Continual-NExT is a benchmark built upon a collection of widely used and publicly available multimodal datasets for both understanding and generation tasks. The benchmark is adopted to evaluate the multimodal continual learning ability of unified generation and understanding MLLMs. It was created by jingyang and last updated on February 19, 2026.
Use Cases
- Benchmarking model performance on multimodal continual learning based on the described collection of datasets.
- Evaluating generalization and forgetting in MLLMs across diverse vision-language tasks.
- Comparing unified model architectures on a standardized suite of multimodal tasks.
Strengths
- Benchmark is built upon widely used and publicly available multimodal datasets, including VQAv2, ImageNet, Flickr30k, OCR-VQA, RefCOCO, and HQEdit.
- Specifically designed to evaluate multimodal continual learning ability, a key challenge for MLLMs.
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
- jingyang
- Collection Method
- Aggregated from multiple publicly available multimodal datasets.
- Freshness
- Last updated 2026-02-19 16:07:04; freshness should be verified.