Toward AI-Friendly Cartography: Understanding How Color Design Influences Foundation Model Spatial Reasoning on Sequential Choropleth Maps

Yonghe Sun; Zhenjia Liu; Hua Liao; Wenjia Xu; Nai Yang; Weihua Dong; Zhiwei Wei

cs.AIAug 16, 2026

Toward AI-Friendly Cartography: Understanding How Color Design Influences Foundation Model Spatial Reasoning on Sequential Choropleth Maps

Authors: Yonghe Sun, Zhenjia Liu, Hua Liao, Wenjia Xu, Nai Yang, Weihua Dong, Zhiwei Wei

Organizations: School of Geographic Sciences, Hunan Normal University, Changsha, China. · Hunan Key Laboratory of Geospatial Big Data Mining and Application, Changsha, China. · School of Information and Communication Engineering, Beijing University of Posts and Telecommunications, Beijing, China. · School of Geography and Information Engineering, China University of Geosciences, Wuhan, China. ·

Abstract

Foundation models (FMs) increasingly support multimodal and geospatial reasoning, yet it remains unclear whether cartographic principles designed for human perception are equally effective for machines. Focusing on sequential choropleth maps, we examine how hue palette, color ordering, and lightness contrast influence FM spatial reasoning. We construct a controlled benchmark of 5,760 maps and 28,800 questions spanning Attribute Identify, Spatial Recognition, Compare, Rank, and Pattern Delineate, and evaluate 21 open-source and proprietary multimodal FMs. Results show that hue choice has limited and inconsistent effects, whereas disrupting sequential color ordering substantially reduces performance, especially for comparison and ranking. Reduced lightness contrast also consistently impairs reasoning, while increasing contrast beyond sufficient separability provides only marginal gains. LoRA fine-tuning improves overall accuracy but preserves these relative sensitivities. Additional factorial experiments further indicate that errors arise from color-and-legend decoding, spatial reasoning, and the integration of thematic attributes with spatial structure. These findings show that conventional sequential ordering and sufficient contrast remain important for machine map understanding and provide empirical guidance for AI-friendly cartographic design.

Figures & tables

Figure 1: The overview of the framework.

Figure 2: Sequential(a) and randomized(b) choropleth encoding under identical spatial distributions. The randomized condition preserves the original color set while disrupting ordinal color–value correspondence.

Figure 3: Examples of standard(a), high-contrast(b), and low-contrast(c) sequential choropleth maps with representative reasoning tasks.

Task dimension	Question Subtype	Question Description
D1. Attribute Identify	attr2region	Q1. Given an attribute level, identify the corresponding region.
	region2attr	Q2. Given a specific region, identify which attribute level it belongs to.
D2. Spatial Recognition	Direction recognition	Q3. Determine the directional spatial relationship between two regions.
	Adjacent recognition	Q4. Determine whether two regions are adjacent or non-adjacent.
D3. Compare	Attribute comparison	Q5. Compare the attribute values of two given regions and determine which is higher, lower, or equal.
D4. Rank	Global rank	Q6. Find the region with the highest (or lowest) attribute value across the entire map.

Table 2: Tasks design, inherited from ChoroplethMap-Bench [ 55 ] .

Participant	Overall	Sequential	Randomized	Standard	High Contrast	Low Contrast
P1	96.8	99.0	92.0	98.0	98.0	97.0
P2	96.6	96.0	94.0	97.0	99.0	97.0
P3	97.8	97.0	98.0	96.0	99.0	99.0
P4	97.2	97.0	94.0	98.0	99.0	98.0
P5	96.4	96.0	96.0	98.0	97.0	95.0
P6	84.2	89.0	76.0	86.0	88.0	82.0

Table 3: Human validation results on a randomly sampled subset of the bench.

Model	Blues	BuGn	BuPu	GnBu	Greens	Greys	OrRd	Oranges	PuBu	PuBuGn	PuRd	Purples	RdPu	Reds	YlGn	YlGnBu	YlOrBr	YlOrRd
Qwen3.5-9B	66.2	70.9	71.2	67.2	69.4	68.8	68.8	68.8	70.9	70.6	68.4	66.9	65.6	69.1	67.5	68.1	67.2	68.8
Qwen3.5-4B	61.9	60.0	64.1	63.7	62.2	62.8	62.5	60.0	59.7	66.6	59.4	63.7	64.7	64.7	65.0	61.3	61.3	62.8
Qwen3.5-2B	46.9	51.2	53.8	48.1	50.0	49.4	46.2	46.2	50.9	50.3	50.3	47.2	50.6	50.6	51.2	49.7	50.3	48.8
Qwen3-VL-8B-Instruct	60.0	62.8	64.1	59.1	67.2	58.8	69.1	60.3	61.6	62.5	57.5	58.8	62.5	65.3	60.9	64.7	61.6	61.6
Qwen3-VL-4B-Instruct	55.9	59.7	62.2	56.6	62.5	55.6	62.5	58.4	55.9	58.4	55.0	56.9	61.3	61.6	59.7	60.3	59.1	56.9
Qwen3-VL-2B-Instruct	46.9	43.8	41.6	42.2	43.8	43.8	44.1	44.1	41.2	44.7	39.1	42.5	44.1	47.5	42.2	41.6	45.3	43.4

