The same recipe in other domains
Qwen3-4B as the agent, held-out test sets. Grey is the agent alone; blue is the gain from advice.
Online shoppingWebShop, share of exactly correct purchases
3.8 → 23.2%
Database questionsBIRD text-to-SQL, execution accuracy
40.0 → 43.1%
Expert document extractionExpertLongBench, checklist coverage
49.7 → 54.5%
Figure 5: Results on three public benchmarks outside finance.