🤖 AI

Do LLMs Know What to Ask and When? Evaluating Multi-Turn Information Seeking

This paper introduces MT-InfoSeek, a controlled evaluation suite demonstrating that while large language models can recognize the need for additional information in underspecified multi-turn tasks, they systematically underestimate the required amount, fail to identify minimal sufficient queries, and struggle with query ordering, revealing a distinct gap between their information-seeking capabilities and current evaluation metrics.

Original authors: Yepeng Huang, Jiawen Zhang, Michelle Dai, Xiaorui Su, Shanghua Gao, Zi Wang, Marinka Zitnik

Published 2026-08-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yepeng Huang, Jiawen Zhang, Michelle Dai, Xiaorui Su, Shanghua Gao, Zi Wang, Marinka Zitnik

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a doctor sitting across from a patient who says only, "I have a pain." To make a diagnosis, the doctor cannot guess; they must ask specific questions to fill in the missing details. Is the pain sharp or dull? Where exactly is it located? How long has it lasted? In the world of artificial intelligence, large language models are often tested on their ability to answer questions, but they are rarely tested on their ability to know when they do not have enough information to answer at all. This is a critical gap. A capable AI should recognize when a situation is incomplete, figure out exactly what is missing, and ask for it before offering a conclusion. If it guesses too soon, it risks being confidently wrong.

Researchers at Harvard University and Google DeepMind have built a new way to test this specific skill. They created a controlled environment where an artificial intelligence is given a puzzle with missing pieces. The puzzle could be a math problem with a hidden number, a logic riddle with a missing fact, or a medical scenario with an unasked symptom. The goal for the AI is not just to solve the puzzle, but to navigate the process of asking the right questions in the right order until the answer becomes clear. The researchers wanted to see if these models could act like a skilled investigator who knows exactly what evidence is needed, or if they would rush to a conclusion with insufficient clues.

The team developed a testing suite called MT-INFOSEEK, which contains over 5,000 distinct problems across five different fields: mathematics, logic, biology, medicine, and general knowledge. In these tests, the AI is given a starting point that is deliberately incomplete. For example, in a medical test, the AI might know a patient has pelvic pain but not know how long the pain has lasted or whether the patient is pregnant. The AI must then engage in a conversation, asking one question at a time to an "oracle"—a system that provides the true answer to the question. The AI must decide when to stop asking and give its final answer. The researchers measured success not just by whether the final answer was correct, but by whether the AI actually gathered enough information to make that answer certain.

The results revealed a significant weakness in current models. While the AI systems generally recognized that they were missing information, they consistently underestimated how much they needed. When the researchers created problems that required two missing pieces of information to solve, the models often guessed that only one piece was missing. In logical puzzles, the models were about four times more likely to underestimate the missing information than to overestimate it. They also struggled to identify the specific questions that would actually solve the problem. Even when the researchers told the models exactly how many questions they needed to ask, the models often failed to pick the right set of questions. They would ask irrelevant questions or stop asking too early, leaving the puzzle unsolved.

The study also looked at the order in which questions were asked. In fields like medicine, the sequence of information matters. A doctor must determine if a pain is acute or chronic before deciding which tests to order next. The researchers found that when AI models asked the right questions but in the wrong order, their final accuracy dropped significantly. This suggests that knowing what to ask is only half the battle; knowing when to ask it is equally important. Furthermore, the researchers discovered that a model could recover from a bad first question. If a model asked a useless question at the start but then continued to ask relevant questions in the following turns, it could still reach the correct solution. However, many models failed to persist, often stopping after a few turns even when they had not gathered enough evidence.

Perhaps the most surprising finding was that a model's ability to ask good questions is separate from its ability to give a good answer. The researchers tested this by taking a conversation where the AI had gathered all the necessary information and then asking a different model to solve the puzzle based on that conversation alone. They found that some models could solve the puzzle perfectly when given the full set of facts, even if they had failed to gather those facts themselves. This means that the failure was not in the reasoning or the knowledge of the model, but in the process of seeking the information. The models were not bad at solving problems; they were bad at knowing what they didn't know.

The researchers concluded that current evaluations of artificial intelligence are incomplete. Measuring only the final accuracy of a model's answer hides the fact that the model may have reached that answer by guessing or by ignoring missing information. To truly understand how well an AI interacts with the world, we must measure its ability to recognize uncertainty, identify what is missing, and acquire that missing information step by step. This new framework provides a way to do exactly that, showing that while today's models are powerful, they still lack the intuitive sense of a human expert who knows when to stop guessing and start asking.

Technical Summary: Do LLMs Know What to Ask and When? Evaluating Multi-Turn Information Seeking

1. Problem Definition

The paper addresses the capability of Large Language Models (LLMs) to operate under underspecification in multi-turn interactions. Unlike ambiguity (where a user's intent is unclear), underspecification occurs when a user's task is clearly defined but lacks sufficient information to determine a unique answer. A capable model must recognize that its context is insufficient, identify the specific missing information required to resolve the target, acquire it through sequential queries, and only respond once the acquired information uniquely determines the answer.

Current evaluation methods often conflate a model's ability to generate a correct answer with its ability to seek information. Final task success can be achieved through guessing or relying on prior knowledge rather than genuine information acquisition. The authors formalize this challenge as solving a $k$ -underspecified Constraint Satisfaction Problem (CSP), where $k$ represents the number of variables jointly required to determine the target variable.

