TLDR: The First Ask Then Answer (FATA) framework is a new interaction paradigm for Large Language Models (LLMs) that guides them to proactively generate multi-dimensional supplementary questions for users before providing a response. This approach addresses incomplete user information by simulating expert consultation patterns, leading to significantly improved response quality, relevance, and stability. Experiments show FATA substantially outperforms baseline methods and offers consistent enhancements over context-enhanced prompts, making LLMs more effective and reliable.
Large Language Models (LLMs) have become incredibly powerful, but they often struggle when users don’t provide all the necessary information. Imagine asking for medical advice without mentioning your symptoms, or seeking financial planning without detailing your current savings. The answers you get might not be very helpful, or even accurate.
To tackle this common problem, researchers Chuanruo Fu and Yuncheng Du from Beijing Information Science and Technology University have introduced an innovative approach called First Ask Then Answer (FATA). This new framework guides LLMs to proactively ask users for more details before generating a response. Instead of waiting for ambiguity, FATA anticipates what information an expert would need and asks for it upfront, all in a single round of questions.
The core idea behind FATA is to bridge the ‘expertise-information gap.’ While LLMs have vast knowledge, everyday users might not know what specific details are crucial for a comprehensive answer. FATA acts like a knowledgeable consultant, transforming the implicit questions an expert would ask into clear, explicit questions for the user. This helps non-experts provide more complete and relevant context, leading to much better responses.
FATA operates in two main stages. First, it generates a comprehensive set of multi-dimensional questions. These questions cover various aspects like contextual background, specific constraints, user preferences, environmental factors, and historical context. This systematic approach ensures that all critical information dimensions are explored. Second, once the user provides the supplementary information, FATA integrates it with the original query to produce a personalized and expert-level answer. This single-turn questioning strategy is designed to be efficient, avoiding the back-and-forth of multi-turn dialogues that can often lead to confusion or context loss.
The advantages of FATA are significant. It’s a proactive system, preventing information gaps rather than just reacting to detected problems. It offers a comprehensive information architecture, moving beyond simple clarification to systematic data gathering. It enhances user capabilities by guiding them through the information-gathering process, effectively democratizing access to professional-grade information structuring. Furthermore, it optimizes interaction efficiency by collecting all necessary information in one go.
From a practical standpoint, FATA is designed for easy deployment. It uses a ‘prompt-only’ approach, meaning it doesn’t require complex fine-tuning or modifications to existing LLMs. This makes it compatible with current systems and readily usable in various production environments. It also includes built-in quality control to prevent excessive questioning and ensure user privacy.
The researchers evaluated FATA against two control methods: a Baseline Prompt (B-Prompt) which simulates typical incomplete user queries, and a Context-Enhanced Expert Prompt (C-Prompt) which represents an ideal scenario where all information is provided upfront. The experiments, conducted across 12 industry domains using models like OpenAI O3, DeepSeek-R1-0528, and Claude 4 Opus Thinking, showed impressive results. FATA significantly outperformed the B-Prompt, with improvements ranging from 27.7% to 47.4% in overall response quality. Claude, in particular, showed the highest responsiveness to FATA, with a 47.4% improvement.
While the improvements over the C-Prompt were more modest (2.1-5.4%), FATA demonstrated superior stability in its responses, achieving consistency rates of 77.8% to 100% across models, compared to C-Prompt’s 0-89% range. This indicates that FATA not only improves quality but also makes LLM responses more reliable and consistent across diverse scenarios.
The most notable improvements with FATA were seen in areas like ‘Persona Recall,’ ‘Relevance,’ and ‘Information Completeness,’ highlighting its strength in understanding and addressing specific user needs. Although ‘Conciseness’ showed smaller gains, this refers to the final answer; the overall interaction involves an additional questioning stage, which is a trade-off for the substantial quality improvements.
Also Read:
FATA represents a significant step forward in human-AI interaction, moving towards a more collaborative and effective dialogue. By empowering LLMs to ask the right questions at the right time, it ensures that users receive more accurate, personalized, and actionable responses. For more details, you can read the full research paper here.