Table 4: Average accuracy (%) under different sequential hue palettes. Repeated-measures ANOVA:

p<0.001

. Post-hoc multiple comparisons reveal no significant differences between individual hue pairs.

Model	Sequential (%)	Randomized (%)	$\Delta$ (%)	Relative Drop (%)
Qwen3.5-9B	68.6	53.6	-15.0	-21.9
Qwen3.5-4B	62.6	49.1	-13.5	-21.6
Qwen3.5-2B	49.5	41.4	-8.1	-16.4
Qwen3-VL-8B-Instruct	62.1	48.0	-14.2	-22.7
Qwen3-VL-4B-Instruct	58.8	45.8	-13.0	-22.1
Qwen3-VL-2B-Instruct	43.4	36.3	-7.1	-16.4

Table 5: Performance comparison between sequential and randomized color encoding.

\Delta

indicates the accuracy drop after disrupting sequential ordering. Paired T-Test:

p<0.001

, Cohen’s

d=1.762

Figure 4: Per-model accuracy under sequential vs. randomized color encoding. Vertical line segments indicate the magnitude of degradation; models are ordered by gap size (largest at left). Dashed horizontal lines show human accuracy under each condition for reference.

Model	Standard (%)	High Contrast (%)	$\Delta_{L+}$ (%)	Low Contrast (%)	$\Delta_{L-}$ (%)
Qwen3.5-9B	70.8	70.9	+0.1	67.6	-3.2
Qwen3.5-4B	63.6	63.6	0.0	60.2	-3.4
Qwen3.5-2B	51.2	52.0	+0.8	49.1	-2.1
Qwen3-VL-8B-Instruct	63.3	64.1	+0.8	60.3	-3.0
Qwen3-VL-4B-Instruct	59.7	60.2	+0.5	56.7	-3.0
Qwen3-VL-2B-Instruct	44.5	44.7	+0.2	42.6	-1.9

Table 6: Performance comparison under different luminance contrast conditions.

\Delta_{L+}

and

\Delta_{L-}

indicate the performance change relative to the standard luminance condition. Friedman test results:

p<0.001

, Cohen’s

f=0.084

. Post-hoc multiple comparisons: Standard vs. Low Contrast (

p=0.004

, Cohen’s

d=0.152

), High vs. Low Contrast (

p<0.001

, Cohen’s

d=0.193

Figure 5: Per-model accuracy under Standard, High Contrast, and Low Contrast conditions, ordered by the High-to-Low range (largest at left). Standard and High Contrast markers nearly overlap for most models, illustrating diminishing returns from contrast enhancement, while Low Contrast is consistently pulled downward. Dashed horizontal lines show human accuracy for reference.

Hypothesis	Contrast	Odds Ratio	$z$	$p$
H1 (Hue)	17 palettes vs. Blues	—	$\|z\|<1.32$	all $>.18$
H2 (main)	Sequential vs. Randomized	1.49	46.96	$<.001$
H2 (random slope)	Sequential vs. Randomized	1.48	5.90	$<.001$
H2 (+ country)	Sequential vs. Randomized	1.49	46.96	$<.001$
H2 $\times$ D1	Sequential vs. Randomized	1.24	10.12	$<.001$
H2 $\times$ D2	Sequential vs. Randomized	1.01	0.45	.656

Table 7: Summary of GLMM robustness analyses at the individual question level. All models include crossed random intercepts for the base map and random intercepts for models nested within model family; the H2 random-slope model additionally includes model-family-specific random slopes for condition. For all rows, odds ratio

>1

indicates higher accuracy under the first-named condition, and

z

is signed accordingly.

