VisuLogic: A Benchmark for Evaluating Visual Reasoning in Multi-modal Large Language Models

The paper introduces VisuLogic, a benchmark of 1,000 human-verified problems designed to evaluate visual reasoning in multimodal large language models. It resides in the General Reasoning and Comprehension Benchmarks leaf, which contains five papers including the original work. This leaf sits within the broader Evaluation and Benchmarking branch, indicating a moderately populated research direction focused on standardized assessment of MLLM capabilities. The benchmark targets a specific gap: measuring genuine vision-centric reasoning without language-based shortcuts, positioning it alongside other evaluation frameworks that probe diverse reasoning skills.

The taxonomy reveals neighboring leaves focused on Visual Perception and Pattern Recognition Benchmarks and Domain-Specific Benchmarks, suggesting the field has organized evaluation efforts along task-type boundaries. Sibling papers in the same leaf include SEED-Bench and MDK12-Bench, which offer broad multimodal understanding and domain-specific mathematical assessments respectively. VisuLogic differentiates itself by emphasizing structured logical inference across six reasoning categories, whereas neighbors tend toward either comprehensive coverage or specialized domains. This positioning reflects an ongoing field tension between general-purpose and capability-specific evaluation design.

Among 30 candidates examined across three contributions, none were identified as clearly refuting the work. The VisuLogic benchmark contribution examined 10 candidates with zero refutable overlaps, as did the data curation pipeline and reasoning taxonomy contributions. This suggests that within the limited search scope, no prior work directly anticipates the specific combination of human-verified visual reasoning problems organized into six categories with explicit controls against language shortcuts. The absence of refutable candidates across all contributions indicates potential novelty, though the search scale means undiscovered prior work remains possible.

Based on the limited literature search of 30 semantically similar papers, the work appears to occupy a distinct position within the evaluation landscape. The taxonomy structure shows a moderately populated leaf with clear boundaries separating general reasoning benchmarks from perception-focused and domain-specific alternatives. However, the analysis cannot rule out relevant prior work outside the top-30 semantic matches or in adjacent research communities not captured by the taxonomy.