# Scientific Question Answering on ScienceQA image

Source: [Scientific Question Answering on ScienceQA image](<https://www.sota2.com/research/sota/scientific-question-answering-on-scienceqa-image>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-07-09T07:11:31.921Z

Evaluation of model performance on the ScienceQA image subset for the Scientific Question Answering task, measured by accuracy.

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 | Scientific Question Answering |
| Dataset (including any stated split) | ScienceQA image |
| Metric | Accuracy |
| Unit | PERCENT |
| Direction | Higher is better |
| Results loaded | 281 of 281 |
| 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

- [Accuracy](<https://www.sota2.com/research/sota/scientific-question-answering-on-scienceqa-image.md?metric=Accuracy>): Higher is better; unit: PERCENT.

## Recorded results

| Method | Accuracy | Result date | Report paper |
| --- | --- | --- | --- |
| InternVL2.5-8B | 98 | Not recorded | 69d5e0fd7d0172140ec759dc |
| FastV | 95.3 | Not recorded | 69d5e0fd7d0172140ec759dc |
| ID-Selection | 94.4 | Not recorded | 69d5e0fd7d0172140ec759dc |
| DART | 93.8 | Not recorded | 69d5e0fd7d0172140ec759dc |
| ID-Selection | 92.7 | Not recorded | 69d5e0fd7d0172140ec759dc |
| DART | 89.4 | Not recorded | 69d5e0fd7d0172140ec759dc |
| FastV | 88.8 | Not recorded | 69d5e0fd7d0172140ec759dc |
| Qwen2.5-VL-7B | 87.4 | Not recorded | 68fa7a6e1e209131cea832ae |
| HiPrune | 87 | Not recorded | 68fa7a6e1e209131cea832ae |
| HiPrune++ | 86.6 | Not recorded | 68fa7a6e1e209131cea832ae |
| HiPrune | 85.7 | Not recorded | 68fa7a6e1e209131cea832ae |
| HiPrune++ | 85.2 | Not recorded | 68fa7a6e1e209131cea832ae |
| HiPrune | 82.8 | Not recorded | 68fa7a6e1e209131cea832ae |
| HiPrune++ | 82.8 | Not recorded | 68fa7a6e1e209131cea832ae |
| S-MFT | 80.4 | Not recorded | 69535c7253fdd5c36173d7fa |
| Qwen2.5-VL-3B | 80.4 | Not recorded | 68fa7a6e1e209131cea832ae |
| FFT | 80.1 | Not recorded | 69535c7253fdd5c36173d7fa |
| PRISM-Qwen-7B (Instruct) | 80.1 | Not recorded | 68fa7c1c1e209131cea8c702 |
| HiPrune | 80 | Not recorded | 68fa7a6e1e209131cea832ae |
| HiPrune | 79.8 | Not recorded | 68fa7a6e1e209131cea832ae |
| HiPrune++ | 79.8 | Not recorded | 68fa7a6e1e209131cea832ae |
| HiPrune++ | 79.7 | Not recorded | 68fa7a6e1e209131cea832ae |
| HiPrune | 79.4 | Not recorded | 68fa7a6e1e209131cea832ae |
| LoRA | 79.1 | Not recorded | 69535c7253fdd5c36173d7fa |
| HiPrune++ | 79.1 | Not recorded | 68fa7a6e1e209131cea832ae |
| PRISM-Qwen-7B (Base) | 78.9 | Not recorded | 68fa7c1c1e209131cea8c702 |
| TwigVLM | 78.1 | Not recorded | 68fa7a1c1e209131cea81398 |
| TwigVLM | 77.9 | Not recorded | 68fa7a1c1e209131cea81398 |
| LLaVA-Qwen-7B (Instruct) | 77.8 | Not recorded | 68fa7c1c1e209131cea8c702 |
| PRISM-Llama-8B | 77.3 | Not recorded | 68fa7c1c1e209131cea8c702 |
| ID-Selection | 77.1 | Not recorded | 69d5e0fd7d0172140ec759dc |
| VisionZip | 77.1 | Not recorded | 68fa7a1c1e209131cea81398 |
| TwigVLM++ | 77.1 | Not recorded | 68fa7a1c1e209131cea81398 |
