FinBACE: Branch-Aware Credit Assignment for Evidence Seeking in Financial Risk Investigations

Financial risk investigations require an agent to request missing evidence and stop when further requests are no longer needed. Grouped reinforcement learning can lose corrective feedback for these decisions: when all sampled responses choose the wrong ask-or-stop branch and receive the same reward, their centered reward advantages vanish. We introduce FinBACE, a branch-aware credit estimator that compares question quality among eligible CONTINUE responses and adds bounded negative credit to all-wrong groups after native advantage estimation. The estimator reuses existing task scores and labels without additional judge calls, a learned critic, or extra rollouts. On offline financial expert trajectories, FinBACE restores stopping under GRPO, while ablations distinguish the roles of question and branch credit. An independent GRPO experiment on the public RealMedConv test set reproduces stopping recovery with similar conditional question-quality scores. These findings identify branch-aware credit assignment as a practical approach to learning evidence selection and stopping.