Source-linked AI summary
Value of Information: A Framework for Human-Agent Communication
Yijiang River Dong, Tiancheng Hu, Zheng Hui, Caiqi Zhang, Ivan Vulić, Andreea Bobu, Nigel Collier
TL;DR
LLM agents face underspecified requests and must decide whether to act or ask for clarification without imposing unnecessary user effort. The paper introduces an inference-time VoI framework that weighs expected utility gain against communication cost, and finds robust performance across diverse tasks without task-specific threshold tuning. Its scope is bounded by predefined actions and questions and a simplified linear communication-cost model.
Problem
LLM agents must balance acting on incomplete user intent against interrupting users, while existing fixed-round and confidence-threshold methods do not explicitly weigh information value against user effort.
Method
The paper uses an inference-time Value of Information policy that estimates expected utility gains from candidate questions and subtracts their communication costs.
Results
Across tasks and communication-cost settings, VoI consistently achieves state-of-the-art utility without task-specific threshold tuning, matching or exceeding manually tuned baselines.
Takeaways & Limitations
The framework provides a parameter-free basis for adaptive communication that balances query ambiguity, task risk, and user cognitive load.
Takeaways & Limitations
The evaluation focuses on deciding when to communicate using predefined actions and clarifying questions, rather than generating open-ended questions.
Abstract
from arXiv · showhide
Large Language Model (LLM) agents deployed for real-world tasks face a fundamental dilemma: user requests are underspecified, yet agents must decide whether to act on incomplete information or interrupt users for clarification. Existing approaches either rely on brittle confidence thresholds that require task-specific tuning, or fail to account for the varying stakes of different decisions. We introduce a decision-theoretic framework that resolves this trade-off through the Value of Information (VoI), enabling agents to dynamically weigh the expected utility gain from asking questions against the cognitive cost imposed on users. Our inference-time method requires no hyperparameter tuning and adapts seamlessly across contexts-from casual games to medical diagnosis. Experiments across four diverse domains (20 Questions, medical diagnosis, flight booking, and e-commerce) show that VoI consistently matches or exceeds the best manually-tuned baselines, achieving up to 1.36 utility points higher in high-cost settings. This work provides a parameter-free framework for adaptive agent communication that explicitly balances task risk, query ambiguity, and user effort.
1 Introduction
Real-world LLM agents must resolve underspecified requests while balancing the risk of acting on incomplete information against the burden of asking users for clarification. The paper proposes Value of Information as a decision-theoretic basis for weighing expected task improvement against communication cost.
- Real-world requests omit latent goals, contexts, and preferences that agents cannot resolve without external input.
- Agents must balance misaligned actions from incomplete information against user frustration and cognitive burden from frequent interruptions.
- Fixed-round and confidence-threshold strategies ignore context-specific information value and require brittle, manually tuned decisions.
- The proposed framework treats clarification as rational action when expected downstream improvement exceeds the user’s time and effort.
- The inference-time VoI method estimates ambiguity, task risk, and cognitive load to dynamically choose between asking and acting.
2 Related Work
Prior work addresses agent execution, proactive clarification, uncertainty gating, reinforcement learning, and pragmatic communication, but these approaches leave gaps in principled, stake-sensitive clarification decisions. The paper situates VoI within RSA-style communication while targeting adaptive downstream decision quality.
- Many agent benchmarks assume complete, unambiguous instructions, emphasizing execution rather than ambiguity resolution before action.
- Prompting can elicit preferences or clarification, but prior work reports that such strategies may remain suboptimal without principled planning or learning.
- Uncertainty-gated methods trigger clarification from confidence or entropy thresholds but may omit the downstream stakes of decisions.
- Reinforcement-learning approaches encourage clarification but are often task-specific and require designed simulations and training pipelines.
- RSA research models communication as rational action, including question selection that trades expected informativeness against asking cost and communication for downstream decisions.
3 Problem Formulation
The paper formulates adaptive communication as a sequential clarify-or-commit process over latent user intent and terminal actions. At each turn, the agent chooses whether another question’s potential uncertainty reduction justifies its user-effort cost.
- Adaptive communication is modeled as sequential decision-making in which an LLM interacts with a user to select an optimal action.
- The agent receives an ambiguous query, represents latent user goals as an unobserved state, and selects among terminal actions.