| TwigVLM++ | 77 | Not recorded | 68fa7a1c1e209131cea81398 |
| VisionZip | 77 | Not recorded | 68fa7a1c1e209131cea81398 |
| LLaVA-Qwen-7B (Base) | 76.7 | Not recorded | 68fa7c1c1e209131cea8c702 |
| Qwen2.5-VL-7B | 76.6 | Not recorded | 69d5e0fd7d0172140ec759dc |
| VisionZip | 76.1 | Not recorded | 68fa7a1c1e209131cea81398 |
| TwigVLM++ | 76.1 | Not recorded | 68fa7a1c1e209131cea81398 |
| ID-Selection | 75.5 | Not recorded | 69d5e0fd7d0172140ec759dc |
| FastV | 75.5 | Not recorded | 68fa7a1c1e209131cea81398 |
| Q2.5VL-7B | 75.3 | Not recorded | 68fa7a1c1e209131cea81398 |
| FastV | 75.3 | Not recorded | 68fa7a1c1e209131cea81398 |
| DART | 75.2 | Not recorded | 69d5e0fd7d0172140ec759dc |
| LLaVA-Llama-8B | 75.2 | Not recorded | 68fa7c1c1e209131cea8c702 |
| TwigVLM | 74.9 | Not recorded | 68fa7a1c1e209131cea81398 |
| FastV | 74 | Not recorded | 69d5e0fd7d0172140ec759dc |
| FastV | 72.8 | Not recorded | 68fa7a1c1e209131cea81398 |
| PRISM-3B | 72.8 | Not recorded | 68fa7c1c1e209131cea8c702 |
| LLaVA-Phi2-3B | 72.7 | Not recorded | 68fa7c1c1e209131cea8c702 |
| DART | 72.6 | Not recorded | 69d5e0fd7d0172140ec759dc |
| S-MFT Attn | 72.4 | Not recorded | 69535c7253fdd5c36173d7fa |
| S-MFT Both | 72.4 | Not recorded | 69535c7253fdd5c36173d7fa |
| M3 | 72.1 | Not recorded | 69b39a110636e5bf4afbcade |
| FastV | 72 | Not recorded | 69d5e0fd7d0172140ec759dc |
| FFT | 71.9 | Not recorded | 69535c7253fdd5c36173d7fa |
| PRISM-13B | 71.8 | Not recorded | 68fa7c1c1e209131cea8c702 |
| S-MFT MLP | 71.7 | Not recorded | 69535c7253fdd5c36173d7fa |
| LLaVA-Vicuna-13B | 71.6 | Not recorded | 68fa7c1c1e209131cea8c702 |
| LoRA | 71.1 | Not recorded | 69535c7253fdd5c36173d7fa |
| ICONS | 70.8 | Not recorded | 69e704be7d0172140ec7eeac |
| FMVR-LLaVA | 70.6 | Not recorded | 69b39a110636e5bf4afbcade |
| Uni-LoRA | 70.3 | Not recorded | 69535c7253fdd5c36173d7fa |
| FMVR-LLaVA | 70.3 | Not recorded | 69b39a110636e5bf4afbcade |
| PyramidDrop | 70.2 | Not recorded | 69b39a110636e5bf4afbcade |
| LLaVA-NeXT-7B | 70.2 | Not recorded | 68fa7a1c1e209131cea81398 |
| TwigVLM++ | 70.2 | Not recorded | 68fa7a1c1e209131cea81398 |
| LLaVA-NeXT-7B | 70.2 | Not recorded | 68fa7a1c1e209131cea81398 |
| TwigVLM++ | 70.2 | Not recorded | 68fa7a1c1e209131cea81398 |
| PDrop | 70.1 | Not recorded | 69a66c182d603e6de3e3ce31 |
| FastV | 70.1 | Not recorded | 69a66c182d603e6de3e3ce31 |
| LLaVA-NeXT-7B | 70.1 | Not recorded | 69ddca487d0172140ec79c86 |
| FMVR-LLaVA | 69.9 | Not recorded | 69b39a110636e5bf4afbcade |
| TwigVLM | 69.9 | Not recorded | 68fa7a1c1e209131cea81398 |
| TwigVLM | 69.9 | Not recorded | 68fa7a1c1e209131cea81398 |
| FastV | 69.9 | Not recorded | 6a1e63e3dbd7150506f85529 |
| FMVR-LLaVA | 69.7 | Not recorded | 69b39a110636e5bf4afbcade |
| SparseVLM | 69.7 | Not recorded | 6a1e63e3dbd7150506f85529 |
| LoRA-One | 69.6 | Not recorded | 69535c7253fdd5c36173d7fa |
| HoloV + RESTORE | 69.6 | Not recorded | 6a1e63e3dbd7150506f85529 |

### InternVL2.5-8B

Accuracy: 98. Result ID: 69d614a747f34ce924beb2be.

