Retrieval-Induced Clarification Suppression in Large Language Models

Retrieval-augmented generation is designed to resolve uncertainty through external evidence. Yet relevant evidence can create a false sense of resolution when the information needed to answer a question must come from the user. We identify and systematically investigate retrieval-induced clarification suppression: the tendency of language models to abandon clarification when retrieved context leaves the underlying underspecification unresolved. Across nine models and three medical and legal datasets, non-resolving retrieval substantially reduces question-asking, including from 83.3% to 27.6% on HealthBench and from 93.3% to 15.0% on Learned Hands for Claude Sonnet 4.6. Strikingly, these reductions frequently occur while independently elicited information need remains unchanged or increases, exposing a dissociation between assessing the need for information and acting on it. Controlled retrieval experiments reveal a second failure pattern, in which models also underestimate the need for clarification, and demonstrate that the framing of identical evidence can alter whether clarification occurs. This diagnosis motivates targeted information-need reflection: a pre-answer assessment of missing user information that does not explicitly instruct the model to ask questions. Reflection substantially restores clarification across five models and both domains. For Sonnet, clarification under retrieval increases from 53.7% to 84.0% on medical queries and from 15.0% to 83.9% on legal queries, whereas matched generic chain-of-thought prompting produces no comparable recovery. A human-validated relevance evaluation further shows that recovered questions target genuinely missing information in several models, while revealing that greater question-asking alone does not guarantee relevant clarification. Our findings expose a consequential failure of retrieval-augmented interaction: acquiring knowledge can inhibit the very information-seeking behavior required to apply it.