Hierarchical semantic classification is essential for taxonomy-driven applications, yet existing classifiers often optimize flat label likelihoods and therefore weakly control parent-child validity, semantic error severity, coarse-to-fine coherence, and out-of-scope rejection. This paper aims to align LLM-based classifiers with the full structured utility of hierarchical prediction rather than treating taxonomy labels as independent targets. We propose TaxoAgent, a utility-aligned policy optimization framework that formulates taxonomy prediction as a verified parent-to-child decision trajectory. Given an input, TaxoAgent first predicts coarse parent labels, expands them into fine-grained child decisions, and verifies the resulting trajectory with a composite utility that jointly measures outcome correctness, hierarchy consistency, semantic proximity, and trajectory coherence; group-relative policy optimization then reinforces high-utility trajectories over low-utility alternatives. Across IOS-APP-ZH, Google-APP-EN, RCV1, WOS, and CLINC150, TaxoAgent consistently improves Micro-F1 and Macro-F1, reduces hierarchy inconsistency, strengthens long-tail classification, and improves OOS-F1 by up to 3.68 points over the strongest prior baseline. These results show that treating taxonomies as structured decision spaces offers a practical and extensible path toward more reliable, adaptable, and utility-aware semantic classifiers.

External memory banks can improve test-time reasoning in frozen large language models without updating backbone parameters. Yet similarity-based retrievers and pointwise selectors score memories independently, ignoring interactions and presentation order. We introduce , an LM-style model for reinforced memory selection that casts context construction as autoregressive index decoding. From a densely retrieved candidate pool, a lightweight Transformer encoder–decoder generates an ordered index sequence [i₁,…,i_k], modeling dependencies among memories while leaving the backbone frozen. This formulation supports standard language-model training. For sparse binary verifiable rewards, we further propose , a confidence-aware extension of group-relative policy optimization that promotes correct, low-confidence rollouts, penalizes confidently incorrect ones, and retains learning signals in reward-homogeneous groups. Across three mathematical reasoning and two reasoning-intensive question-answering benchmarks with three Qwen backbones,  achieves an average score of 84.9 on Qwen-plus, exceeding the strongest memory-augmented baseline by 2.3 points.