[Reporting paper](<https://www.sota2.com/research/paper/69d5e0fd7d0172140ec759dc.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7bab1e209131cea89c7f.md>)

[Implementation](<https://github.com/OpenGVLab/InternVL>)

Recorded evaluation configuration: {"Base Model":"InternVL2.5-8B","Token Reduction Ratio":"0%"}. This record does not assert independent reproduction.

### FastV

Accuracy: 95.3. Result ID: 69d614a747f34ce924beb2bf.

[Reporting paper](<https://www.sota2.com/research/paper/69d5e0fd7d0172140ec759dc.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa74ae1e209131cea632b7.md>)

Recorded evaluation configuration: {"Base Model":"InternVL2.5-8B","Token Reduction Ratio":"77.8%"}. This record does not assert independent reproduction.

### ID-Selection

Accuracy: 94.4. Result ID: 69d614a747f34ce924beb2c1.

[Reporting paper](<https://www.sota2.com/research/paper/69d5e0fd7d0172140ec759dc.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69d5e0fd7d0172140ec759dc.md>)

Recorded evaluation configuration: {"Base Model":"InternVL2.5-8B","Token Reduction Ratio":"77.8%","Importance Source":"LLM second-layer cross-modal attention"}. This record does not assert independent reproduction.

### DART

Accuracy: 93.8. Result ID: 69d614a747f34ce924beb2c0.

[Reporting paper](<https://www.sota2.com/research/paper/69d5e0fd7d0172140ec759dc.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa79461e209131cea7cfd7.md>)

Recorded evaluation configuration: {"Base Model":"InternVL2.5-8B","Token Reduction Ratio":"77.8%"}. This record does not assert independent reproduction.

### ID-Selection

Accuracy: 92.7. Result ID: 69d614a747f34ce924beb2c4.

[Reporting paper](<https://www.sota2.com/research/paper/69d5e0fd7d0172140ec759dc.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69d5e0fd7d0172140ec759dc.md>)

Recorded evaluation configuration: {"Base Model":"InternVL2.5-8B","Token Reduction Ratio":"88.9%","Importance Source":"LLM second-layer cross-modal attention"}. This record does not assert independent reproduction.

### DART

Accuracy: 89.4. Result ID: 69d614a747f34ce924beb2c3.

[Reporting paper](<https://www.sota2.com/research/paper/69d5e0fd7d0172140ec759dc.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa79461e209131cea7cfd7.md>)

Recorded evaluation configuration: {"Base Model":"InternVL2.5-8B","Token Reduction Ratio":"88.9%"}. This record does not assert independent reproduction.

### FastV

Accuracy: 88.8. Result ID: 69d614a747f34ce924beb2c2.

[Reporting paper](<https://www.sota2.com/research/paper/69d5e0fd7d0172140ec759dc.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa74ae1e209131cea632b7.md>)

Recorded evaluation configuration: {"Base Model":"InternVL2.5-8B","Token Reduction Ratio":"88.9%"}. This record does not assert independent reproduction.

### Qwen2.5-VL-7B

Accuracy: 87.4. Result ID: 69e757e847f34ce924ce6fab.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa773a1e209131cea71d3c.md>)

Recorded evaluation configuration: {"Token Budget":"100%"}. This record does not assert independent reproduction.

### HiPrune

Accuracy: 87. Result ID: 69e757e847f34ce924ce6fac.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-7B","Token Budget":"33.3%"}. This record does not assert independent reproduction.

### HiPrune++

Accuracy: 86.6. Result ID: 69e757e847f34ce924ce6faf.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-7B","Token Budget":"33.3%"}. This record does not assert independent reproduction.

### HiPrune

Accuracy: 85.7. Result ID: 69e757e847f34ce924ce6fad.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-7B","Token Budget":"22.2%"}. This record does not assert independent reproduction.

### HiPrune++

Accuracy: 85.2. Result ID: 69e757e847f34ce924ce6fb0.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-7B","Token Budget":"22.2%"}. This record does not assert independent reproduction.

### HiPrune

Accuracy: 82.8. Result ID: 69e757e847f34ce924ce6fae.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-7B","Token Budget":"11.1%"}. This record does not assert independent reproduction.

### HiPrune++

Accuracy: 82.8. Result ID: 69e757e847f34ce924ce6fb1.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-7B","Token Budget":"11.1%"}. This record does not assert independent reproduction.

