GAGR-Lab: Evaluating Joint Spatial-Geometric and Analytic Function Reasoning

Jingyao Zhang; Yun Li; Lu Han

cs.AIOct 7, 2026

GAGR-Lab: Evaluating Joint Spatial-Geometric and Analytic Function Reasoning

Authors: Jingyao Zhang, Yun Li, Lu Han

Abstract

Joint spatial-geometric and analytic function reasoning requires translating a perceived spatial configuration into a symbolic function whose executed curve satisfies geometric constraints. We present GAGR-Lab, a framework for measuring this capability through Cartesian game scenes, explicit function semantics, and authoritative Rust trajectory execution. It distinguishes spatial perception, metric grounding, geometric relations, function interpretation, function construction, and constrained synthesis. We specify four configurable scene-difficulty presets and a prospective 24-cell diagnostic design, while reporting only the subset actually evaluated. A bounded pilot of one hosted model (Llama 3.2 11B Vision Instruct) using two API credentials as execution replicas yields 72 balanced games with 432 attempts, 429 valid provider responses, and no target hits; exploratory ordinary-function prompt variants also fail to hit, while the structured localization interface yields no scoreable outputs. A privileged analytic search control independently succeeds on 600 directional cases from 300 generated scenes, with exact repeatability and 1,200 successful vertical-reflection or translation checks. The framework separates serving reliability, symbolic compliance, and geometric success, and preserves exact model-visible inputs and realized paths. A staged protocol outlines diagnostic calibration, held-out replication, multi-model comparison, and paired robustness tests. The contribution is an operational research framework with an executed pilot and a clearly identified prospective study plan; the full difficulty matrix and comparative model results remain untested.

Figures & tables

Dimension	Diagnostic question	Proposed observable score
Entity perception	Which objects are shooter, target, and obstacles?	Entity accuracy; pixel-region error
Metric grounding	Where are those objects in world coordinates?	Coordinate error; scoreable-output coverage
Geometric relations	What are their directions, slopes, distances, and blocking relations?	Numeric error; relation accuracy
Function to geometry	What trajectory will a supplied function produce under anchoring?	Predicted-path error; outcome accuracy
Geometry to function	Which function realizes specified endpoint relationships?	Engine hit; target approach; validity
Constrained synthesis	Which function reaches the target while satisfying obstacle and path constraints?	Hit before collision; clearance; constraint satisfaction

Table 1: Proposed component tasks for joint spatial-geometric and analytic function reasoning. Independent component scores are not all available in the current pilot.

Preset	Player $y$ span	Obstacles	Size multiplier	Corridor half-width
Easy	$\pm 3$	0–1	0.80	4.00
Medium	$\pm 6$	2–4	1.00	3.00
Hard	$\pm 9$	4–8	1.25	2.00
Extreme	$\pm 11$	6–12	1.50	1.25

Table 2: Implemented default difficulty controls in world units, except the dimensionless size multiplier. These presets have not been evaluated as a four-level model difficulty sweep in this study.

Component task	Easy	Medium	Hard	Extreme
Entity perception	P	P	P	P
Metric grounding	P	P	P	P
Geometric relations	P	P	P	P
Function to geometry	P	P	P	P
Geometry to function	P	P	P	P
Constrained synthesis	P	P	P	P

Table 3: Prospective 24-cell diagnostic matrix. P means planned, not passed or evaluated. The current custom-profile pilot is reported separately and is not substituted for this matrix.

Family	Input	Games	Attempts	HTTP 200	Valid
Horizontal	P	12	72	71	71
Horizontal	P+C	12	72	71	71
Slanted	P	12	72	72	72
Slanted	P+C	12	72	72	72
One obstacle	P	12	72	71	71
One obstacle	P+C	12	72	72	72

Table 4: Primary pilot counts. P is the pixel condition; P+C adds exact endpoint coordinates. Repeated assignments share scene geometry and are not independent scene samples.

One-shot condition	Requests	Format valid	Engine valid	Hits
Canonical	24	24	24	0
No full examples	24	24	24	0
Anchoring equation	24	24	24	0
Metric JSON	36	0	0	—

Table 5: Exploratory diagnostics. Metric JSON actions were rejected before execution, so their hit outcome and localization accuracy are unavailable, not zero.

