ViC-Bench is a dataset for benchmarking Multimodal Large Language Models (MLLMs) on visual reasoning tasks. It was created by meituan-longcat and last updated on May 30, 2025. The dataset likely contains tasks designed to evaluate models using Visual-Interleaved Chain-of-Thought reasoning.
Use Cases
- Benchmarking MLLM reasoning capabilities based on the Visual-Interleaved Chain-of-Thought (VI-CoT) concept.
- Evaluating model performance on tasks requiring step-wise intermediate visual states.
- Comparing model advancements on multimodal reasoning benchmarks.
- Training models to update understanding based on visual information.
- Analyzing the impact of free-style versus fixed intermediate visual states on model performance.
Strengths
- Dataset is focused on a specific and emerging evaluation paradigm (Visual-Interleaved Chain-of-Thought).
- Last update timestamp is recent (2025-05-30 05:19:26).
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is unknown, which may limit suitability assessment.
- Description metadata is limited; actual data quality requires manual inspection after download.
Provenance
- Source
- meituan-longcat