Fu & Du 2025 — First Ask Then Answer (FATA)

Core Contribution

Proposes a two-stage interaction paradigm where LLMs proactively generate supplementary questions before answering, rather than reactively clarifying only when ambiguity is detected.

Key Insight

The paper identifies an expertise-information gap: AI systems possess extensive knowledge capabilities, but non-expert users lack the domain frameworks necessary to formulate comprehensive queries. Users can’t anticipate what information experts would consider essential.

The solution: leverage AI to externalize expert consultation patterns, transforming implicit professional knowledge-gathering into explicit, accessible questioning frameworks.

Framework Components

Stage 1 (F1): Systematic question generation across five dimensions:

  1. Contextual background
  2. Constraints (resource, time, regulatory)
  3. Preferences (goals, priorities, trade-offs)
  4. Environmental factors
  5. Historical context

Stage 2 (F2): Context-enhanced response generation integrating original query with collected supplementary information.

Key Design Decisions

  • Single-turn question generation: all clarifying questions produced at once rather than multi-turn back-and-forth
  • Proactive rather than reactive: doesn’t wait for detected ambiguity
  • Prompt-only approach: no fine-tuning required, compatible with existing LLMs
  • Built-in constraints: prevents over-questioning, protects privacy

Experimental Results

  • 27.7–47.4% improvement over baseline prompts
  • 2.1–5.4% improvement over expert-level prompts with complete context
  • Stability rates improved from 11–22% to 77.8–100%
  • Tested across 12 industry domains, 300 test cases

Limitations Acknowledged

  • Simulated user responses (not actual humans)
  • Single-turn may miss complex interdependencies
  • Dependent on underlying model quality
  • User comprehension of questions varies

Extractions for heyMHK

  • CLAM (Kuhn et al. 2022) - selective clarification
  • Self-Ask (Press et al. 2022) - model self-questioning
  • Chain-of-Thought prompting
  • RAG limitations