### S-MFT

Accuracy: 80.4. Result ID: 6999366d5698665ec49fbc4a.

[Reporting paper](<https://www.sota2.com/research/paper/69535c7253fdd5c36173d7fa.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69535c7253fdd5c36173d7fa.md>)

Recorded evaluation configuration: {"Language Model Backbone":"Qwen2.5-14B","Vision Encoder":"SigLIP-ViT-Large"}. This record does not assert independent reproduction.

### Qwen2.5-VL-3B

Accuracy: 80.4. Result ID: 69e757e847f34ce924ce6fa4.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa773a1e209131cea71d3c.md>)

Recorded evaluation configuration: {"Token Budget":"100%"}. This record does not assert independent reproduction.

### FFT

Accuracy: 80.1. Result ID: 6999366d5698665ec49fbc48.

[Reporting paper](<https://www.sota2.com/research/paper/69535c7253fdd5c36173d7fa.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa72421e209131cea5554e.md>)

[Implementation](<https://github.com/DLCV-BUAA/TinyLLaVA-Bench>)

Recorded evaluation configuration: {"Language Model Backbone":"Qwen2.5-14B","Vision Encoder":"SigLIP-ViT-Large"}. This record does not assert independent reproduction.

### PRISM-Qwen-7B (Instruct)

Accuracy: 80.1. Result ID: 6a1d4a7ffbc373592c16aae2.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7c1c1e209131cea8c702.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7c1c1e209131cea8c702.md>)

Recorded evaluation configuration: {"Model":"PRISM-Qwen-7B (Instruct)"}. This record does not assert independent reproduction.

### HiPrune

Accuracy: 80. Result ID: 69e757e847f34ce924ce6fa6.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-3B","Token Budget":"22.2%"}. This record does not assert independent reproduction.

### HiPrune

Accuracy: 79.8. Result ID: 69e757e847f34ce924ce6fa5.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-3B","Token Budget":"33.3%"}. This record does not assert independent reproduction.

### HiPrune++

Accuracy: 79.8. Result ID: 69e757e847f34ce924ce6fa9.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-3B","Token Budget":"22.2%"}. This record does not assert independent reproduction.

### HiPrune++

Accuracy: 79.7. Result ID: 69e757e847f34ce924ce6fa8.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-3B","Token Budget":"33.3%"}. This record does not assert independent reproduction.

### HiPrune

Accuracy: 79.4. Result ID: 69e757e847f34ce924ce6fa7.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-3B","Token Budget":"11.1%"}. This record does not assert independent reproduction.

### LoRA

Accuracy: 79.1. Result ID: 6999366d5698665ec49fbc49.

[Reporting paper](<https://www.sota2.com/research/paper/69535c7253fdd5c36173d7fa.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6b121e209131cea2bf59.md>)

[Implementation](<https://github.com/microsoft/LoRA>)

Recorded evaluation configuration: {"Language Model Backbone":"Qwen2.5-14B","Vision Encoder":"SigLIP-ViT-Large"}. This record does not assert independent reproduction.

### HiPrune++

Accuracy: 79.1. Result ID: 69e757e847f34ce924ce6faa.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7a6e1e209131cea832ae.md>)

Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-3B","Token Budget":"11.1%"}. This record does not assert independent reproduction.

### PRISM-Qwen-7B (Base)

Accuracy: 78.9. Result ID: 6a1d4a7ffbc373592c16aae0.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7c1c1e209131cea8c702.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7c1c1e209131cea8c702.md>)

Recorded evaluation configuration: {"Model":"PRISM-Qwen-7B (Base)"}. This record does not assert independent reproduction.

### TwigVLM

Accuracy: 78.1. Result ID: 69dc9a8e47f34ce924c338c1.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

Recorded evaluation configuration: {"Tokens Retained Percentage":"22.2%"}. This record does not assert independent reproduction.

### TwigVLM

Accuracy: 77.9. Result ID: 69dc9a8e47f34ce924c338bd.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

Recorded evaluation configuration: {"Tokens Retained Percentage":"33.3%"}. This record does not assert independent reproduction.

### LLaVA-Qwen-7B (Instruct)

Accuracy: 77.8. Result ID: 6a1d4a7ffbc373592c16aae1.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7c1c1e209131cea8c702.md>)

Recorded evaluation configuration: {"Model":"LLaVA-Qwen-7B (Instruct)"}. This record does not assert independent reproduction.

