PaperlayerView exact revision on arXiv

Source-linked AI summary

Active Task Disambiguation with LLMs

Katarzyna Kobalczyk, Nicolas Astorga, Tennison Liu, Mihaela van der Schaar

arXiv:2502.04485v1cs.CLcs.AIcs.LG

TL;DR

Ambiguously specified tasks are common in real-world LLM use, but benchmarks typically assume complete problem statements. The paper formalizes task ambiguity and uses Bayesian Experimental Design to select information-gaining clarifying questions, reporting better disambiguation than implicit-reasoning baselines.

  • Problem

    LLM benchmarks typically assume complete, unambiguous problem statements, although real-world tasks may be underspecified and user intentions may differ from model preferences.

    Source: §1 · p. 1

  • Method

    The paper formalizes task ambiguity and applies Bayesian Experimental Design to select clarifying questions by estimating information gain over compatible solutions.

    Source: §1 · p. 2, §2 · p. 3, §2.1 · p. 5

  • Results

    The BED-based method improves task disambiguation over baseline strategies relying on implicit reasoning about which question to ask.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Clarifications can narrow the compatible solution set and reduce the risk of unsatisfactory responses.

    Source: §2 · p. 3

    The work focuses on requirement elicitation, leaving ambiguity detection and deciding when to stop querying for future research; its strategies also require more LLM calls than baselines.

    Source: §5 · p. 10

Abstract

from arXiv · show

Despite the impressive performance of large language models (LLMs) across various benchmarks, their ability to address ambiguously specified problems--frequent in real-world interactions--remains underexplored. To address this gap, we introduce a formal definition of task ambiguity and frame the problem of task disambiguation through the lens of Bayesian Experimental Design. By posing clarifying questions, LLM agents can acquire additional task specifications, progressively narrowing the space of viable solutions and reducing the risk of generating unsatisfactory outputs. Yet, generating effective clarifying questions requires LLM agents to engage in a form of meta-cognitive reasoning, an ability LLMs may presently lack. Our proposed approach of active task disambiguation enables LLM agents to generate targeted questions maximizing the information gain. Effectively, this approach shifts the load from implicit to explicit reasoning about the space of viable solutions. Empirical results demonstrate that this form of question selection leads to more effective task disambiguation in comparison to approaches relying on reasoning solely within the space of questions.

1 INTRODUCTION

LLMs face ambiguously specified tasks in real-world use, where inferred preferences can miss users’ intentions. The paper proposes selecting clarifying questions through explicit reasoning about viable solutions and reports improved disambiguation over implicit-reasoning baselines.

1 INTRODUCTION

  • Natural-language ambiguity and deliberate underspecification mean real-world tasks may be ill-defined, risking misinterpretation of the problem setter’s intentions.The introduction contrasts this with benchmarks that typically assume complete, unambiguous problem statements.Source: §1 · p. 1
  • When user intentions differ from the average preferences encoded in an LLM, its output may not meet the user’s needs, with particular risks in safety-critical decisions.The introduction names medical diagnosis and treatment as examples.Source: §1 · p. 1
  • Clarifying questions can narrow viable solutions and reduce unsatisfactory responses, but choosing a good question requires reasoning about the model’s solution distribution.The paper hypothesizes that this meta-cognitive ability may exceed many modern LLMs’ capabilities.Source: §1 · p. 2
  • The proposed BED approach estimates question utility through information gain, shifting question selection from implicit reasoning to explicit sampling from the solution space.The paper frames this as an approach to efficient information acquisition for LLM tasks expressed in natural language.Source: §1 · p. 2
  • The BED-based strategy improves on baseline strategies that rely on implicit reasoning about which clarifying question to ask.The paper evaluates the strategies on an illustrative game of 20 questions and a real-world code-generation application.Source: §1 · p. 2

2 FORMALISM & BACKGROUND

The section defines task ambiguity as a mismatch between requirement-satisfying and acceptable solutions, then frames clarification as selecting questions that reduce uncertainty over viable solutions. Under a uniform solution prior, questions with balanced answer partitions maximize expected information gain.

