GeoSym-Bench

A controlled evaluation benchmark for quantitative chart reasoning with fine-grained error attribution.

GeoSym-Bench is the companion benchmark for GeoSym: Structured Evidence Fusion for Reliable Quantitative Chart Reasoning.

Overview

GeoSym-Bench comprises 460 clean-rendered mathematical function graph images across 8 function families, each with complete algebraic ground-truth annotations spanning 10 evaluation dimensions. The benchmark enables precise attribution of reasoning errors to specific pipeline stages — a capability not supported by existing chart reasoning benchmarks.

Property	Value
Total samples	460
Image resolution	1440 x 1440 px
Function families	8
Evaluation dimensions	10 (9 objective + 1 judge-scored)
Ground truth	Exact algebraic annotations
Language	English (questions), bilingual (EN/ZH)

Function Families

Category	Family	General Form	Samples
Elementary	Linear	y = ax + b	59
Elementary	Quadratic	y = a(x-h)^2 + k	58
Elementary	Absolute value	y = a|x-h| + k	58
Polynomial	Cubic	y = a(x-h)^3 + b(x-h) + k	58
Piecewise	Piecewise linear	y = s_i(x - h_i) + k_i	59
Transcendental	Exponential	y = a * b^(x-h) + k	57
Transcendental	Logarithmic	y = a * log_2(x-h) + k	51
Special	Rational	y = a/(x-h) + k	60

Evaluation Dimensions

Each sample is scored across 10 equally weighted dimensions (D1-D10):

Dimension	Target	Type
D1: Intercepts	x- and y-intercepts	Objective
D2: Extrema	Local minima/maxima with type labels	Objective
D3: Predictions	Function values at specified x-values	Objective
D4: Rate of change	Average slope over a defined interval	Objective
D5: Monotonicity	Increasing/decreasing intervals	Objective
D6: Symmetry	Even, odd, or neither	Objective
D7: Expression	Closed-form functional representation	Objective
D8: Domain	Mathematical domain	Objective
D9: Range	Mathematical range	Objective
D10: Inference quality	Factual accuracy, coherence, evidence support	Judge-scored

Dataset Structure

GeoSym-Bench/
├── images/
│   ├── absolute-clean_render/    (58 PNG files)
│   ├── cubic-clean_render/       (58 PNG files)
│   ├── exponential-clean_render/ (57 PNG files)
│   ├── linear-clean_render/      (59 PNG files)
│   ├── logarithmic-clean_render/ (51 PNG files)
│   ├── piecewise_linear-clean_render/ (59 PNG files)
│   ├── quadratic-clean_render/   (58 PNG files)
│   └── rational-clean_render/    (60 PNG files)
└── ground_truth/
    └── (same subfolder structure, 460 JSON files)

Each JSON ground-truth file contains:

• chart_bbox, page_size, layout — rendering metadata
• x_range, y_range, x_tick_values, y_tick_values — axis specifications
• intersections — exact curve-grid intersection coordinates
• functions — parameters, expression, function type
• question.ground_truth — all 10 evaluation dimensions
• benchmark_meta — function family, noise profile, seed

Usage

# Direct download
from huggingface_hub import snapshot_download
path = snapshot_download(repo_id="lihaonan0716/GeoSym-Bench", repo_type="dataset")

Source Code

The GeoSym framework, evaluation suite, and prompt templates are available at: https://github.com/lihaonan0716/GeoSym

License

This dataset is released under the Creative Commons Attribution 4.0 International License (CC-BY 4.0).

Citation

@misc{wang2026geosym,
  title     = {GeoSym: Structured Evidence Fusion for Reliable Quantitative Chart Reasoning},
  author    = {Wang, Yongqing and Zhang, Qisheng and Guo, Linyan and Li, Haonan and Sun, Tianjun and Yu, Samson S.},
  year      = {2026},
  note      = {Manuscript in preparation. Software and dataset available online},
  url       = {https://github.com/lihaonan0716/GeoSym}
}