### PRISM-Llama-8B

Accuracy: 77.3. Result ID: 6a1d4a7ffbc373592c16aae4.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7c1c1e209131cea8c702.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7c1c1e209131cea8c702.md>)

Recorded evaluation configuration: {"Model":"PRISM-Llama-8B"}. This record does not assert independent reproduction.

### ID-Selection

Accuracy: 77.1. Result ID: 69d614a747f34ce924beb2ba.

[Reporting paper](<https://www.sota2.com/research/paper/69d5e0fd7d0172140ec759dc.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69d5e0fd7d0172140ec759dc.md>)

Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-7B","Token Reduction Ratio":"77.8%","Importance Source":"LLM second-layer cross-modal attention"}. This record does not assert independent reproduction.

### VisionZip

Accuracy: 77.1. Result ID: 69dc9a8e47f34ce924c338bc.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa760f1e209131cea6b33b.md>)

[Implementation](<https://github.com/dvlab-research/VisionZip>)

Recorded evaluation configuration: {"Tokens Retained Percentage":"33.3%"}. This record does not assert independent reproduction.

### TwigVLM++

Accuracy: 77.1. Result ID: 69dc9a8e47f34ce924c338c2.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

Recorded evaluation configuration: {"Tokens Retained Percentage":"22.2%"}. This record does not assert independent reproduction.

### TwigVLM++

Accuracy: 77. Result ID: 69dc9a8e47f34ce924c338be.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

Recorded evaluation configuration: {"Tokens Retained Percentage":"33.3%"}. This record does not assert independent reproduction.

### VisionZip

Accuracy: 77. Result ID: 69dc9a8e47f34ce924c338c0.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa760f1e209131cea6b33b.md>)

[Implementation](<https://github.com/dvlab-research/VisionZip>)

Recorded evaluation configuration: {"Tokens Retained Percentage":"22.2%"}. This record does not assert independent reproduction.

### LLaVA-Qwen-7B (Base)

Accuracy: 76.7. Result ID: 6a1d4a7ffbc373592c16aadf.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7c1c1e209131cea8c702.md>)

Recorded evaluation configuration: {"Model":"LLaVA-Qwen-7B (Base)"}. This record does not assert independent reproduction.

### Qwen2.5-VL-7B

Accuracy: 76.6. Result ID: 69d614a747f34ce924beb2b7.

[Reporting paper](<https://www.sota2.com/research/paper/69d5e0fd7d0172140ec759dc.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa773a1e209131cea71d3c.md>)

Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-7B","Token Reduction Ratio":"0%"}. This record does not assert independent reproduction.

### VisionZip

Accuracy: 76.1. Result ID: 69dc9a8e47f34ce924c338c4.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa760f1e209131cea6b33b.md>)

[Implementation](<https://github.com/dvlab-research/VisionZip>)

Recorded evaluation configuration: {"Tokens Retained Percentage":"11.1%"}. This record does not assert independent reproduction.

### TwigVLM++

Accuracy: 76.1. Result ID: 69dc9a8e47f34ce924c338c6.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

Recorded evaluation configuration: {"Tokens Retained Percentage":"11.1%"}. This record does not assert independent reproduction.

### ID-Selection

Accuracy: 75.5. Result ID: 69d614a747f34ce924beb2bd.

[Reporting paper](<https://www.sota2.com/research/paper/69d5e0fd7d0172140ec759dc.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69d5e0fd7d0172140ec759dc.md>)

Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-7B","Token Reduction Ratio":"88.9%","Importance Source":"LLM second-layer cross-modal attention"}. This record does not assert independent reproduction.

### FastV

Accuracy: 75.5. Result ID: 69dc9a8e47f34ce924c338bb.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa74ae1e209131cea632b7.md>)

Recorded evaluation configuration: {"Tokens Retained Percentage":"33.3%"}. This record does not assert independent reproduction.

### Q2.5VL-7B

Accuracy: 75.3. Result ID: 69dc9a8e47f34ce924c338ba.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa70d41e209131cea4d774.md>)

[Implementation](<https://github.com/QwenLM/Qwen-VL>)

Recorded evaluation configuration: {"Tokens Retained Percentage":"100%"}. This record does not assert independent reproduction.

### FastV

Accuracy: 75.3. Result ID: 69dc9a8e47f34ce924c338bf.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa74ae1e209131cea632b7.md>)

Recorded evaluation configuration: {"Tokens Retained Percentage":"22.2%"}. This record does not assert independent reproduction.

### DART

Accuracy: 75.2. Result ID: 69d614a747f34ce924beb2b9.

