tylerxdurden/TACIT-benchmark

TACIT Benchmark v0.1.0

Transformation-Aware Capturing of Implicit Thought

A programmatic visual reasoning benchmark for evaluating generative and discriminative capabilities of multimodal models across 10 tasks and 6 reasoning domains.

Author: Daniel Nobrega Medeiros  |  arXiv paper  |  GitHub

Overview

TACIT presents visual puzzles that require genuine spatial, logical, and structural reasoning — not pattern matching on text. Each puzzle is generated programmatically with deterministic seeding, ensuring full reproducibility. Evaluation is programmatic (no LLM-as-judge): solutions are verified through computer vision algorithms (pixel sampling, SSIM, BFS path detection, color counting).

Key Features

• 6,000 puzzles across 10 tasks and 3 difficulty levels
• Dual-track evaluation: generative (produce a solution image) and discriminative (select from candidates)
• Multi-resolution: every puzzle rendered at 512px, 1024px, and 2048px
• Deterministic: seeded generation (seed=42) for exact reproducibility
• Programmatic verification: CV-based solution checking, no subjective evaluation

Task Examples

All examples below show medium difficulty puzzles at 512px resolution.

01 — Multi-layer Mazes

    

Navigate through multiple maze layers connected by portals (colored dots).

02 — Raven's Progressive Matrices

    

Identify the missing panel in a 3×3 matrix governed by transformation rules.

03 — Cellular Automata Forward Prediction

    

Given a rule and initial state, predict the next state of a cellular automaton.

04 — Cellular Automata Inverse Inference

    

Given an initial and final state, identify which rule was applied.

05 — Visual Logic Grids

    

Complete a constraint-satisfaction grid using visual clues.

06 — Planar Graph k-Coloring

    

Color graph nodes so no adjacent nodes share the same color.

07 — Graph Isomorphism Detection

    

Determine whether two graphs have the same structure despite different layouts.

08 — Unknot Detection

    

Determine whether a knot diagram can be untangled into a simple loop.

09 — Orthographic Projection Identification

    

Match a 3D object to its correct orthographic projection views.

10 — Isometric Reconstruction

    

Reconstruct a 3D isometric view from orthographic projections.

Tasks

#	Task	Domain	Easy	Medium	Hard
01	Multi-layer Mazes	Spatial Reasoning	8×8, 1 layer	16×16, 2 layers, 2 portals	32×32, 3 layers, 5 portals
02	Raven's Progressive Matrices	Abstract Reasoning	1 rule	2 rules	3 rules, compositional
03	Cellular Automata Forward	Causal Reasoning	8×8, 1 step	16×16, 3 steps	32×32, 5 steps
04	Cellular Automata Inverse	Causal Reasoning	8×8, 4 rules	16×16, 8 rules	32×32, 16 rules
05	Visual Logic Grids	Logical Reasoning	4×4, 6 constraints	5×5, 10 constraints	6×6, 16 constraints
06	Planar Graph k-Coloring	Graph Theory	6 nodes, k=4	12 nodes, k=4	20 nodes, k=3
07	Graph Isomorphism	Graph Theory	5 nodes	8 nodes	12 nodes
08	Unknot Detection	Topology	3 crossings	6 crossings	10 crossings
09	Orthographic Projection	Spatial Reasoning	6 faces	10 faces, 1 concavity	16 faces, 3 concavities
10	Isometric Reconstruction	Spatial Reasoning	6 faces	10 faces, 1 ambiguity	16 faces, 2 ambiguities

Each task has 200 puzzles per difficulty level (easy / medium / hard) = 600 per task, 6,000 total.

