List your company

Structured tax-form completion on TaxCalcBench TY24 (test)

55Case 1 AccuracyDAPO
40.4444.224851.78May 8, 2026

Evaluation results

MethodLinks
DAPOBackbone=Qwen3-8B2026.0555045—55
GRPOBackbone=Qwen3-8B2026.0554046—54
Zero-shotBackbone=Qwen3-8B2026.0553047—53
GRPOBackbone=Qwen3-4B2026.0552048—52
GSPOBackbone=Qwen3-8B2026.0552048—52
DPA-GRPOBackbone=Qwen3-4B2026.0552140653
DAPOBackbone=Qwen3-4B2026.0547053—47
GSPOBackbone=Qwen3-4B2026.0546054—46
DPA-GRPOBackbone=Qwen3-8B2026.054313301157
Zero-shotBackbone=Qwen3-4B2026.0541059—41