[Reporting paper](<https://www.sota2.com/research/paper/69d5e0fd7d0172140ec759dc.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa79461e209131cea7cfd7.md>)

Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-7B","Token Reduction Ratio":"77.8%"}. This record does not assert independent reproduction.

### LLaVA-Llama-8B

Accuracy: 75.2. Result ID: 6a1d4a7ffbc373592c16aae3.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7c1c1e209131cea8c702.md>)

Recorded evaluation configuration: {"Model":"LLaVA-Llama-8B"}. This record does not assert independent reproduction.

### TwigVLM

Accuracy: 74.9. Result ID: 69dc9a8e47f34ce924c338c5.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

Recorded evaluation configuration: {"Tokens Retained Percentage":"11.1%"}. This record does not assert independent reproduction.

### FastV

Accuracy: 74. Result ID: 69d614a747f34ce924beb2b8.

[Reporting paper](<https://www.sota2.com/research/paper/69d5e0fd7d0172140ec759dc.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa74ae1e209131cea632b7.md>)

Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-7B","Token Reduction Ratio":"77.8%"}. This record does not assert independent reproduction.

### FastV

Accuracy: 72.8. Result ID: 69dc9a8e47f34ce924c338c3.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa74ae1e209131cea632b7.md>)

Recorded evaluation configuration: {"Tokens Retained Percentage":"11.1%"}. This record does not assert independent reproduction.

### PRISM-3B

Accuracy: 72.8. Result ID: 6a1d4a7ffbc373592c16aada.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7c1c1e209131cea8c702.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7c1c1e209131cea8c702.md>)

Recorded evaluation configuration: {"Model":"PRISM-3B"}. This record does not assert independent reproduction.

### LLaVA-Phi2-3B

Accuracy: 72.7. Result ID: 6a1d4a7ffbc373592c16aad9.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7c1c1e209131cea8c702.md>)

Recorded evaluation configuration: {"Model":"LLaVA-Phi2-3B"}. This record does not assert independent reproduction.

### DART

Accuracy: 72.6. Result ID: 69d614a747f34ce924beb2bc.

[Reporting paper](<https://www.sota2.com/research/paper/69d5e0fd7d0172140ec759dc.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa79461e209131cea7cfd7.md>)

Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-7B","Token Reduction Ratio":"88.9%"}. This record does not assert independent reproduction.

### S-MFT Attn

Accuracy: 72.4. Result ID: 6999366d5698665ec49fbcff.

[Reporting paper](<https://www.sota2.com/research/paper/69535c7253fdd5c36173d7fa.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69535c7253fdd5c36173d7fa.md>)

Recorded evaluation configuration: {"Backbone":"Phi-2-2.7B","Epoch":"0.46","Time (h)":"5.21"}. This record does not assert independent reproduction.

### S-MFT Both

Accuracy: 72.4. Result ID: 6999366d5698665ec49fbd01.

[Reporting paper](<https://www.sota2.com/research/paper/69535c7253fdd5c36173d7fa.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69535c7253fdd5c36173d7fa.md>)

Recorded evaluation configuration: {"Backbone":"Phi-2-2.7B","Epoch":"0.64","Time (h)":"8.12"}. This record does not assert independent reproduction.

### M3

Accuracy: 72.1. Result ID: 69b3bd0c47f34ce9249d633f.

[Reporting paper](<https://www.sota2.com/research/paper/69b39a110636e5bf4afbcade.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa74451e209131cea60a91.md>)

Recorded evaluation configuration: {"#Vision Tokens":"180"}. This record does not assert independent reproduction.

### FastV

Accuracy: 72. Result ID: 69d614a747f34ce924beb2bb.

[Reporting paper](<https://www.sota2.com/research/paper/69d5e0fd7d0172140ec759dc.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa74ae1e209131cea632b7.md>)

Recorded evaluation configuration: {"Base Model":"Qwen2.5-VL-7B","Token Reduction Ratio":"88.9%"}. This record does not assert independent reproduction.

### FFT

Accuracy: 71.9. Result ID: 6999366d5698665ec49fbcfa.

[Reporting paper](<https://www.sota2.com/research/paper/69535c7253fdd5c36173d7fa.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa72421e209131cea5554e.md>)

[Implementation](<https://github.com/DLCV-BUAA/TinyLLaVA-Bench>)

Recorded evaluation configuration: {"Backbone":"Phi-2-2.7B","Epoch":"0.86","Time (h)":"13.53"}. This record does not assert independent reproduction.

