Task-Level VOI Score

• Task-level VOI is a decision-theoretic score that measures the expected improvement in optimal utility after acquiring clarifying information.
• It computes net gain by comparing downstream action utility before and after potential information acquisition while accounting for explicit query costs.
• This framework enhances human–agent communication by replacing fixed uncertainty thresholds with cost-sensitive, adaptive query policies.

Task-level Value of Information (VOI) is a decision-theoretic score assigned to a candidate information-acquisition step—such as a clarification question, measurement, sensing action, inspection, or transmission—by comparing the expected downstream value before and after acquiring that information, usually net of an explicit acquisition cost. In recent work on human–agent communication, the score is used to resolve whether a LLM agent should clarify an underspecified request or commit immediately, thereby balancing task risk, query ambiguity, and user effort without manual threshold tuning (Dong et al., 10 Jan 2026). Across adjacent literatures, closely related quantities appear as Expected Value of (Imperfect) Information in Bayesian decision analysis, as a closed-loop versus open-loop value gap in POMDP planning, and as mutual information in hidden Markov models, indicating that the term denotes a family of formally related but task-specific constructs rather than a single universal formula (Langtry et al., 2024, Laouar et al., 1 Apr 2026, Wang et al., 2021).

1. Decision-theoretic foundations

In the human–agent communication formulation, the latent user condition is represented by a finite state space $\Theta$, with current belief $b(\theta)$ over the true state $\theta$, terminal actions $a \in A$, task utility $U(\theta,a)$, and question cost $c(q)$. The expected utility of taking action $a$ under the current belief is

$$E[U(a)\mid b] = \sum_{\theta\in\Theta} b(\theta)\,U(\theta,a),$$

and the best immediate action is

$a^*=\arg\max_{a\in A} E[U(a)\mid b].$

Its value is

$$V(b)=\max_{a\in A}\sum_{\theta} b(\theta)\,U(\theta,a).$$

A candidate question $b(\theta)$0 with answer space $b(\theta)$1 induces updated beliefs $b(\theta)$2 and a pre-posterior value

$b(\theta)$3

with

$b(\theta)$4

The net task-level score is then

$b(\theta)$5

The resulting “clarify-or-commit” policy asks $b(\theta)$6 only if $b(\theta)$7; otherwise it commits to the current best terminal action (Dong et al., 10 Jan 2026).

This construction makes the benefit term explicitly downstream and task-dependent. The gain is not “informativeness” in the abstract, but the expected improvement in the best attainable utility after the posterior update. The subtraction of $b(\theta)$8 converts a pure pre-posterior gain into a communication policy, so that a question is warranted only when its expected decision improvement exceeds the user’s cognitive burden. The stated assumptions are finite or discrete $b(\theta)$9 and $\theta$0, closed-ended questions, a linear cost model $\theta$1, and a utility function $\theta$2 that is known or computable (Dong et al., 10 Jan 2026).

2. Human–agent communication as a VOI problem

The central motivation is that LLM agents operating on real-world tasks receive underspecified requests yet must decide whether to act on incomplete information or interrupt for clarification. The proposed framework treats communication itself as a costly decision. Current belief $\theta$3 induces uncertainty over utilities $\theta$4; $\theta$5 captures the best expected utility available immediately, while $\theta$6 captures the expected best utility after branching on possible answers to question $\theta$7. Their difference is the classical value of information, and the additional subtraction of $\theta$8 operationalizes a user-aware clarification policy (Dong et al., 10 Jan 2026).

This formulation differs from brittle confidence-threshold methods because the decision boundary is not an externally tuned threshold on uncertainty. It is instead endogenous to the utility structure and the question cost. A high-stakes task with asymmetric utilities can justify clarification at lower epistemic ambiguity than a low-stakes task, while higher cognitive cost can suppress questioning even under uncertainty. The paper characterizes the procedure as parameter-free and inference-time, precisely because the decision criterion is $\theta$9 rather than a manually set confidence threshold (Dong et al., 10 Jan 2026).

A broader implication, suggested by the formulation, is that ambiguity alone is insufficient to trigger interaction. A question can reduce posterior uncertainty yet still have non-positive net VOI if the answer is unlikely to change the optimal action, if the action utilities are nearly flat, or if the cost of interruption dominates the expected utility gain. This distinguishes VOI-driven communication from uncertainty-minimization heuristics.

3. Estimation and inference-time computation

Practical estimation proceeds by eliciting or simulating the components of the score. The belief $a \in A$0 is obtained by prompting the LLM for a probability distribution over $a \in A$1 given dialogue history $a \in A$2. The answer likelihood $a \in A$3 is estimated by simulating how a user with latent state $a \in A$4 would answer question $a \in A$5, using the LLM on a small sample of $a \in A$6 values. The posterior $a \in A$7 is then recomputed by re-prompting with the augmented history $a \in A$8. Utilities are task-defined—for example, $a \in A$9 for a correct medical diagnosis and $U(\theta,a)$0 otherwise—and the question cost may be a fixed constant or a more sophisticated cognitive-load model (Dong et al., 10 Jan 2026).