	D1		D2		D3		D4		D5
Model	Seq.	Rand.	Seq.	Rand.	Seq.	Rand.	Seq.	Rand.	Seq.	Rand.
Qwen3.5-9B	84.8	80.3	75.2	75.9	59.7	35.3	76.0	37.8	47.2	38.7
Qwen3.5-4B	74.8	71.4	70.5	68.8	55.9	33.6	66.9	33.4	44.7	38.5
Qwen3.5-2B	64.5	60.8	56.4	55.9	42.4	35.2	56.5	28.0	28.0	27.0
Qwen3-VL-8B-Instruct	75.4	68.5	73.2	72.5	54.4	32.9	75.1	34.7	32.5	31.6
Qwen3-VL-4B-Instruct	66.7	62.7	66.2	66.7	50.9	32.0	72.2	33.1	38.0	34.4

Table 8: Performance comparison between sequential and randomized encoding across different task dimensions. Wilcoxon Signed-Rank Test, D1 (

p<0.001

, Cohen’s

d=0.217

), D3 (

p<0.001

, Cohen’s

d=1.287

), D4 (

p<0.001

, Cohen’s

d=1.843

). Paired T-Test: D2 (

p=0.476

, Cohen’s

d=0.16

), D5 (

p<0.001

, Cohen’s

d=0.98

	D1			D2			D3			D4			D5
Model	Std.	L+	L-	Std.	L+	L-	Std.	L+	L-	Std.	L+	L-	Std.	L+	L-
Qwen3.5-9B	90.3	93.0	85.0	74.3	74.1	73.4	61.4	60.9	59.5	79.8	78.1	73.9	48.2	48.4	46.2
Qwen3.5-4B	79.5	82.3	75.0	65.6	66.1	65.9	58.7	57.0	54.9	68.4	67.4	61.5	45.8	45.2	43.6
Qwen3.5-2B	64.3	69.0	60.7	58.2	57.3	58.6	43.2	43.3	44.2	60.9	61.3	53.2	29.5	29.3	29.1
Qwen3-VL-8B-Instruct	78.4	81.8	70.4	70.1	70.9	70.7	55.3	55.9	53.9	78.0	78.5	72.9	34.8	33.3	33.7
Qwen3-VL-4B-Instruct	68.1	72.8	61.5	66.5	66.8	67.4	50.9	50.3	47.3	75.8	73.4	70.9	37.4	37.9	36.3

Table 9: Performance comparison under different luminance contrast conditions across task dimensions. Friedman test results: D1 (

p<0.001

, Cohen’s

f=0.153

), D3 (

p=0.007

, Cohen’s

f=0.056

), and D5 (

p=0.229

, Cohen’s

f=0.03

). Repeated-measures ANOVA results: D2 (

p=0.024

), D4 (

p<0.001

); post-hoc multiple comparisons reveal no significant differences in D2 and D4.

Condition	Sequential	Randomized	$\Delta$
Baseline (clean)	59.1	46.3	-12.8
+ Gaussian noise	58.5	45.8	-12.7
+ Rotation	37.5	31.7	-5.9
+ Resolution reduction	35.8	31.6	-4.2

Table 10: InternVL3.5-8B accuracy (%) under input degradation, compared with the clean-map baseline (Table 5 ).

\Delta

denotes the accuracy drop after disrupting sequential ordering.

Figure 6: Results of the matched $2\times 2$ factorial experiment. Columns indicate whether region attribute values and ordinal classes were supplied as structured text, and rows indicate whether centroid coordinates and adjacency relations were supplied. Each quadrant reports accuracy (%) for three models across the five task dimensions. Overall denotes accuracy averaged across the three models and five dimensions.

Model	Sequential	Randomized	Reversed
InternVL3.5-8B	59.1	46.3	40.7
Qwen3.5-9B	68.6	53.6	48.7
GLM-4.6V-Flash	64.3	51.9	50.2

Table 11: Overall accuracy (%) under the original sequential and reversed (lighter = higher) sequential color conventions, compared with randomized encoding.

Figure 7: Accuracy (%) under original, randomized, and reversed color conventions, by task dimension (D1–D5), for InternVL3.5-8B, Qwen3.5-9B, and GLM-4.6V-Flash.

Figure 8: Accuracy before and after LoRA fine-tuning under different cartographic conditions for QWEN3.5-9B, GLM-4.6V-Flash, and InternVL3.5-8B. (a) Sequential vs. randomized. (b) Standard, high-contrast, and low-contrast luminance conditions.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Model	No-Image	Blank-Image	Shuffled + Blank-Image
InternVL3.5-8B	28.1	27.2	27.1
Qwen3.5-9B	27.8	27.5	27.2
Avg.	27.95	27.35	27.15

Appendix

Table 12: Accuracy (%) under no-image, blank-image, and shuffled-answer control baselines, averaged across all five task dimensions (D1–D5). Weighted chance level

\approx

27.8%.

Figure 9: Per-model, per-subtype accuracy gap between sequential and randomized encoding (percentage points; positive = higher accuracy under sequential encoding). Models are ordered by mean gap (largest at top). Subtype labels follow Table 2 .

Figure 10: Per-model, per-subtype accuracy change under contrast manipulation. Left: Standard minus Low Contrast (degradation under low contrast). Right: High Contrast minus Standard (improvement under high contrast). Models are ordered by mean Standard-minus-Low gap (largest at top).