Offline family	Directional cases	Direct-line hits	Search-control hits
Horizontal	200	200	200
Slanted	200	200	200
One obstacle	200	0	200

Table 6: Independent engine controls. The search uses exact geometry and up to nine engine previews, which the model does not receive. These are execution checks, not inter-agent rankings.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 1: Executed model trajectories (red) and privileged analytic controls (blue) on two primary scenes. The model shots exit the world; the controls enter the target. The obstacle is shown in black. Coordinates and trajectories come from the archived scene and authoritative traces, rather than from model explanations. Both panels use the same world scale.

Legacy cohort	Games	Attempts	Engine valid	Hits	Draws
H1	6	34	26	4	2
H2	12	180	141	4	8
H3	12	221	9	1	11

Appendix

Table 7: Retrospective legacy counts. Engine valid refers to the historical parser/extraction policy, not the current strict full-response output gate. Rows are descriptive cohorts, not comparable benchmark scores.

Explore similar work

Apr 13, 2026cs.AI

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

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.

Arun Sharma

University of Minnesota, Twin Cities

Tool-Grounded Reasoning Atlas Agent2Agent

Ask PDF

Jun 3, 2026cs.LG

From Symbolic to Geometric: Enabling Spatial Reasoning in Large Language Models

Recent large language models (LLMs) often appear to exhibit spatial reasoning ability; however, this capability is largely \emph{symbolic}, arising from pattern matching over spatial language rather than true \emph{geometric} reasoning over space. Because LLMs operate on discrete tokens, they lack native support for continuous spatial representations, explicit geometric computation, and structured spatial operators. To address this limitation, we introduce the \emph{Spatial Language Model (SLM)}, the first multimodal LLM that treats location information as a first-class modality and enables geometric spatial reasoning within the model's inference process. SLM directly operates on learned spatial representations rather than textual descriptions of spatial relations. To support effective training, we construct a \emph{Spatial Instruction Dataset} that aligns spatial representations, atomic geometric operations, and natural language instructions. We further propose a new benchmark named \emph{SpatialEval}, which is designed to evaluate spatial reasoning across attributes, distance, topology, and relative-position tasks. Extensive experiments show that SLM significantly outperforms existing LLM-based approaches that rely on symbolic reasoning via prompt engineering or textual abstraction, demonstrating the benefits of integrating geometric spatial representations for robust spatial reasoning. Our instruction dataset, evaluation benchmark, model training codes, and models' checkpoints can be found at: \hyperlink{https://github.com/chuchen2017/SLM}{https://github.com/chuchen2017/SLM}.

Chen Chu, Bita Azarijoo, Li Xiong +2

University of Southern California Los Angeles, CA, USA · Emory University Atlanta, GA, USA · Novateur Research Solutions Ashburn, Virginia, USA

Spatial Reasoning LLM Reasoning Strategies Multimodal Large Language Models Symbol-

Ask PDF

May 19, 2026cs.AI

GeoX: Mastering Geospatial Reasoning Through Self-Play and Verifiable Rewards

Geospatial reasoning requires solving image-grounded problems over the complex spatial structure of a scene. However, developing this capability is hindered by the cost of annotating a vast and combinatorial question space. We propose GeoX, a self-play framework that acquires spatial logic through executable programs that yield verifiable rewards, without relying on large-scale human-curated data Given a satellite or aerial image, our framework employs a single multimodal policy that proposes spatial problems as executable programs and solves them under three reasoning modes-abduction, deduction, and induction-over spatial primitives and an image understanding tool. A verifier executes each program to covert a reward signal that jointly optimizes the two roles via reinforcement learning. GeoX consistently improves its base VLMs by up to 5.5 points on average, matching or exceeding conventional baselines trained on millions of curated data. Along-side the proposed method, we release a benchmark for geospatial understanding accumulated through self-play.

Kyeongjin Ahn, Seungeon Lee, Krishna P. Gummadi +1

KAIST · MPI-SWS · MPI-SP

Geospatial Reasoning Geogs-Slam Self-Play Inductive Bias Verifiable Rewards

Ask PDF