Chart Question Answering on CharXiv
76.4Reasoning ScoreGPT-5.4
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| GPT-5.4Model Group=Proprietary models2026.05 | 76.4 | 80.5 | 77.7 | 3.7 | |
| Claude Sonnet 4.6Model Group=Proprietary models2026.05 | 75.7 | 70.4 | 68.2 | 3.2 | |
| Gemini 2.5 ProModel Group=Proprietary models2026.05 | 75.7 | 73.9 | 68.9 | 7.3 | |
| GPT-5.4 miniModel Group=Proprietary models2026.05 | 70.9 | 70.2 | 68.9 | 1.9 | |
| GPT-52026.05 | 69.1 | 90.03 | 85.84 | — | |
| Gemini-2.5-Pro2026.05 | 68.4 | 91.22 | 86.66 | — | |
| Gemini 2.5 FlashModel Group=Proprietary models2026.05 | 68.2 | 65.9 | 62.2 | 6 | |
| GPT-4.1 miniModel Group=Proprietary models2026.05 | 62.8 | 48.4 | 42.6 | 13.7 | |
| Claude Haiku 4.5Model Group=Proprietary models2026.05 | 61.5 | 58 | 55.4 | 4.8 | |
| Claude-3.5-Sonnet2026.05 | 60.2 | 84.3 | 79.48 | — | |
| GPT-4.1Model Group=Proprietary models2026.05 | 57.4 | 54.4 | 54.1 | 0.6 | |
| Qwen3-VL-8B-InstructModel Group=Open-source models2026.05 | 53.4 | 51 | 48.6 | 4.9 | |
| GPT-4o2026.05 | 47.1 | 84.45 | 76.98 | — | |
| Qwen3-VL + ChartCFModel Size=8B, # Training Data=4K pairs (DPO)2026.05 | 46.5 | 82.58 | 75.36 | — | |
| Chart-R1Model Size=7B, # Training Data=228K (SFT) + 30K (RL)2026.05 | 46.2 | 62 | 58.84 | — | |
| Qwen2.5-VL + ChartCFModel Size=7B, # Training Data=4K pairs (DPO)2026.05 | 44.4 | 75.08 | 68.94 | — | |
| Gemma 4 E4B ITModel Group=Open-source models2026.05 | 43.2 | 41.4 | 35.8 | 15.7 | |
| Qwen3-VLModel Size=8B2026.05 | 43.1 | 81.25 | 73.62 | — | |
| GPT-4oModel Group=Proprietary models2026.05 | 42.6 | 39.4 | 42.6 | -7.5 | |
| InternVL3.5 + ChartCFModel Size=8B, # Training Data=4K pairs (DPO)2026.05 | 42.3 | 80.47 | 72.84 | — | |
| Qwen3-VL + ECDModel Size=8B, # Training Data=300K (SFT)2026.05 | 42.2 | 81.93 | 73.98 | — | |
| Qwen2.5-VLModel Size=7B2026.05 | 41.2 | 66.4 | 61.36 | — | |
| InternVL3.5 + ECDModel Size=8B, # Training Data=300K (SFT)2026.05 | 40.3 | 77.03 | 69.68 | — | |
| Qwen2.5-VL + ECDModel Size=7B, # Training Data=300K (SFT)2026.05 | 40.2 | 74.2 | 67.4 | — | |
| InternVL3.5Model Size=8B2026.05 | 40 | 75 | 68 | — | |
| Pixtral-12BModel Group=Open-source models2026.05 | 39.2 | 32.4 | 34.5 | -6.1 | |
| Qwen2.5-VL-7B-InstructModel Group=Open-source models2026.05 | 34.5 | 35.2 | 31.1 | 13.3 | |
| GPT-4o mini2026.05 | 34.1 | 74.92 | 66.76 | — | |
| InternVL3-8BModel Group=Open-source models2026.05 | 33.8 | 30.8 | 31.8 | -3 | |
| ReachQA (InternVL2-8B)Model Size=8B, # Training Data=20K (SFT)2026.05 | 32.7 | 54.83 | 50.4 | — | |
| LLaVA-OneVision-Qwen2-7BModel Group=Open-source models2026.05 | 20.3 | 19.7 | 27 | -27 | |
| ChartGemmaModel Size=3B, # Training Data=123K (SFT)2026.05 | 12.5 | 21.3 | 19.54 | — | |
| TinyChartModel Size=3B, # Training Data=1.36M (SFT)2026.05 | 8.3 | — | — | — |