guinansu/AnyBudgetSearchAgent-Qwen3-4B-softgate-step250

AnyBudget soft gate (solvability-weighted anytime reward, phi(p)=p; dense budget grid 0-30 + 40/52/66, left-endpoint weights), Qwen3-4B multi-turn search agent, step 250 (resumed from vanilla-GRPO step 200)

initial commit