This paper presents the concept of "Speaking with Intent (SWI)," which explicitly generates intents in large-scale language models (LLMs) to capture the model's underlying intentions and provide high-level plans that guide subsequent analysis and actions. SWI aims to enhance the inference and generation quality of LLMs by mimicking the intentional and purposeful thinking of humans. Extensive experiments on text summarization, multi-task question answering, and mathematical reasoning benchmarks demonstrate the effectiveness and generalizability of SWI compared to direct generation without explicit intent. We validate the generalizability of SWI across a variety of experimental settings, and human evaluations validate the consistency, effectiveness, and interpretability of the generated intents. Consequently, we propose that enhancing LLMs with explicit intents offers a novel approach to enhancing LLM generation and inference performance.
