TACIT Benchmark: A Programmatic Visual Reasoning Benchmark for Generative and Discriminative Models

Recent image generation systems increasingly combine multimodal understanding, reasoning, and synthesis, suggesting that they may do more than render plausible scenes. Yet existing evaluations emphasize aesthetics, prompt alignment, compositionality, or text-based answers, leaving unclear whether these systems can solve visual problems and faithfully express solutions in pixels. We introduce GenPuzzle, a benchmark for reasoning-centric image generation. GenPuzzle contains 2,005 problems across 12 tracks, spanning pattern completion, spatial construction, mazes, Sudoku, nonograms, tangrams, board games, matchstick puzzles, orthographic projection, and mathematical visual proof. Each task provides a visual puzzle and requires an image output that preserves the input state while executing a logically valid solution. GenPuzzle uses task-specific evaluation protocols: discrete grid outputs are transcribed and verified programmatically, while visually complex outputs are assessed with tiered, multidimensional, or binary multimodal large language model (MLLM) rubrics. We further select the automatic judge by measuring agreement with human reference scores. Across three frontier generators, the strongest model reaches only 40.57 Macro Overall, revealing frequent failures in logic, geometry, state preservation, and instruction execution. GenPuzzle provides a testbed for measuring progress from image rendering toward visual problem solving.

Reliable visual reasoning requires composing multiple visual observations and returning consistent answers to logically equivalent questions. We introduce Hob-VL, a benchmark for visually grounded Boolean reasoning. Hob-VL comprises two tasks: (1) evaluating whether a Boolean rule holds in an image, and (2) identifying the (unique) object satisfying a Boolean description. Hob-VL contains 6,000 human-verified balanced Yes/No questions, each defined by a Boolean combination of ten visual statements, across 1,000 generated scenes and 46 diverse labeled photographs, along with 1,000 object-identification questions over the same photographs. Our question families are deliberately constructed to challenge reasoning through misleading local cues and nested logical operations, and include symbolic and structured natural-language presentations. Across eight model configurations with thinking disabled or minimized, Boolean accuracy ranges from 48.52% to 50.57%, while the identification accuracy reaches at most 43.0%. A thinking-enabled GLM configuration achieves uneven gains while retaining substantial errors and inconsistencies. Hob-VL exposes these failures through executable reference answers and matched evaluations.

Text-to-Image (T2I) generative models have made significant progress in producing high-fidelity images from natural-language prompts. Alongside this progress, numerous benchmarks have been developed to evaluate capabilities such as attribute binding, numeracy, spatial reasoning, and commonsense reasoning. In this work, we propose a new perspective on T2I evaluation by casting it as a constraint-satisfaction problem. We introduce a logical framework that formally specifies the benchmark domain, including objects, attributes, and relations, alongside the constraints that generated images must satisfy. We evaluated four state-of-the-art T2I models across 9 classes of constraints ranging from a single-object-attribute to multi-object relationships constraints involving logical operators ($\forall, \exists, count, \rightarrow$). Our results reveal a substantial gap between single-object and complex conditional multi-object relationship constraint classes, with the best-performing model achieving 98% accuracy on the former and just 53% on the latter. We assess compositional generalisation, finding that the best-performing model`s accuracy drops from 84% on single-constraint prompts to 20% when satisfying multiple constraints simultaneously. Identifying unsatisfiable constraints is also challenging, with the best model correctly recognizing only 23% of such prompts. Building on this framework, we introduce LogicVal, an evaluation methodology that combines neural visual detection with logical reasoning to assess constraint satisfaction in generated images. On CLEVR, LogicVal achieves ROC-AUC scores ranging from 0.88 to 0.98, indicating strong alignment with human judgments. These findings underscore the value of logic-based evaluation for characterizing T2I model capabilities and limitations.