialQuery: Benchmarking Geometry-Grounded Multi-Instance Spatial Reasoning in Vision-Language Models

Vision-language models (VLMs) achieve strong semantic understanding but remain unreliable in metric spatial reasoning, particularly when queries require comparing multiple instances of the same object category. We study this problem through the Closest-Instance Distance Query (CIDQ), where a model must identify the nearest visible candidate to a unique reference object and estimate their gravity-aligned floor-plane distance. We introduce SPATIALQUERY, a training- free framework for CIDQ reasoning from a single RGB image, together with SPATIALQUERY-1M, a benchmark containing over one million RGB-only question-answer pairs from 200 indoor scenes. SPATIALQUERY recovers instance-level metric geometry and transforms it into a canonical Bird's-Eye View through Scene Cubifying, which represents objects as uniformly sized, category-coded blocks to emphasize their relative floor- plane locations. We further propose Uncertainty-Aware Chain-of-Thought (UA-CoT) prompting, which incorporates geometry- derived per-instance uncertainty into the VLM reasoning process. Without task-specific fine-tuning or architectural modification, SPATIALQUERY with Qwen3-VL-8B achieves a Floor-MAE of 0.259 m, an Unc-Acc@0.3 m of 90.5%, and a proximity-decision accuracy of 84.18%, outperforming fine-tuned spatial specialists, general-purpose VLMs, and closed-source frontier models. Code, benchmark resources, and an interactive demo are available at https://namhai1810.github.io/SpatialQuery/.

Hai Nguyen, Tung Vu, Cong Tran

Posts and Telecommunications Institute of Technology, Hanoi, Vietnam

Spatial Reasoning Approximate Nearest-Neighbor Search

Ask PDF