Evaluation Tracks

Track 1 — Generative

The model receives a puzzle image and must produce a solution image (e.g., a solved maze, colored graph, completed matrix). Verification is fully programmatic using computer vision:

Task	Verification Method
Maze	BFS path detection on rendered solution
Raven	SSIM comparison (threshold 0.997)
CA Forward / Inverse	Pixel sampling of cell states
Logic Grid	Pixel sampling of grid cells
Graph Coloring	Occlusion-aware node color sampling
Graph Isomorphism	Color counting + structural validation
Unknot	Color region counting
Ortho Projection	Pixel sampling of projection views
Iso Reconstruction	SSIM comparison (threshold 0.99999)

Track 2 — Discriminative

The model receives a puzzle image plus 4 distractor images and 1 correct solution, and must identify the correct answer. This is a 5-way multiple-choice visual task.

Dataset Structure

snapshot/
├── metadata.json              # Generation config and parameters
├── README.md                  # This file
├── task_01_maze/
│   ├── task_info.json         # Task parameters
│   ├── easy/
│   │   ├── 512/               # 512px resolution
│   │   │   ├── puzzle_0000.png
│   │   │   ├── solution_0000.png
│   │   │   ├── distractors_0000/
│   │   │   │   ├── distractor_00.png
│   │   │   │   ├── distractor_01.png
│   │   │   │   ├── distractor_02.png
│   │   │   │   └── distractor_03.png
│   │   │   ├── puzzle_0001.png
│   │   │   ├── solution_0001.png
│   │   │   ├── distractors_0001/
│   │   │   │   └── ...
│   │   │   └── ... (200 puzzles)
│   │   ├── 1024/              # 1024px resolution
│   │   │   └── ... (same structure)
│   │   └── 2048/              # 2048px resolution
│   │       └── ... (same structure)
│   ├── medium/
│   │   └── ... (same structure)
│   └── hard/
│       └── ... (same structure)
├── task_02_raven/
│   └── ...
└── ... (10 tasks total)

File Naming Convention

• puzzle_NNNN.png — the input puzzle image
• solution_NNNN.png — the ground-truth solution (Track 1 target)
• distractors_NNNN/distractor_0X.png — 4 wrong answers (Track 2 candidates)

Statistics

Metric	Value
Total puzzles	6,000
Total PNG files	108,008
Resolutions	512, 1024, 2048 px
Difficulties	easy, medium, hard
Distractors per puzzle	4
Dataset size	~3.9 GB
Generation seed	42

Usage

Loading with Hugging Face

from datasets import load_dataset

# Load full dataset
ds = load_dataset("tylerxdurden/TACIT-benchmark")

# Or download specific files
from huggingface_hub import hf_hub_download

puzzle = hf_hub_download(
    repo_id="tylerxdurden/TACIT-benchmark",
    filename="task_01_maze/easy/1024/puzzle_0000.png",
    repo_type="dataset",
)

Using the Evaluation Harness

from tacit.registry import GENERATORS

# Regenerate a specific puzzle (deterministic)
gen = GENERATORS["maze"]
puzzle = gen.generate(seed=42, difficulty="easy", index=0)

# Verify a candidate solution (Track 1)
is_correct = gen.verify(puzzle, candidate_png=model_output_bytes)

See the GitHub repository for full evaluation documentation.

Reasoning Domains

The 10 tasks span 6 reasoning domains, chosen to probe different aspects of visual cognition:

1. Spatial Reasoning — Mazes, orthographic projection, isometric reconstruction
2. Abstract Reasoning — Raven's progressive matrices
3. Causal Reasoning — Cellular automata (forward prediction and inverse inference)
4. Logical Reasoning — Visual logic grids
5. Graph Theory — Graph coloring, graph isomorphism
6. Topology — Unknot detection

Citation

@misc{medeiros_2026,
    author       = {Daniel Nobrega Medeiros},
    title        = {TACIT-benchmark},
    year         = 2026,
    url          = {https://huggingface.co/datasets/tylerxdurden/TACIT-benchmark},
    doi          = {10.57967/hf/7904},
    publisher    = {Hugging Face}
}

License

Apache 2.0