Asking the Right Question: Epistemic Inquiry as a Learnable Reasoning Skill for Scientific Discovery

Amin Mansouri, Caglar Gulcehre, Philippe Schwaller

Track: Track 1: Original Research/Position/Education/Attention Track

Keywords: epistemic questioning, question generation, reasoning, chain-of-thought, science QA

TL;DR: Epistemic question-asking is a learnable, transferable reasoning skill that improves LLM performance across science, math, and coding, and a key capability for genuine human–AI scientific collaboration.

Abstract: Effective scientific reasoning depends not only on extending chains of thought, but on identifying what is missing, surfacing implicit assumptions and posing the right questions to unlock a solution. This epistemic dimension is central not just to discovery, but to scientific partnership: a genuine AI co-scientist must be able to articulate its knowledge gaps in a form that human experts can engage with, including gaps that can only be filled by tacit domain knowledge no model possesses. We introduce Chain-of-Questions (CoQ), a prompting and training framework in which a model explicitly generates targeted epistemic questions before attempting to solve a problem, producing a structured, inspectable representation of what it does not know. Across science, chemistry, mathematics, and coding benchmarks, CoQ consistently outperforms chain-of-thought prompting at matched token budgets, and question decompositions trained on one domain transfer to held-out chemistry targets without any chemistry training, suggesting that the skill of decomposing uncertainty into questions is a portable reasoning skill, and not a benchmark artifact. Together, these results point toward a concrete capability that separates a scientific tool from a scientific collaborator: not the ability to answer, but the ability to ask.

Submission Number: 164