Structured tax-form completion on TaxCalcBench TY24 (train)
90Train-A ScoreDPA-GRPO
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| DPA-GRPOBackbone=Qwen3-8B2026.05 | 90 | 68 | 58 | |
| Zero-shotBackbone=Qwen3-8B2026.05 | 84 | 42 | 42 | |
| GRPOBackbone=Qwen3-8B2026.05 | 84 | 63 | 32 | |
| DAPOBackbone=Qwen3-8B2026.05 | 84 | 63 | 53 | |
| GSPOBackbone=Qwen3-8B2026.05 | 84 | 68 | 58 | |
| GRPOBackbone=Qwen3-4B2026.05 | 74 | 47 | 32 | |
| DAPOBackbone=Qwen3-4B2026.05 | 74 | 37 | 21 | |
| GSPOBackbone=Qwen3-4B2026.05 | 74 | 37 | 21 | |
| DPA-GRPOBackbone=Qwen3-4B2026.05 | 74 | 58 | 37 | |
| Zero-shotBackbone=Qwen3-4B2026.05 | 53 | 32 | 26 |