Structured tax-form completion on TaxCalcBench TY24 (test)
55Case 1 AccuracyDAPO
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| DAPOBackbone=Qwen3-8B2026.05 | 55 | 0 | 45 | — | 55 | |
| GRPOBackbone=Qwen3-8B2026.05 | 54 | 0 | 46 | — | 54 | |
| Zero-shotBackbone=Qwen3-8B2026.05 | 53 | 0 | 47 | — | 53 | |
| GRPOBackbone=Qwen3-4B2026.05 | 52 | 0 | 48 | — | 52 | |
| GSPOBackbone=Qwen3-8B2026.05 | 52 | 0 | 48 | — | 52 | |
| DPA-GRPOBackbone=Qwen3-4B2026.05 | 52 | 1 | 40 | 6 | 53 | |
| DAPOBackbone=Qwen3-4B2026.05 | 47 | 0 | 53 | — | 47 | |
| GSPOBackbone=Qwen3-4B2026.05 | 46 | 0 | 54 | — | 46 | |
| DPA-GRPOBackbone=Qwen3-8B2026.05 | 43 | 13 | 30 | 11 | 57 | |
| Zero-shotBackbone=Qwen3-4B2026.05 | 41 | 0 | 59 | — | 41 |