2. Methodology: MT-INFOSEEK Framework

The authors introduce MT-INFOSEEK, a controlled evaluation suite comprising 5,251 problems and 9,006 task instances across five domains: mathematics, logic, biology, medicine, and general knowledge (via a 20 Questions game).

2.1 Formalization

A problem is defined as a tuple $P = \langle X, D, C, A, Y \rangle$ :

$X$ : Set of variables with domains $D$ .
$C$ : Logical constraints.
$A$ : Observed partial assignment (initial context).
$Y$ : Target variable.
Underspecification: The problem is underspecified if $Y$ is not uniquely determined by $A$ and $C$ .
Minimal $k$ -Sufficient Set ( $k$ -MSS): A subset of unassigned variables $S$ of size $k$ such that knowing their values uniquely determines $Y$ , and no smaller subset suffices.
Task Instance: A specific realization $(P, y)$ where $y$ is the hidden ground-truth value of $Y$ .

2.2 Evaluation Protocols

The framework evaluates models along three axes, separating information seeking from answer generation:

$k$ -Prediction: Given $P$ , the model predicts the degree of underspecification ( $\hat{k}$ ).
Missing Variables Identification: Given $P$ (and optionally the true $k$ ), the model selects a minimal sufficient set of variables.
Sequential Task-Solving: The model interacts with an oracle over multiple turns, querying one variable at a time, observing the value, and deciding when to stop and answer.

2.3 Datasets

Structured Domains (Logic-Q-MT, GSME-Q-MT/Ext, GeneReg-MT, ClinGuide-MT): Variables and constraints are predefined. The model selects from a finite set of queryable variables.
- Logic-Q-MT: Extends logical reasoning problems to multi-turn settings.
- GSME-Q-MT/Ext: Grade-school math problems with masked quantities; Ext version features deeper dependency structures.
- GeneReg-MT: Boolean gene regulatory networks where the model must identify initial states to determine steady states.
- ClinGuide-MT: Diagnostic decision trees where query order matters due to conditional dependencies.
Unstructured Domain (20Q): The model generates free-form yes/no questions to deduce a hidden target from a candidate set.

2.4 Metrics

Crucially, the paper introduces Final Sufficiency as a primary metric. This measures whether the acquired information (the set of variables queried) uniquely determines the target, independent of whether the model's final answer is correct. This separates the ability to seek information from the ability to reason to an answer. Other metrics include query coverage, order correctness (in ClinGuide-MT), and question informativeness (Normalized Expected Information Gain in 20Q).

3. Key Results

The evaluation covers closed and open-weight LLMs (e.g., GPT-5, Gemini-3, Qwen3, gpt-oss).

3.1 Underestimation of Missing Information

Models generally recognize that context is insufficient (rarely predicting $k=0$ for underspecified problems) but systematically underestimate the degree of underspecification.
Accuracy drops sharply as $k$ increases. Specifically, in logical problems (Logic-Q-MT) at $k = 2$ , models under-predict the degree of missing information about four times as often as they over-predict it.

3.2 Failure to Identify Minimal Sufficient Sets

Even when provided with the true $k$ , models struggle to identify the specific variables that form a minimal sufficient set. Exact-set accuracy is low (below 0.4 for $k=2$ ) and drops further for higher $k$ .
Estimating how much information is missing does not imply knowing which information is needed.

3.3 Interaction Dynamics and Recovery

Turn Budget: Allocating queries across more turns generally improves final sufficiency compared to issuing all queries in a single turn, provided the model does not stop prematurely.
Recovery: A first-turn query that misses the minimal sufficient set is often recoverable. Interactions where the first query is incorrect but the model continues to search for relevant variables reach sufficiency at rates comparable to those with correct first queries.
Late-Turn Performance: For highly underspecified problems ( $k \ge 3$ ), success depends less on early-turn correctness and more on sustained information acquisition and late-turn query correctness.

3.4 The Importance of Order and Integration

Order Dependencies: In clinical tasks (ClinGuide-MT), querying the correct variables in the wrong order significantly reduces final accuracy, even if all necessary information is eventually acquired.
Evidence Integration: In 20Q, models with similar question informativeness (Expected Information Gain) show substantial differences in final accuracy. This indicates that success depends not just on asking good questions, but on effectively integrating the answers to narrow the candidate space.
Entropy Reduction: While models generally reduce uncertainty, they often plateau in later turns, failing to distinguish between closely related remaining candidates.

4. Contributions and Significance

The paper makes three primary contributions:

Formalization: It formalizes multi-turn information seeking as solving $k$ -underspecified CSPs, providing ground truth for the variables needed and the point of sufficiency.
Benchmark (MT-INFOSEEK): It introduces a comprehensive evaluation suite spanning mathematics, logic, biology, medicine, and general knowledge, allowing for controlled variation in the degree of underspecification.
Evaluation Methodology: It proposes Final Sufficiency as a metric to separate the ability to acquire information from the ability to generate answers.

Significance:
The authors argue that current LLM evaluations, which rely on final accuracy, fail to capture interactive competence. A model may answer correctly by guessing or using priors without ever identifying the missing information. By measuring whether the acquired information actually determines the target, MT-INFOSEEK reveals that current models:

Underestimate the complexity of missing information.
Fail to plan minimal query strategies.
Often stop before acquiring sufficient information.
Struggle with ordered dependencies and evidence integration.

The paper concludes that the ability to seek information over multiple turns is a distinct capability from answer generation and is not adequately measured by existing benchmarks. This work provides a diagnostic framework to understand where models fail in interactive settings, suggesting that future improvements must focus on recognizing underspecification and planning sequential information acquisition, not just on reasoning capabilities.