Figure 11: Example real-world thematic maps from the MapQA-derived source data (Kaiser Family Foundation health and healthcare indicators), rendered at the US state level prior to the controlled synthetic map generation pipeline.

Figure 12: Additional example choropleth maps from CHROMA, spanning five thematic categories (Obesity Population, Software Usage Rate, Birth Population, Unemployment Population, Birth Rate) and multiple color rendering variants used in the benchmark construction pipeline.

	D1		D2		D3		D4		D5
Model	Color	Value	Color	Value	Color	Value	Color	Value	Color	Value
Qwen3.5-9B	84.8	64.0	75.2	76.2	59.7	52.4	76.0	55.4	47.2	38.5
GLM-4.6V-Flash	81.6	54.6	66.1	67.2	55.9	53.8	76.8	62.5	41.1	36.5
InternVL3.5-8B	74.0	49.1	65.1	63.4	51.0	53.0	75.5	61.3	29.8	27.5
Avg.	80.1	55.9	68.8	68.9	55.5	53.1	76.1	59.7	39.4	34.2

Appendix

Table 13: Supplementary comparison of accuracy (%) under the sequential color-encoding condition (Color) and the value-in-region, color-free condition (Value) across the five task dimensions (D1–D5). The value-in-region condition additionally requires numerical-text recognition and region–value association.

Data Visualization

Data visualization studies how visual representations help people explore, interpret, communicate, and reason about data.

See all 121 papers

Do Maps Still Matter for Machines: Revisiting the Role of Choropleth Maps in Foundation Model Spatial Understanding
Jul 20, 2026Zhiwei Wei, Yonghe Sun, Zhenjia Liu +5Data Visualization Spatial Reasoning Benchmarks
aSchool of Geographic Sciences, Hunan Normal University, Changsha, China · bHunan Key Laboratory of Geospatial Big Data Mining and Application, Changsha, China · cSchool of Information and Communication Engineering, Beijing University of Posts and Telecommunications, Beijing, China +3
MindTopo: Can Foundation Models Reason in Topological Space?
Sep 10, 2026Yunfei Ge, Anbang Liu, Qineng Wang +9Spatial Reasoning Benchmarks AI Agent Benchmarks
Northwestern University · Microsoft Research · Stanford University
Human-AI Perceptual Alignment by Playing Hues and Cues
Aug 7, 2026Nuria Alabau-Bosque, Jorge Vila-Tomás, Paula Daudén-Oliver +3VLM Evaluation Vision-Language Grounding
Image Processing Lab, Universitat de València, Carrer del Catedrátic José Beltrán Martinez, Paterna, 46980, Spain

Jul 20, 2026J/K move · Enter open · S save

Do Maps Still Matter for Machines: Revisiting the Role of Choropleth Maps in Foundation Model Spatial Understanding

Zhiwei Wei, Yonghe Sun, Zhenjia Liu, Wenjia Xu+4

aSchool of Geographic Sciences, Hunan Normal University, Changsha, China · bHunan Key Laboratory of Geospatial Big Data Mining and Application, Changsha, China · cSchool of Information and Communication Engineering, Beijing University of Posts and Telecommunications, Beijing, China+3

Data Visualization Spatial Reasoning Benchmarks Visual Spatial Reasoning Geospatial Reasoning Benchmarks Geospatial Foundation Models Geospatial Reasoning+2

Open paper AskPDF arXiv

Spatial understanding is crucial for foundation models (FMs), and maps have long helped humans organize and reason about geographic information. This study examines whether choropleth maps remain useful for machine spatial understanding when models can directly process structured geodata. We introduce ChoroplethMap-Bench, a controlled benchmark containing 2,400 synthetic choropleth maps, corresponding GeoJSON data, and 12,000 questions across five cognitive dimensions: Identify, Spatial Recognition, Compare, Rank, and Delineate. We evaluate 22 open-source and proprietary models under three input conditions: Data Only, Map Only, and Data + Map. The results show that maps substantially improve spatial reasoning, especially when combined with symbolic data and for tasks requiring higher-level understanding of spatial patterns. We further analyze the effects of map type, color hue, and spatial structure, as well as prompting strategies, language, geographic context, decoding settings, classification methods, and response stability. Overall, the Data + Map condition achieves the strongest performance, demonstrating that maps remain valuable external representations for foundation model spatial reasoning.

Abstract

Data Visualization

Explore similar work

Do Maps Still Matter for Machines: Revisiting the Role of Choropleth Maps in Foundation Model Spatial Understanding

MindTopo: Can Foundation Models Reason in Topological Space?

Human-AI Perceptual Alignment by Playing Hues and Cues