Intent Understanding on Weibo Post-Sync
76.98AccuracyIntPro
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| IntProBackbone=Qwen3-4B, Training strategy=IntPro (Ours)2026.02 | 76.98 | 62.45 | 76.12 | |
| Naive GRPOBackbone=Qwen3-4B, Training strategy=Naive GRPO2026.02 | 75.87 | 60.14 | 75.24 | |
| SFT with Answer-onlyBackbone=Qwen3-4B, Training strategy=SFT with Answer-only2026.02 | 74.89 | 59.83 | 74.55 | |
| IntProBackbone=Qwen2.5-3B, Training strategy=IntPro (Ours)2026.02 | 74.06 | 60.95 | 74.12 | |
| SFT with Retrieval-conditionedBackbone=Qwen3-4B, Training strategy=SFT with Retrieval-conditioned2026.02 | 72.14 | 58.21 | 71.78 | |
| IntProBackbone=Llama3.2-3B, Training strategy=IntPro (Ours)2026.02 | 70.32 | 55.15 | 69.04 | |
| SFT with Answer-onlyBackbone=Qwen2.5-3B, Training strategy=SFT with Answer-only2026.02 | 69.9 | 52.52 | 68.81 | |
| Naive GRPOBackbone=Qwen2.5-3B, Training strategy=Naive GRPO2026.02 | 69.21 | 56.31 | 69.06 | |
| SFT with Retrieval-conditionedBackbone=Qwen2.5-3B, Training strategy=SFT with Retrieval-conditioned2026.02 | 67.32 | 51.18 | 66.74 | |
| BERT-baseBackbone=BERT-base2026.02 | 67.27 | 33.45 | 60.36 | |
| SFT with Answer-onlyBackbone=Llama3.2-3B, Training strategy=SFT with Answer-only2026.02 | 67.26 | 51.23 | 65.89 | |
| Naive GRPOBackbone=Llama3.2-3B, Training strategy=Naive GRPO2026.02 | 67.22 | 51.25 | 67.13 | |
| GPT-4o (w/ retrieval)Backbone=GPT-4o, Retrieval usage=true2026.02 | 65.18 | 42.34 | 58.76 | |
| SFT with Retrieval-conditionedBackbone=Llama3.2-3B, Training strategy=SFT with Retrieval-conditioned2026.02 | 64.58 | 49.86 | 63.42 | |
| RoBERTa-baseBackbone=RoBERTa-base2026.02 | 62.69 | 30.8 | 55.72 | |
| Qwen3-30B-A3B (w/ retrieval)Backbone=Qwen3-30B-A3B, Retrieval usage=true2026.02 | 62.45 | 38.92 | 55.34 | |
| GPT-4oBackbone=GPT-4o2026.02 | 42.16 | 38.58 | 45.49 | |
| Qwen3-30B-A3BBackbone=Qwen3-30B-A3B2026.02 | 41.05 | 30.78 | 36.76 |