2 FORMALISM & BACKGROUND

  • Task ambiguity means the solutions satisfying stated requirements outnumber the acceptable solutions; it differs from model uncertainty, which can be high even for an unambiguous task.The paper defines H as the set satisfying requirements and calls a task ambiguous when H is a proper superset of the ground-truth acceptable set.Source: §2 · pp. 2–3
  • Preference optimization aligns outputs with population preferences, so individual users whose intentions differ from the average may receive unacceptable solutions.The paper contrasts this approach with interactive elicitation, which seeks additional specifications from the user.Source: §2 · p. 3, Figure 2 · p. 3
  • Interactive elicitation adds each question and oracle answer to the task specification, progressively shrinking compatible solutions toward the true set without fine-tuning.The paper seeks a question strategy that concentrates the agent’s solution distribution around the acceptable set with minimal interaction cost.Source: §2 · p. 3

2.1 WHAT MAKES A QUESTION INFORMATIVE?

  • Bayesian Experimental Design evaluates a question by the expected reduction in entropy over possible solutions after the answer is received.Each question induces answer-specific subsets of compatible solutions, which determine the remaining uncertainty.Source: §2.1 · pp. 3–4
  • The strategy queries requirements as common grounding between unknown setter preferences and the agent’s potentially biased distribution, and approximates the unknown distribution as uniform over compatible solutions.It assumes question-answer pairs add requirements and reduce the compatible solution set.Source: §2.1 · p. 4
  • The analysis assumes truthful oracle answers and the law of total probability; human errors or “I don’t know” answers are treated as unmodeled random noise.The authors note that alternatives could penalize questions expected to elicit uncertain answers.Source: §2.1 · p. 4
  • Under a uniform distribution over compatible solutions, expected information gain is maximized by questions whose answers partition that set into equally sized subsets.For binary questions, this means splitting the set in half; questions with more answers can yield greater information gain, but may impose greater mental load on users.Source: §2.1 · p. 5, Figure 3 · p. 5
  • Questions with more possible answers can yield greater information gain, but balanced questions with few answers trade off information against the user’s mental load.Enumerating solutions can produce highly informative questions, but requires users to examine each possible solution.Source: §2.1 · p. 5

3 THE METHOD: ACTIVE TASK DISAMBIGUATION

Active task disambiguation explicitly scores candidate questions by estimated information gain, using sampled solutions and their pseudo-answers to select a question. Its utility accounts for answer cost, while the approach depends on representative solution samples and accurate requirement checks.

3 THE METHOD: ACTIVE TASK DISAMBIGUATION

  • The method selects the candidate question with the highest estimated utility, explicitly evaluating questions against self-generated solutions rather than relying on implicit question-space reasoning.The authors frame this as bootstrapping question generation with the LLM’s potentially stronger solution-generation ability.Source: §3 · pp. 5–6
  • Question utility is defined to reflect estimated information gain while accounting for the cost of obtaining an oracle answer.The cost function assigns an expected answer cost to a candidate question.Source: §3 · p. 6
  • At each iteration, the procedure samples candidate solutions and questions, obtains pseudo-answers, estimates question utility, then records the user’s answer and updates the task specification.Pseudo-answers may be generated by the problem-solving LLM or evaluated with an external tool, depending on the solution type.Source: §3 · p. 6
  • The estimated information gain and resulting question selection directly depend on the sampled solutions.The authors encourage diversity to make samples closer to uniform over requirement-compatible solutions, using increased temperature or instructions to generate diverse, representative solutions.Source: §3 · p. 6
  • The approach requires accurate checks that generated solutions satisfy requirements and that pseudo-answers are compatible with those solutions; hallucinations make these checks challenging without external grounding.The authors report that refined prompting, such as chain-of-thought or self-reflection, works well in practice.Source: §3 · p. 6

4 EXPERIMENTS

The experiments test whether selecting questions by reasoning over possible solutions improves task disambiguation, and whether that advantage depends on solution diversity and evaluation noise. They evaluate iterative yes-or-no questioning in an animal-guessing game and active code generation using unit tests.

4 EXPERIMENTS

  • The experiments test whether explicit reasoning through sampled solutions improves question selection, especially when samples cover the compatible solution space and EIG evaluation noise is minimal.The study contrasts a setting without an external evaluator with one where a tool grounds requirement compatibility.Source: §4 · p. 7

4.1.1 THE MAIN EXPERIMENT

  • EIG-uniform outperforms implicit and implicit-ToT questioning in the animal game; its advantage over those strategies is smaller for GPT-4o-mini than GPT-3.5-turbo.EIG-logprobs significantly underperforms EIG-uniform, consistent with the LLM’s generative distribution favoring solutions that do not necessarily match the ground-truth solution set.Source: §4.1.1 · p. 8, Figure 4 · p. 8

4.1.2 ADDITIONAL STUDIES

  • Across the first six iterations, EIG-uniform questions leave smaller subsets of compatible animals, and estimates from 20 samples proxy ground-truth EIG approximated using 500 animals.The 20 Questions experiment uses 15 arbitrary animal tasks, 10 iterations, and no fixed candidate-animal list.Source: §4.1.1 · p. 7, §4.1.2 · p. 8, Figure 5 · p. 8
  • Adding a diversifying statement to the animal-sampling prompt leads to more diverse samples, as assessed using conditional mutual information across 330 questions.The study varies sample count and compares sampling with and without the statement; Table 1 reports the diversity results.Source: §4.1.2 · pp. 8–9, Table 1 · p. 9
  • As requirements accumulate, the LLM becomes more prone to hallucinating incompatible animal solutions; filtering rejected samples improves the accuracy of the solutions used.Figure 6 compares rejections after one filtering step and EIG-uniform rankings with zero, one, or two filtering steps.Source: §4.1.2 · p. 9, Figure 6 · p. 9