### PRISM-13B

Accuracy: 71.8. Result ID: 6a1d4a7ffbc373592c16aade.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7c1c1e209131cea8c702.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7c1c1e209131cea8c702.md>)

Recorded evaluation configuration: {"Model":"PRISM-13B"}. This record does not assert independent reproduction.

### S-MFT MLP

Accuracy: 71.7. Result ID: 6999366d5698665ec49fbd00.

[Reporting paper](<https://www.sota2.com/research/paper/69535c7253fdd5c36173d7fa.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69535c7253fdd5c36173d7fa.md>)

Recorded evaluation configuration: {"Backbone":"Phi-2-2.7B","Epoch":"0.54","Time (h)":"7.33"}. This record does not assert independent reproduction.

### LLaVA-Vicuna-13B

Accuracy: 71.6. Result ID: 6a1d4a7ffbc373592c16aadd.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7c1c1e209131cea8c702.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa732e1e209131cea5af59.md>)

[Implementation](<https://llava-vl.github.io/>)

Recorded evaluation configuration: {"Model":"LLaVA-Vicuna-13B"}. This record does not assert independent reproduction.

### LoRA

Accuracy: 71.1. Result ID: 6999366d5698665ec49fbcfb.

[Reporting paper](<https://www.sota2.com/research/paper/69535c7253fdd5c36173d7fa.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6b121e209131cea2bf59.md>)

[Implementation](<https://github.com/microsoft/LoRA>)

Recorded evaluation configuration: {"Backbone":"Phi-2-2.7B","Epoch":"0.67","Time (h)":"8.15"}. This record does not assert independent reproduction.

### ICONS

Accuracy: 70.8. Result ID: 69e7436d47f34ce924cdd400.

[Reporting paper](<https://www.sota2.com/research/paper/69e704be7d0172140ec7eeac.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa795d1e209131cea7d48d.md>)

Recorded evaluation configuration: {"Training Data Percentage":"20%","Data Selection Constraint":"Full Data Used before Selection"}. This record does not assert independent reproduction.

### FMVR-LLaVA

Accuracy: 70.6. Result ID: 69b3bd0c47f34ce9249d6344.

[Reporting paper](<https://www.sota2.com/research/paper/69b39a110636e5bf4afbcade.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69b39a110636e5bf4afbcade.md>)

Recorded evaluation configuration: {"#Vision Tokens":"2880"}. This record does not assert independent reproduction.

### Uni-LoRA

Accuracy: 70.3. Result ID: 6999366d5698665ec49fbcfe.

[Reporting paper](<https://www.sota2.com/research/paper/69535c7253fdd5c36173d7fa.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7c301e209131cea8ced0.md>)

Recorded evaluation configuration: {"Backbone":"Phi-2-2.7B","Epoch":"0.72","Time (h)":"9.94"}. This record does not assert independent reproduction.

### FMVR-LLaVA

Accuracy: 70.3. Result ID: 69b3bd0c47f34ce9249d6343.

[Reporting paper](<https://www.sota2.com/research/paper/69b39a110636e5bf4afbcade.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69b39a110636e5bf4afbcade.md>)

Recorded evaluation configuration: {"#Vision Tokens":"720"}. This record does not assert independent reproduction.

### PyramidDrop

Accuracy: 70.2. Result ID: 69b3bd0c47f34ce9249d63ed.

[Reporting paper](<https://www.sota2.com/research/paper/69b39a110636e5bf4afbcade.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa772c1e209131cea71934.md>)

Recorded evaluation configuration: {"#Vision Tokens":"192"}. This record does not assert independent reproduction.

### LLaVA-NeXT-7B

Accuracy: 70.2. Result ID: 69dc9a8e47f34ce924c33997.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

Recorded evaluation configuration: {"Retained Tokens":"2880","Base Model":"LLaVA-NeXT-7B"}. This record does not assert independent reproduction.

### TwigVLM++

Accuracy: 70.2. Result ID: 69dc9a8e47f34ce924c3399d.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

Recorded evaluation configuration: {"Retained Tokens":"640","Base Model":"LLaVA-NeXT-7B"}. This record does not assert independent reproduction.

### LLaVA-NeXT-7B

Accuracy: 70.2. Result ID: 69dc9a8e47f34ce924c33a23.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

Recorded evaluation configuration: {"Retained Tokens":"2880","Base Model":"LLaVA-NeXT-7B"}. This record does not assert independent reproduction.