- At each turn, the agent either asks a question and updates dialogue history or terminates dialogue and commits to an action.
- The clarify-or-commit policy is the central decision, because each question can reduce uncertainty while imposing additional user effort.
- The objective measures committed-action utility against communication cost representing the user’s time and effort.
4 Methods
The method evaluates questions by their expected improvement in downstream decision utility and subtracts communication cost. An inference-time LLM instantiation estimates beliefs and answer probabilities, using constrained closed-ended questions for tractable computation.
- Value of Information Framework: The VoI policy addresses heuristic failures by reasoning about downstream consequences and decision stakes when selecting clarification.
- Value of Information Framework: The agent maintains beliefs over latent intents and chooses the immediate action with maximum expected utility under its current belief.
- Value of Information Framework: For a candidate question, the agent updates beliefs for each possible answer and averages the resulting decision values using answer probabilities.
- Instantiation with LLMs: Closed-ended multiple-choice or yes-no questions constrain the answer space to make expected-value computation tractable.
- Value of Information Framework: VoI is the difference between expected utility after asking and utility from acting immediately, while net VoI subtracts communication cost.
- Value of Information Framework: The agent asks the highest-positive-net-value question and commits when no question’s expected gain exceeds its cost.
- Instantiation with LLMs: At inference time, the LLM estimates belief distributions, simulates one-step user responses, and marginalizes over latent hypotheses.
5 Experimental Setup
The experiments evaluate adaptive clarification against non-adaptive and confidence-based baselines across ambiguous tasks with different stakes, using the proposed VoI algorithm and several LLMs.
- Baselines and method: The study compares No-Question, Fixed-Round, Adaptive Prompting, and Confidence Thresholding baselines against VoI for adaptive communication.The baselines respectively commit immediately, ask a fixed number of questions, rely on heuristic self-assessment, or use a tunable confidence threshold.
- Tasks: Mixed-Stakes 20 Questions tests strategic information gathering in low-stakes animal guessing and high-stakes medical diagnosis variants.The animal task has 100 candidates and utility U = 1 for a correct guess, while medical diagnosis has 15 diseases and utility U = 10 for a correct diagnosis.
- Tasks: Flight Recommendation evaluates preference elicitation by predicting a user’s choice among three held-out flights from five rounds of selection history.Each flight is represented by eight features, including price, stops, and airline.
- Tasks: Ambiguous WebShop removes details from otherwise well-specified shopping instructions, requiring the agent to decide whether to act on partial information.This setting simulates underspecified real-world user queries.
- Baselines and method: VoI generates targeted questions, estimates their expected utility gain, subtracts communication cost, and stops when the best question has non-positive value.After each answer, the agent updates its belief and eventually commits to the action with highest expected utility.
- Models: The evaluation includes GPT-4.1 and Gemini-2.5-Flash as representative leading LLMs.
6 Results
Across the evaluated tasks and communication costs, VoI generally matches or exceeds tuned baselines while adapting question counts to task ambiguity and communication cost. Results also examine baseline behavior, cost sensitivity, and calibration limitations.
- Main results: VoI consistently achieves state-of-the-art utility across tasks and communication-cost settings without task-specific threshold tuning.Figure 2 reports that VoI automatically identifies operating points matching or exceeding the best-tuned baseline in nearly all settings.
- Main results: 14.14 utility in Mixed 20Q at c = 0.01 exceeds the best confidence-thresholding baseline’s 11.49 at τ = 0.90.The result illustrates VoI’s ability to select an effective utility-cost balance dynamically.
- Main results: Near-zero accuracy for No Question on ambiguous Mixed 20Q shows the value of clarification, while VoI stops early when communication costs become prohibitively high.VoI therefore adapts in both underspecified and high-cost conditions.
- Baseline behavior: Fixed-round utility follows an inverted-U pattern because additional questions first help and then communication costs outweigh their information benefits.The optimal question count varies by task and cost.
- Baseline behavior: Confidence thresholding can match VoI when correctly tuned, but its optimal τ is sensitive to each task and cost combination.VoI matches or exceeds this performance without manual tuning.
- Ablation study: VoI matches or exceeds the strongest grid-searched baselines across the Mixed 20-Question cost sweep and adapts smoothly as communication cost changes.The comparison uses four baselines tuned over nine threshold settings.
