VisualPuzzles: Decoupling Multimodal Reasoning Evaluation from Domain Knowledge

Yueqi Song, Tianyue Ou, Yibo Kong, Zecheng Li, Graham Neubig, Xiang Yue
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:114860-114932, 2026.

Abstract

Current multimodal benchmarks often conflate reasoning with domain knowledge, making it difficult to isolate and evaluate reasoning abilities in non-expert settings. To address this, we introduce VisualPuzzles, a benchmark that targets visual reasoning while deliberately minimizing reliance on specialized knowledge. VisualPuzzles consists of diverse puzzles spanning 5 categories: algorithmic, analogical, deductive, inductive, and spatial reasoning. One major source of our questions is manually translated logical reasoning questions from the Chinese Civil Service Examination. Experiments show that VisualPuzzles requires significantly less intensive domain-specific knowledge and more complex reasoning compared to benchmarks like MMMU, enabling us to better evaluate genuine multimodal reasoning. Evaluations show that state-of-the-art multimodal large language models consistently lag behind human performance on VisualPuzzles, and that strong performance on knowledge-intensive benchmarks does not necessarily translate to success on reasoning-focused, knowledge-light tasks. Additionally, reasoning enhancements such as scaling up inference compute (with "thinking" modes) yield inconsistent gains across models and task types, and we observe no clear correlation between model size and performance. We also found that models exhibit different reasoning and answering patterns on VisualPuzzles compared to benchmarks with heavier emphasis on knowledge. VisualPuzzles offers a clearer lens through which to evaluate reasoning capabilities beyond factual recall and domain knowledge.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-song26s, title = {{V}isual{P}uzzles: Decoupling Multimodal Reasoning Evaluation from Domain Knowledge}, author = {Song, Yueqi and Ou, Tianyue and Kong, Yibo and Li, Zecheng and Neubig, Graham and Yue, Xiang}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {114860--114932}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/song26s/song26s.pdf}, url = {https://proceedings.mlr.press/v306/song26s.html}, abstract = {Current multimodal benchmarks often conflate reasoning with domain knowledge, making it difficult to isolate and evaluate reasoning abilities in non-expert settings. To address this, we introduce VisualPuzzles, a benchmark that targets visual reasoning while deliberately minimizing reliance on specialized knowledge. VisualPuzzles consists of diverse puzzles spanning 5 categories: algorithmic, analogical, deductive, inductive, and spatial reasoning. One major source of our questions is manually translated logical reasoning questions from the Chinese Civil Service Examination. Experiments show that VisualPuzzles requires significantly less intensive domain-specific knowledge and more complex reasoning compared to benchmarks like MMMU, enabling us to better evaluate genuine multimodal reasoning. Evaluations show that state-of-the-art multimodal large language models consistently lag behind human performance on VisualPuzzles, and that strong performance on knowledge-intensive benchmarks does not necessarily translate to success on reasoning-focused, knowledge-light tasks. Additionally, reasoning enhancements such as scaling up inference compute (with "thinking" modes) yield inconsistent gains across models and task types, and we observe no clear correlation between model size and performance. We also found that models exhibit different reasoning and answering patterns on VisualPuzzles compared to benchmarks with heavier emphasis on knowledge. VisualPuzzles offers a clearer lens through which to evaluate reasoning capabilities beyond factual recall and domain knowledge.} }
Endnote
%0 Conference Paper %T VisualPuzzles: Decoupling Multimodal Reasoning Evaluation from Domain Knowledge %A Yueqi Song %A Tianyue Ou %A Yibo Kong %A Zecheng Li %A Graham Neubig %A Xiang Yue %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-song26s %I PMLR %P 114860--114932 %U https://proceedings.mlr.press/v306/song26s.html %V 306 %X Current multimodal benchmarks often conflate reasoning with domain knowledge, making it difficult to isolate and evaluate reasoning abilities in non-expert settings. To address this, we introduce VisualPuzzles, a benchmark that targets visual reasoning while deliberately minimizing reliance on specialized knowledge. VisualPuzzles consists of diverse puzzles spanning 5 categories: algorithmic, analogical, deductive, inductive, and spatial reasoning. One major source of our questions is manually translated logical reasoning questions from the Chinese Civil Service Examination. Experiments show that VisualPuzzles requires significantly less intensive domain-specific knowledge and more complex reasoning compared to benchmarks like MMMU, enabling us to better evaluate genuine multimodal reasoning. Evaluations show that state-of-the-art multimodal large language models consistently lag behind human performance on VisualPuzzles, and that strong performance on knowledge-intensive benchmarks does not necessarily translate to success on reasoning-focused, knowledge-light tasks. Additionally, reasoning enhancements such as scaling up inference compute (with "thinking" modes) yield inconsistent gains across models and task types, and we observe no clear correlation between model size and performance. We also found that models exhibit different reasoning and answering patterns on VisualPuzzles compared to benchmarks with heavier emphasis on knowledge. VisualPuzzles offers a clearer lens through which to evaluate reasoning capabilities beyond factual recall and domain knowledge.
APA
Song, Y., Ou, T., Kong, Y., Li, Z., Neubig, G. & Yue, X.. (2026). VisualPuzzles: Decoupling Multimodal Reasoning Evaluation from Domain Knowledge. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:114860-114932 Available from https://proceedings.mlr.press/v306/song26s.html.

Related Material