# Visual Question Answering on VQAv2 3K curated MS COCO (test) Source: [Visual Question Answering on VQAv2 3K curated MS COCO (test)]() Collection index: [llms.txt]() Data updated: 2026-04-21T12:18:06.223Z Evaluation of model performance on the specific 3K curated subset of the VQAv2 dataset, tested on the official test split, focusing on Visual Question Answering capabilities. A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage. | Scope | Recorded value | | --- | --- | | Task | Visual Question Answering | | Dataset (including any stated split) | VQAv2 3K curated MS COCO (test) | | Metric | Relative Performance Drop (%) | | Unit | PERCENT | | Direction | Lower is better | | Results loaded | 10 of 10 | | Coverage | All results currently returned for this benchmark | Evaluation settings may differ between results. The original papers and recorded configurations are needed to establish comparability. A shared metric name alone does not establish an identical protocol. ## Metrics - [Relative Performance Drop (%)](): Lower is better; unit: PERCENT. ## Recorded results | Method | Relative Performance Drop (%) | Result date | Report paper | | --- | --- | --- | --- | | DLA | 7.4 | Not recorded | 69e704d47d0172140ec7efea | | DLA | 9.8 | Not recorded | 69e704d47d0172140ec7efea | | LLM-Knowledge | 16.5 | Not recorded | 69e704d47d0172140ec7efea | | Group Patching | 18 | Not recorded | 69e704d47d0172140ec7efea | | LLM-Knowledge | 19.4 | Not recorded | 69e704d47d0172140ec7efea | | Group Patching | 20.05 | Not recorded | 69e704d47d0172140ec7efea | | QRNCA | 20.8 | Not recorded | 69e704d47d0172140ec7efea | | QRNCA | 24.5 | Not recorded | 69e704d47d0172140ec7efea | | HONES | 27.3 | Not recorded | 69e704d47d0172140ec7efea | | HONES | 36.5 | Not recorded | 69e704d47d0172140ec7efea | ### DLA Relative Performance Drop (%): 7.4. Result ID: 69e76af147f34ce924cef6a2. [Reporting paper]() Recorded evaluation configuration: {"Base Model":"LLaVA-1.5-7B","Masking Ratio":"top-1%"}. This record does not assert independent reproduction. ### DLA Relative Performance Drop (%): 9.8. Result ID: 69e76af147f34ce924cef6a7. [Reporting paper]() Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-7B","Masking Ratio":"top-1%"}. This record does not assert independent reproduction. ### LLM-Knowledge Relative Performance Drop (%): 16.5. Result ID: 69e76af147f34ce924cef6a5. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Base Model":"LLaVA-1.5-7B","Masking Ratio":"top-1%"}. This record does not assert independent reproduction. ### Group Patching Relative Performance Drop (%): 18. Result ID: 69e76af147f34ce924cef6a3. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Base Model":"LLaVA-1.5-7B","Masking Ratio":"top-1%"}. This record does not assert independent reproduction. ### LLM-Knowledge Relative Performance Drop (%): 19.4. Result ID: 69e76af147f34ce924cef6aa. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-7B","Masking Ratio":"top-1%"}. This record does not assert independent reproduction. ### Group Patching Relative Performance Drop (%): 20.05. Result ID: 69e76af147f34ce924cef6a8. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-7B","Masking Ratio":"top-1%"}. This record does not assert independent reproduction. ### QRNCA Relative Performance Drop (%): 20.8. Result ID: 69e76af147f34ce924cef6a4. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Base Model":"LLaVA-1.5-7B","Masking Ratio":"top-1%"}. This record does not assert independent reproduction. ### QRNCA Relative Performance Drop (%): 24.5. Result ID: 69e76af147f34ce924cef6a9. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-7B","Masking Ratio":"top-1%"}. This record does not assert independent reproduction. ### HONES Relative Performance Drop (%): 27.3. Result ID: 69e76af147f34ce924cef6a6. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Base Model":"LLaVA-1.5-7B","Masking Ratio":"top-1%"}. This record does not assert independent reproduction. ### HONES Relative Performance Drop (%): 36.5. Result ID: 69e76af147f34ce924cef6ab. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-7B","Masking Ratio":"top-1%"}. This record does not assert independent reproduction. Page 1 of 1. This is an index page, not the full collection.