Dialogue Fragment Retrieval on WeChat real-domain (test)
60.53Fragment Order ConsistencyF2RVLM
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| F2RVLM#Params=3B, Zero-shot inference=false2026.06 | 60.53 | 61.18 | |
| Qwen2-VL#Params=7B, Zero-shot inference=false2026.06 | 55.28 | 59.49 | |
| Qwen2.5-VL#Params=7B, Zero-shot inference=false2026.06 | 54.81 | 58.39 | |
| Qwen2.5-VL#Params=72B, Zero-shot inference=true2026.06 | 45.98 | 48.49 | |
| GPT-4o#Params=Closed-Source, Zero-shot inference=true2026.06 | 43.17 | 48.21 | |
| Claude-Sonnet-4#Params=Closed-Source, Zero-shot inference=true2026.06 | 42.29 | 46.64 | |
| Gemini-2.5-Flash#Params=Closed-Source, Zero-shot inference=true2026.06 | 37.67 | 41.44 | |
| MiMo-VL-RL#Params=7B, Zero-shot inference=false2026.06 | 34.69 | 36.69 | |
| Doubao-Seed-1.6#Params=Closed-Source, Zero-shot inference=true2026.06 | 27.89 | 31.06 | |
| DeepSeek-VL2-Small#Params=16B, Zero-shot inference=false2026.06 | 12.52 | 13.67 |