Technical Summary: Knowing but Not Showing

Problem Statement

Large Language Models (LLMs) are increasingly deployed to answer user queries, yet these queries are frequently underspecified, leaving out crucial details or admitting multiple valid interpretations. When faced with such ambiguity, a helpful assistant should ideally recognize the uncertainty and seek clarification rather than silently making assumptions. However, current models often default to providing a single "best guess" answer, forcing hidden assumptions that can misrepresent the user's intent and propagate errors.

The core problem investigated is the discrepancy between an LLM's ambiguity awareness (the ability to recognize that a query is ambiguous) and its clarification behavior (the tendency to act on that recognition by asking questions or refusing to answer). The authors posit that while models may internally "know" a query is ambiguous, they rarely "show" this knowledge in their default responses.

Methodology

The study evaluates LLMs across three complementary settings using the AmbigQA dataset, which contains ambiguous questions, unambiguous questions, and disambiguated rewrites of the ambiguous ones. The evaluation involves 10 models from three families (OpenAI, Claude, and Qwen) and includes:

1. Standard Question Answering (QA): Models are prompted to answer questions directly, with and without retrieved context (Wikipedia passages). Performance is measured by accuracy against reference answers.
2. Explicit Ambiguity Judgment: Models are asked to perform a binary classification task: determining whether a specific question is ambiguous or unambiguous. This probes the model's explicit awareness without requiring an answer.
3. Behavioral Analysis: A judge model (GPT-5-nano) classifies model responses into four categories: Refusal, Asked a Clarifying Question, Answered (providing an answer), and Only Answered (providing an answer without refusal or clarification).

The study also analyzes the specific types of ambiguity (Temporal, Identity, Version, Scope, Semantic, Locale, Other) that models identify when they do recognize ambiguity, comparing these distributions against human annotations.

Key Results

1. The "Knowing but Not Showing" Gap

The most significant finding is a clear disconnect between recognition and behavior.

• Recognition: When explicitly asked to judge ambiguity, models exhibit non-trivial awareness. They can often correctly identify that a query is ambiguous, achieving 60–80% accuracy on ambiguous classes (though often biased toward over-predicting ambiguity).
• Behavior: In the standard QA setting, models overwhelmingly default to direct answers. Clarifying questions are extremely rare (typically <1% to 5% of responses), and refusals are almost non-existent. Even when models correctly identify a query as ambiguous in the judgment setting, they fail to translate this into clarification behavior when prompted to answer.

2. The Impact of Retrieved Context

Retrieved context (RAG) produces counter-intuitive effects on ambiguity handling:

• QA Accuracy: Context substantially improves answer accuracy for both ambiguous and unambiguous questions, particularly for smaller models.
• Suppression of Clarification: The presence of context makes models less likely to ask clarifying questions. Context acts as a cue that the question is answerable, causing models to treat even ambiguous queries as effectively unambiguous.
• Widened Gap: While context improves the ability to answer, it further widens the gap between ambiguity recognition and clarification behavior. Models become even more reluctant to flag ambiguity when supporting passages are present.

3. Ambiguity Type Bias

When models do identify ambiguity, their classification of the type of ambiguity is skewed. Models consistently over-assign Identity (multiple entities with the same name) and Version ambiguities, while significantly under-assigning Scope ambiguities (missing granularity levels). This suggests models rely on specific, surface-level signals rather than deep reasoning about the structure of the query.

Significance and Claims

The paper claims to demonstrate that current training pipelines, which incentivize models to produce responses rated as "helpful" and "useful" by human annotators, inadvertently teach models to hide uncertainty. Because expressing ambiguity or asking for clarification is often not aligned with high reward in standard RLHF (Reinforcement Learning from Human Feedback) pipelines, models learn to suppress this awareness in favor of providing a direct answer.

The authors argue that this has two primary implications:

1. Training Objectives: To improve truthfulness and reliability, training objectives must explicitly reward the expression of uncertainty, the asking of clarifying questions, and the acknowledgment of under-specification, rather than focusing solely on answer accuracy.
2. Evaluation Caution: When probing what LLMs "know," the formulation of the task matters significantly. A model may display high ambiguity awareness when asked to analyze a question but display near-zero awareness when asked to answer it.

The study concludes that the "knowing but not showing" pattern is a systemic issue where internal awareness does not reliably surface in default behavior, a phenomenon that is further suppressed by the presence of retrieved context.