The inference-time algorithm initializes history $U(\theta,a)$1 with the user request $U(\theta,a)$2, computes a belief $U(\theta,a)$3, and iterates for at most $U(\theta,a)$4 turns. At each turn it computes the current value

$U(\theta,a)$5

generates a small candidate set of questions $U(\theta,a)$6, and for each $U(\theta,a)$7 marginalizes over simulated answers $U(\theta,a)$8. For each answer it recomputes a belief $U(\theta,a)$9, evaluates the corresponding optimal terminal value, averages these values to obtain $c(q)$0, and sets $c(q)$1. The agent then asks the maximizing question if its VOI is positive; otherwise it breaks and returns the current optimal action (Dong et al., 10 Jan 2026).

The reported per-turn complexity is $c(q)$2 LLM inferences for belief estimation and answer simulation. The paper notes that in practice $c(q)$3 and $c(q)$4 is small, such as yes/no or three to four choices, which makes the procedure real-time feasible. By construction, no manual threshold $c(q)$5 is required. Changing the task utility $c(q)$6 and cost $c(q)$7 changes the behavior directly, without retuning a separate interaction hyperparameter (Dong et al., 10 Jan 2026).

4. Worked example and empirical behavior

A concrete medical-diagnosis example instantiates the framework with $c(q)$8, utility $c(q)$9 for a correct diagnosis and $a$0 for a wrong one, and question cost $a$1. With belief $a$2, the immediate value is

$a$3

For the question “Do you have a fever?” with $a$4, the model estimates $a$5 and $a$6, which implies $a$7 and $a$8. The posterior after “Yes” is proportional to $a$9 and normalizes to $$E[U(a)\mid b] = \sum_{\theta\in\Theta} b(\theta)\,U(\theta,a),$$0, yielding value $$E[U(a)\mid b] = \sum_{\theta\in\Theta} b(\theta)\,U(\theta,a),$$1. The posterior after “No” is proportional to $$E[U(a)\mid b] = \sum_{\theta\in\Theta} b(\theta)\,U(\theta,a),$$2 and normalizes to $$E[U(a)\mid b] = \sum_{\theta\in\Theta} b(\theta)\,U(\theta,a),$$3, yielding value $$E[U(a)\mid b] = \sum_{\theta\in\Theta} b(\theta)\,U(\theta,a),$$4. Hence

$$E[U(a)\mid b] = \sum_{\theta\in\Theta} b(\theta)\,U(\theta,a),$$5

and

$$E[U(a)\mid b] = \sum_{\theta\in\Theta} b(\theta)\,U(\theta,a),$$6

so the question should be asked (Dong et al., 10 Jan 2026).

The empirical study spans four domains: 20 Questions, medical diagnosis, flight booking, and e-commerce. Across these settings, the framework is reported to consistently match or exceed the best manually tuned baselines, with gains of up to $$E[U(a)\mid b] = \sum_{\theta\in\Theta} b(\theta)\,U(\theta,a),$$7 utility points in high-cost settings. The same underlying procedure is stated to adapt from low-stakes games to high-stakes tasks by changing the utility scale and question cost rather than by retuning an interaction threshold (Dong et al., 10 Jan 2026).

A common misunderstanding is that the score is merely a proxy for uncertainty reduction. The worked example shows otherwise: the crucial quantity is the expected increase in optimal downstream utility. The answer “Yes” and “No” both sharpen belief, but their contribution enters only through the value of the best action after each posterior, weighted by answer probability and offset by the communication cost.

5. Related formulations in adjacent domains

In Bayesian decision analysis, the standard Expected Value of (Imperfect) Information compares pre-posterior utility with prior utility. The “On-Policy” extension replaces an intractable exact optimizer by a decision policy $$E[U(a)\mid b] = \sum_{\theta\in\Theta} b(\theta)\,U(\theta,a),$$8 that maps a distribution over uncertain parameters to an action, defining $$E[U(a)\mid b] = \sum_{\theta\in\Theta} b(\theta)\,U(\theta,a),$$9 and estimating it by Monte Carlo over prior samples, measurements, posterior updates, and policy evaluations (Langtry et al., 2024). In the district-energy application used to motivate that extension, load uncertainty was found to affect system operating costs by $a^*=\arg\max_{a\in A} E[U(a)\mid b].$0 and optimal design by $a^*=\arg\max_{a\in A} E[U(a)\mid b].$1, yet building monitoring reduced overall costs by less than $a^*=\arg\max_{a\in A} E[U(a)\mid b].$2 on average and was not economically worthwhile in that case (Langtry et al., 2024).

