Current Mixture-of-Experts (MoE) models typically rely on a static Top-k routing mechanism, activating a fixed number of experts for every token regardless of its difficulty. This uniform allocation ignores the inherent variability in linguistic information density that simple functional words require significantly less processing capacity than content-heavy terms. The static routing results in computational waste for easy tokens and potential under-fitting for difficult ones. To address this issue, we propose **CAR**, a Complexity-Aware Routing framework that aligns the computational budget with the intrinsic information density of input tokens. Diverging from previous dynamic methods that rely on unstable heuristics, we ground our routing decision in a dual-signal objective combining Shannon Surprisal and Predictive Entropy to objectively measure token complexity. We implement a decoupled architecture where a Complexity Perception (CP) module utilizes a Recurrent Complexity Memory (RCM) to estimate these signals from a teacher model, guiding a Dynamic Expert Allocator (DEA) to adjust k on the fly. Extensive experiments demonstrate that CAR consistently outperforms both static baselines and SOTA dynamic routing methods. Notably, our approach achieves superior accuracy on reasoning-intensive benchmarks (e.g., boosting MMLU by 0.8 and GSM8K by 1.5 over static routing) while reducing the average number of activated parameters by approximately 10%. Our code is available at https://github.com/Acking646/car.

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.