4.2 ACTIVE CODE GENERATION–REQUIREMENTS AS UNIT TESTS

  • The code-generation experiment uses an external tool to evaluate requirement compatibility, with generated test cases or function inputs serving as questions and answers becoming executable unit tests.The two question types are binary assertions answered True or False, and open inputs answered with expected outputs.Source: §4.2 · p. 9
  • Across HumanEval and APPS, EIG-based strategies yield higher code-sample accuracy with fewer queried test cases; open questions can be more informative than binary questions but may increase human answering effort.The reported Figure 7 accuracy averages 48 tasks, three run seeds, and three evaluation seeds after four elicited requirements.Source: §4.2 · pp. 9–10, Figure 7 · p. 10

5 DISCUSSION

Clarifying questions selected by direct utility estimation are reported as more efficient than those generated through zero-shot prompting. The work focuses on requirement elicitation, while ambiguity detection and deciding when to stop querying remain separate challenges.

5 DISCUSSION

  • Zero-shot clarifying questions are less efficient than questions selected by direct utility estimation over self-generated solutions.The authors suggest current LLM skills in generating efficient clarifying questions are underdeveloped.Source: §5 · p. 10
  • The study focuses on requirement elicitation, leaving ambiguity detection and deciding when to stop querying as additional challenges.The authors identify both as equally important aspects of handling ambiguously specified tasks.Source: §5 · p. 10
  • The question-generating strategies require more LLM calls than baselines, while the authors assume query-selection computation is negligible relative to the value of reducing ambiguity.They expect this assumption to become more valid as LLM token-generation costs decline.Source: §5 · p. 10
  • GPT-4o-mini zero-shot accuracy is ~35% on APPS versus ~70% on HumanEval, consistent with APPS being more challenging.The experiments filter the two benchmarks to 48 non-trivial tasks that do not approach 100% zero-shot accuracy.Source: §5 · p. 10

REPRODUCIBILITY

The appendix records experiment-reproduction resources and situates active task disambiguation among related approaches. Its distinguishing setting has the agent generate questions over unconstrained natural-language solutions rather than select from fixed pools or feature spaces.

REPRODUCIBILITY

  • Reproduction materials include prompts, a pseudo-algorithm, results for four models, and code for the Section 4.2 experiments.The models are GPT-3.5-turbo, GPT-4o-mini, Llama-3-8B (Instruct), and Llama-3-70B (Instruct).Source: REPRODUCIBILITY · p. 11

A EXTENDED RELATED WORK

  • Unlike conventional active learning, this setup has the agent generate questions and elicit binary requirements to filter its own solutions.The authors describe active task elicitation as a generalization of active learning, while noting that distribution shifts after added requirements may not align with the ground truth.Source: A EXTENDED RELATED WORK · p. 16
  • Prior clarification work includes trained question generators and zero-shot prompting; this work encourages explicit inference-time reasoning over viable outputs.The cited approaches include Rao & Daumé III (2018; 2019), Min et al. (2020), Krasheninnikov et al. (2022), Kuhn et al. (2022), and Li et al. (2023).Source: A EXTENDED RELATED WORK · p. 16
  • Unlike preference elicitation over predetermined objects or fixed feature spaces, this setting iteratively samples solutions from unconstrained natural language.This allows questions about hypothetical solutions at varying levels of granularity, including aspects difficult to anticipate before dialogue.Source: A EXTENDED RELATED WORK · p. 16
  • Related approaches include uncertainty-driven information seeking, active acquisition with costs, and self-improvement rewards for useful clarifying questions.The cited methods are UoT, Astorga et al. (2024a), and STaR-GATE (Andukuri et al., 2024).Source: A EXTENDED RELATED WORK · pp. 16–17

B POTENTIAL APPLICATIONS OF ACTIVE TASK DISAMBIGUATION

The paper envisions active task disambiguation in educational, software, creative, scientific, robotic, and healthcare settings. It also lists possible extensions across additional decision-making domains.