In finite-horizon POMDP planning, task-level VOI is defined as the value gap between closed-loop planning, which branches on observations, and open-loop planning, which marginalizes them out:

$a^*=\arg\max_{a\in A} E[U(a)\mid b].$3

The VOIMCP algorithm introduces open-loop and closed-loop meta-actions and conditionally prunes observation branches when the expected closed-loop gain is within a $a^*=\arg\max_{a\in A} E[U(a)\mid b].$4-fraction of the closed-loop value. The paper provides a bounded-regret theorem for the adaptive VOI operator and polynomial convergence results for VOIMCP (Laouar et al., 1 Apr 2026).

In hidden Markov models, VOI is formalized differently, as mutual information between the current hidden state and the available observation sequence:

$a^*=\arg\max_{a\in A} E[U(a)\mid b].$5

For the noisy Ornstein–Uhlenbeck process, the framework gives closed-form expressions, high- and low-SNR approximations, worst-case statistics under Poisson sampling with FCFS M/M/1 delivery, and a threshold sampling policy that maximizes average VOI under a rate constraint (Wang et al., 2021).

In delayed LQG networked control, the one-step delay-dependent VoI for loop $a^*=\arg\max_{a\in A} E[U(a)\mid b].$6 reduces in practice to the proxy

$a^*=\arg\max_{a\in A} E[U(a)\mid b].$7

which is used to rank transmissions when only a subset of loops can access the channel. The paper states that this scheduling policy outperformed periodical triggering and an Age-of-Information-based policy under transmission delay (Wang et al., 2021). In network inspection of binary components, the local score for inspecting component $a^*=\arg\max_{a\in A} E[U(a)\mid b].$8 is

$a^*=\arg\max_{a\in A} E[U(a)\mid b].$9

with net VOI obtained by subtracting inspection cost $$V(b)=\max_{a\in A}\sum_{\theta} b(\theta)\,U(\theta,a).$$0. For a single component, evaluation is $$V(b)=\max_{a\in A}\sum_{\theta} b(\theta)\,U(\theta,a).$$1 once $$V(b)=\max_{a\in A}\sum_{\theta} b(\theta)\,U(\theta,a).$$2 is known, while a heuristic that fixes the other components’ actions at prior optima reduces ranking all $$V(b)=\max_{a\in A}\sum_{\theta} b(\theta)\,U(\theta,a).$$3 components to $$V(b)=\max_{a\in A}\sum_{\theta} b(\theta)\,U(\theta,a).$$4 (Lin et al., 2021).

These variants share a common structure—comparing decision quality with and without candidate information—but differ in what counts as “value”: expected utility improvement, closed-loop planning gain, or uncertainty reduction measured by mutual information.

6. Computation, bounds, and interpretive issues

Several computational strategies have been proposed when exact VOI evaluation is expensive. In collaborative human–machine sensing, deep convolutional networks were trained on image encodings of belief states and candidate queries to estimate discretized VOI bins. In the reported toy problem, CNN VOI selection was approximately $$V(b)=\max_{a\in A}\sum_{\theta} b(\theta)\,U(\theta,a).$$5 correct, average posterior entropy was $$V(b)=\max_{a\in A}\sum_{\theta} b(\theta)\,U(\theta,a).$$6–$$V(b)=\max_{a\in A}\sum_{\theta} b(\theta)\,U(\theta,a).$$7 lower than an AMDP baseline, and the MAP-error distribution was more concentrated near zero; in the indoor semantic-search task, VOI-selection accuracy was $$V(b)=\max_{a\in A}\sum_{\theta} b(\theta)\,U(\theta,a).$$8, and 16 candidate queries were evaluated in much less than $$V(b)=\max_{a\in A}\sum_{\theta} b(\theta)\,U(\theta,a).$$9 ms at runtime (Lore et al., 2015). In rational anytime estimation for measurement selection, each candidate VOI $b(\theta)$00 is represented by a Gaussian belief $b(\theta)$01 with uncertainty update $b(\theta)$02, and exact recomputation is triggered only when the expected gain $b(\theta)$03 exceeds the recomputation cost $b(\theta)$04 (Tolpin et al., 2010).

For long-horizon partially observable deterioration problems, structural-health formulations distinguish step-wise VOI from life-cycle VOI inside a POMDP. There, the step-wise pre-cost quantity is bounded by

$b(\theta)$05

and the paper emphasizes that a POMDP policy inherently leverages net VOI to guide observational actions at every decision step (Andriotis et al., 2019). This is consistent with a broader point visible across the literature: pre-cost VOI and net VOI are different objects. Pre-cost information value can be nonnegative under an optimal policy, while the operational decision to inspect, query, monitor, or transmit depends on whether that gain exceeds the relevant cost term.

This distinction resolves a common misconception that more information is always worth obtaining. In the human–agent communication framework, the agent asks only when $b(\theta)$06 and otherwise commits (Dong et al., 10 Jan 2026). In the district-energy case, monitoring reduced cost by less than $b(\theta)$07 on average and still failed the economic test because that benefit was below the measurement cost (Langtry et al., 2024). A plausible implication is that task-level VOI is most informative when it is treated not as a generic measure of informativeness, but as a cost-sensitive decision score tied to a specific action space, belief update, and utility model.