### TwigVLM++

Accuracy: 70.2. Result ID: 69dc9a8e47f34ce924c33a29.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

Recorded evaluation configuration: {"Retained Tokens":"640","Base Model":"LLaVA-NeXT-7B"}. This record does not assert independent reproduction.

### PDrop

Accuracy: 70.1. Result ID: 69a834bf6dbb7075b94eb7e9.

[Reporting paper](<https://www.sota2.com/research/paper/69a66c182d603e6de3e3ce31.md>)

Recorded evaluation configuration: {"Retained Tokens":"128"}. This record does not assert independent reproduction.

### FastV

Accuracy: 70.1. Result ID: 69a834bf6dbb7075b94eb7f0.

[Reporting paper](<https://www.sota2.com/research/paper/69a66c182d603e6de3e3ce31.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa74ae1e209131cea632b7.md>)

Recorded evaluation configuration: {"Retained Tokens":"64"}. This record does not assert independent reproduction.

### LLaVA-NeXT-7B

Accuracy: 70.1. Result ID: 69de25ce47f34ce924c6a920.

[Reporting paper](<https://www.sota2.com/research/paper/69ddca487d0172140ec79c86.md>)

Recorded evaluation configuration: {"Base Model":"LLaVA-NeXT-7B","Retained Tokens":"2880","Pruning Ratio":"100%"}. This record does not assert independent reproduction.

### FMVR-LLaVA

Accuracy: 69.9. Result ID: 69b3bd0c47f34ce9249d63f9.

[Reporting paper](<https://www.sota2.com/research/paper/69b39a110636e5bf4afbcade.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69b39a110636e5bf4afbcade.md>)

Recorded evaluation configuration: {"#Vision Tokens":"9"}. This record does not assert independent reproduction.

### TwigVLM

Accuracy: 69.9. Result ID: 69dc9a8e47f34ce924c3399c.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

Recorded evaluation configuration: {"Retained Tokens":"640","Base Model":"LLaVA-NeXT-7B"}. This record does not assert independent reproduction.

### TwigVLM

Accuracy: 69.9. Result ID: 69dc9a8e47f34ce924c33a28.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7a1c1e209131cea81398.md>)

Recorded evaluation configuration: {"Retained Tokens":"640","Base Model":"LLaVA-NeXT-7B"}. This record does not assert independent reproduction.

### FastV

Accuracy: 69.9. Result ID: 6a1ebf0afbc373592c1a072f.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e63e3dbd7150506f85529.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa74ae1e209131cea632b7.md>)

Recorded evaluation configuration: {"VTR Type":"Text-aware","Tokens Retained":"64","Reduction Ratio":"11.1%"}. This record does not assert independent reproduction.

### FMVR-LLaVA

Accuracy: 69.7. Result ID: 69b3bd0c47f34ce9249d63fb.

[Reporting paper](<https://www.sota2.com/research/paper/69b39a110636e5bf4afbcade.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69b39a110636e5bf4afbcade.md>)

Recorded evaluation configuration: {"#Vision Tokens":"144"}. This record does not assert independent reproduction.

### SparseVLM

Accuracy: 69.7. Result ID: 6a1ebf0afbc373592c1a0731.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e63e3dbd7150506f85529.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa79181e209131cea7c1b1.md>)

Recorded evaluation configuration: {"VTR Type":"Text-aware","Tokens Retained":"64","Reduction Ratio":"11.1%"}. This record does not assert independent reproduction.

### LoRA-One

Accuracy: 69.6. Result ID: 6999366d5698665ec49fbcfd.

[Reporting paper](<https://www.sota2.com/research/paper/69535c7253fdd5c36173d7fa.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa79941e209131cea7e64d.md>)

Recorded evaluation configuration: {"Backbone":"Phi-2-2.7B","Epoch":"0.60","Time (h)":"7.69"}. This record does not assert independent reproduction.

### HoloV + RESTORE

Accuracy: 69.6. Result ID: 6a1ebf0afbc373592c1a0722.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e63e3dbd7150506f85529.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6a1e63e3dbd7150506f85529.md>)

Recorded evaluation configuration: {"VTR Type":"Text-agnostic","Tokens Retained":"192","Reduction Ratio":"33.3%"}. This record does not assert independent reproduction.

Page 1 of 4. This is an index page, not the full collection.

- [Next index page](<https://www.sota2.com/research/sota/scientific-question-answering-on-scienceqa-image.md?metric=Accuracy&page=2>): Follow this link to discover more pages.
