To address the limited controllability, unstable output consistency, and weakly constrained decision processes of large language models in text classification tasks, this work proposes a controllable prompt-driven text classification method that establishes an end-to-end unified modeling framework from instruction alignment to constrained decoding. Text classification is reformulated as an instruction-conditioned generative discriminative problem. Input texts and task instructions are jointly encoded to form a unified internal representation that integrates textual semantics with classification constraints. On this basis, a category semantic alignment mechanism is introduced to ensure that the model explicitly follows category boundaries and decision criteria defined by the instructions, thereby reducing classification inconsistency caused by prompt variation or implicit bias. To further improve output reliability, a structured constrained decoding strategy is designed to restrict the generation space to a predefined set of valid categories, preventing redundant text or invalid outputs from interfering with classification results. Comparative analysis under unified data and evaluation settings demonstrates that the proposed method achieves more consistent advantages in classification accuracy, discriminative stability, and overall separability. These findings indicate that deeply integrating instruction understanding and output control into the decision process of large language models effectively transforms their generative capacity into stable, interpretable, and controllable text classification capability, providing a systematic solution for building reliable intelligent text analysis systems.

Hallucinations in large language models (LLMs) remain a critical challenge for deploying these systems in high-stakes domains. We introduce Calib-Abstain, a framework that combines token-level uncertainty estimation with multi-path consistency checking to generate calibrated confidence scores. When confidence falls below a learned threshold, the system triggers tiered mitigation actions: retrieval augmentation, query reformulation, or abstention with explanation. Our approach employs temperature scaling and isotonic regression to minimize Expected Calibration Error (ECE). Experiments on mathematical reasoning (GSM8K, MATH), code generation (HumanEval), and general knowledge tasks (MMLU) demonstrate that Calib-Abstain reduces hallucination rates among answered queries by 24--31% while maintaining 80--88% coverage, with ECE improvements of 0.08--0.12 across datasets.