B POTENTIAL APPLICATIONS OF ACTIVE TASK DISAMBIGUATION

  • In tutoring, software engineering, and creative work, clarifying questions may help surface learning objectives, reduce requirement misinterpretation, and align generated outputs with evolving user intentions.The passage connects these examples to personalized education, software development, and content generation, design, and games.Source: B POTENTIAL APPLICATIONS OF ACTIVE… · p. 17
  • The authors envision applications in research, robotics, and healthcare, alongside legal reasoning, financial advice, emergency planning, policy support, and cybersecurity.Examples include clarifying experimental constraints, mission goals, diagnostic criteria, and treatment protocols.Source: B POTENTIAL APPLICATIONS OF ACTIVE… · p. 17

C ACTIVE TASK DISAMBIGUATION

The appendix sketches an iterative requirement-elicitation procedure and supplies prompt templates for the 20 Questions game and coding tasks. These templates instantiate game questions as yes/no queries and coding questions as test cases or sample inputs.

C ACTIVE TASK DISAMBIGUATION

  • The elicitation loop samples candidate questions and solutions, estimates question utility, obtains an answer, converts it into a requirement, and updates the problem statement.The procedure runs over a specified number of iterations; the methods for obtaining model distributions and transforming answers are application-specific.Source: C ACTIVE TASK DISAMBIGUATION · p. 18

D.1.1 PROMPT TEMPLATES

  • For 20 Questions, prompts generate diverse animal solutions and informative yes/no questions, then transform question-answer pairs into affirmative requirements.A separate prompt instructs an oracle to answer Yes, No, or Pass for a given animal and question.Source: §D.1.1 · p. 19

D.2.1 PROMPT TEMPLATES

  • For coding tasks, prompts elicit binary assertion tests or open-ended sample inputs, with separate instructions for generating the solution function.The binary format requests assertions; the open format requests function calls for expected outputs.Source: §D.2.1 · pp. 20–21

D.1.2 EXPERIMENTAL DETAILS

The main setup tests elicited requirements using 15 ground-truth animals. A separate study constructs a diverse 500-animal list to examine question partitioning across a larger solution sample.

D.1.2 EXPERIMENTAL DETAILS

  • The main setup uses 15 animals as ground-truth solutions, running 10 iterations with five seeds and generating 20 solutions and five candidate questions.Source: §D.1.2 · p. 20
  • For the larger partitioning study, the authors construct a diverse list of 500 animals stratified across six animal categories.The list contains 100 mammals, 100 birds, 100 invertebrates, 50 amphibians, 50 reptiles, and 100 fish.Source: §D.1.2 · p. 20
  • The 500-animal list is formed by subsampling category-specific animals generated in the main experiments through iteration six.GPT-4o-mini was prompted to stratify the animal kingdom into categories; the authors use the resulting stratification to ensure diversity.Source: §D.1.2 · p. 20

D.2.2 EXPERIMENTAL DETAILS

The code-generation experiments use filtered HumanEval and APPS tasks to ensure the prompts contain ambiguity relevant to eliciting additional requirements.

D.2.2 EXPERIMENTAL DETAILS

  • The experiments use HumanEval and APPS, retaining tasks that cannot be solved with 100% zero-shot accuracy without additional requirements.Source: §D.2.2 · p. 21
  • APPS tasks are further filtered to exclude prompts containing example input-output pairs.Source: §D.2.2 · p. 21
  • The filtered datasets contain 48 HumanEval tasks and 47 APPS tasks.Source: §D.2.2 · p. 21

E ADDITIONAL RESULTS

The section points readers to Table 2 for the final ranking results in the game of 20 questions.

E ADDITIONAL RESULTS

  • Table 2 presents the final ranking results for the game of 20 questions.Source: §E.1 · p. 21

E.2.1 FULL RESULTS

Across HumanEval and APPS, EIG-based strategies outperform their non-EIG counterparts, and open questions yield larger information gains than binary questions. The section also discusses ambiguity in the HumanEval evaluation and the increased LLM-call costs of question elicitation.

E.2.1 FULL RESULTS

E.2.2 AMBIGUITY IN HUMANEVAL

  • Low zero-shot accuracy does not establish that a HumanEval prompt is inherently ambiguous; appended examples may bias outputs toward correct solutions, and Table 5 samples ambiguous prompts.Source: §E.2.2 · pp. 22–23, Table 5 · p. 24

F COSTS OF QUESTION ELICITATION

  • EIG-based question generation uses more LLM calls than zero-shot baselines, a cost that may constrain real-world applications.The authors compare effective calls per interaction for key strategies.Source: F COSTS OF QUESTION ELICITATION · p. 25

F.1 20 QUESTIONS

  • For 20 questions, the baseline uses O(1) calls, implicit-ToT O(M), and EIG O(N + M + NM).EIG adds calls to sample candidate questions and estimate their information gain across sampled solutions.Source: §F.1 · p. 25

F.2 CODE GENERATION

  • For code generation, the baseline uses O(1) calls and EIG O(N + M), because code execution supplies answers without extra LLM calls.Source: §F.2 · p. 25

Related papers

Loading 2502.04485v1…