- Calibration analysis: Calibration is better on Animal Guessing than Medical Diagnosis, where symptom complexity and noise are suspected to reduce calibration quality.Despite this calibration difference, VoI remains empirically effective against tuned baselines.
- Limitations: Improved calibration under missing context is identified as a potential direction for further improving VoI performance.
7 Conclusion
The paper presents VoI as an inference-time framework for deciding when LLM agents should act or ask, balancing query ambiguity, task risk, and user cognitive load. Across diverse domains, the approach consistently outperforms non-adaptive and heuristic baselines without task-specific threshold tuning.
- The VoI agent decides when to act or ask by weighing expected utility gain against communication cost.The method explicitly calculates the expected utility gain of a potential question at inference time.
- Across medical diagnosis and online shopping, VoI consistently outperforms non-adaptive and heuristic-based baselines.
- VoI avoids brittle, task-specific threshold tuning that limits other adaptive communication methods.
- The framework aims to support LLM agents as thoughtful communicators by formally representing when information is valuable.This is presented as a foundation for more aligned, efficient, and collaborative human-AI systems.
Limitations
The evaluation isolates the decision of when to communicate by using predefined actions and clarifying questions, while adopting a simplified linear communication-cost model. Extending the framework to open-ended question generation and richer cognitive-cost models remains future work.
- The study evaluates communication selection with predefined actions and clarifying questions rather than open-ended question generation.This controls for question-generation quality and isolates the VoI-based selection policy.
- The framework’s open-ended dialogue extension remains an important next step.The authors present the current selection principle as a foundation for more sophisticated generative components.
- The model uses a linear communication cost, c(H) = T · c, as a simplified first-order approximation of cognitive load.The authors note that richer cognitive-cost models require further empirical user research.
- The VoI decision rule can incorporate more sophisticated cost functions as they are developed.
Ethical Considerations
The paper emphasizes that adaptive communication must preserve user agency while accounting for burden, trust, frustration, and privacy. Questions should not impose unnecessary interruptions or elicit sensitive information without appropriate safeguards.
- Users should retain the ability to decline questions or proceed without clarification according to their own judgment.
- Frequent or intrusive clarifying questions can increase cognitive burden, erode trust, and cause frustration.The paper therefore calls for cost models that penalize unnecessary interruptions.
- Reducing ambiguity may require eliciting sensitive information such as medical symptoms or personal preferences.
- Future implementations should incorporate data minimization principles and privacy safeguards.These safeguards are intended to prevent uncertainty reduction from compromising user privacy or comfort.
A Case Study: VoI is Risk-Aware
The case study shows that VoI adapts questioning to task stakes, unlike confidence-based methods that primarily track uncertainty. It asks more in high-stakes diagnosis and stops earlier in low-stakes guessing when further information is not worth its cost.
- Risk-aware communication: The comparison uses a low-stakes animal-guessing task with reward=1 and a high-stakes medical-diagnosis task with reward=10, holding communication cost at c = 0.05.
- High-stakes task: In medical diagnosis, VoI continues asking moderate-information questions because their expected value outweighs the communication cost.It continues several rounds after confidence thresholding stops, while ambiguity remains and the baseline reaches an incorrect diagnosis.
- Low-stakes task: In animal guessing, VoI stops earlier because the potential utility gain from additional questions does not justify cumulative communication cost.This avoids unnecessary cognitive load in the low-risk task.
- Risk-aware principle: VoI unifies uncertainty reduction with stake-weighted utility, whereas confidence-based methods address only uncertainty.Its decision principle values information according to its expected impact on the final task utility.
B Main Results in Tables
This section presents figures and tables comparing communication strategies across Mixed 20 Questions, medical diagnosis, flight recommendation, and Webshop tasks. The methods include direct prompting, auto-stop, confidence thresholding, and VoI-based question generation or batch answer simulation.
- The broader result displays compare methods and thresholds across Mixed 20Q, Flight Recommendation, and Webshop, with utility accounting for language-model performance and question costs.
- The figures cover Animal 20 Questions and Medical Diagnosis, including direct prompting, auto-stop, confidence thresholding, and two VoI procedures.
- The results are organized around comparisons among direct prompting, auto-stop, confidence thresholding, and VoI variants.
- Additional visualizations examine airline preference matching through prior estimation, posterior estimation with options, and candidate preference questions.