Beyond Abstention: Modeling Clarification-Seeking in LLMs under Concept Incongruence

TL;DR: What if LLMs didn’t just answer or abstain when faced with impossible or incongruent prompts, but instead asked clarifying questions like people do? The first step would be to train or prompt LLMs to recognize concept boundary violations and respond by seeking clarification, measuring the effect on downstream accuracy and user trust.

Research Question: Can LLMs be trained or prompted to recognize concept incongruence and actively seek clarifications, rather than defaulting to answering or abstaining?

Hypothesis: Explicitly modeling a “clarification-seeking” state in LLMs will reduce both erroneous responses and inappropriate abstentions when facing concept-incongruent prompts, as compared to current models that only answer or abstain.

Experiment Plan: Setup: Construct a suite of prompts involving concept boundary violations (e.g., “Draw a unicorn with two horns”) across domains, not just temporal or role-play.
Methodology: Fine-tune or prompt LLMs to, upon detecting incongruence, generate clarification questions (“Did you mean a unicorn or a two-horned horse?”).
Data: Collect both user ratings of model helpfulness and quantitative metrics (accuracy, abstention, clarification rate).
Analysis: Compare rates of clarification-seeking, inappropriate answering, and user satisfaction between baseline and intervention models.

References:

• Bai, X., Peng, I., Singh, A., & Tan, C. (2025). Concept Incongruence: An Exploration of Time and Death in Role Playing. arXiv.org.
• Sun, C.-E., Oikarinen, T. P., Ustun, B., & Weng, T.-W. (2024). Concept Bottleneck Large Language Models. International Conference on Learning Representations.