SAGE: Semantic Alignment and Geometric Enhancement for Efficient Few-Shot Intent Detection

Few-shot intent detection (FSID) requires recognizing user intents with limited labeled examples, a critical challenge in dialogue systems. Existing pretrained models suffer from representation anisotropy, where sentence embeddings concentrate in narrow conical regions, resulting in insufficient angular margins between intent prototypes and degraded classification boundaries. While Large Language Models (LLMs) demonstrate strong few-shot capabilities, their prohibitive computational costs and high latency preclude practical deployment in real-time applications. To address these limitations, we propose Semantic Alignment and Geometric Enhancement (SAGE), a multi-stage framework that refines intent representations through geometric regularization. SAGE comprises three modules: (1) Semantic Structure Parsing calibrates intent prototypes via semantic anchors. (2) Structured Consistency Regularization enforces isotropic distributions to maximize inter-class separability. (3) Iterative Adaptive Distillation smooths decision boundaries through manifold optimization. Extensive experiments on standard FSID benchmarks demonstrate that SAGE achieves accuracy comparable to or exceeding that of state-of-the-art LLM-based methods (e.g., 2.3% improvement on BANKING77 under the 5-shot setting), while reducing inference latency by 3 to 4 orders of magnitude. Our work provides an efficient and theoretically principled solution for scalable intent understanding, bridging the gap between representation quality and computational efficiency.