Information-Consistency Plan Selection

• The topic defines information-consistency-based plan selection as choosing plans that remain coherent under uncertainty and incomplete world descriptions, ensuring broad reusability.
• It leverages diverse methodologies such as evidential support in possible-world planning, conflict-free probabilistic argumentation, and credal robustness in Quasi-Bayesian frameworks.
• Practical applications include autonomous driving, LLM web agents, and model selection, highlighting strategies for robust decision-making in uncertain environments.

Information-consistency-based plan selection is an Editor’s term for a family of planning and decision procedures in which a candidate plan is preferred, retained, or executed because it remains coherent with the available information under uncertainty, incompleteness, stochasticity, or model ambiguity. Across the literature, the operative notion of consistency varies: a plan may be consistent with multiple possible worlds represented by evidential support, with a conflict-free set of structured arguments, with an entire credal set of probability distributions, with KL-bounded deviations from a prior policy and a Bayesian model posterior, or with repeated-run robustness metrics that distinguish one-off success from repeatable task completion (Mansell, 2013, Morveli-Espinoza et al., 2020, Cozman et al., 2013, Grau-Moya et al., 2016, Zambrano et al., 28 May 2026). In each case, plan selection is not treated as the recovery of a uniquely “true” plan; rather, the selected plan is one that remains usable, justified, or stable relative to the informational structure imposed by the method.

1. Conceptual scope and recurrent selection logic

The common planning setting is one in which the agent does not possess a complete, static, and unambiguous description of the world. In U-Plan, the agent has only an incomplete and uncertain description of the initial state and therefore cannot construct one exact initial state (Mansell, 2013). In structured-argument approaches, incompatible procedural goals must be selected under uncertainty in beliefs, actions, and intermediate goals (Morveli-Espinoza et al., 2020). In Quasi-Bayesian planning, subjective assessments are imprecise and are represented by a nonempty convex set of probability distributions rather than a single prior (Cozman et al., 2013). In KL-regularized MDP planning, both action selection and transition modeling are constrained relative to reference distributions (Grau-Moya et al., 2016). In LLM web agents, the relevant uncertainty is execution stochasticity and the brittleness of plan representations in interactive web environments (Zambrano et al., 28 May 2026).

These formulations differ in ontology and formalism, but they share a recurring selection logic. First, the planner represents uncertainty explicitly: possible worlds, probabilistic intervals, credal sets, posterior model distributions, or repeated-run outcomes. Second, candidate plans are evaluated not only by task achievement but also by whether they remain valid, admissible, or reliable across that uncertainty set. Third, information acquisition, further search, or stronger penalties are introduced only when the current informational state is insufficient to discriminate among alternatives. This suggests that information-consistency-based plan selection is less a single algorithm than a selection principle: plans are filtered by their compatibility with structured uncertainty before or alongside ordinary utility or reward optimization.

A related but broader analogue appears in information-criterion work on model selection. PanIC defines consistency through penalized empirical risk minimization, with a penalty that vanishes yet separates larger models sufficiently fast (Nguyen, 2023). In exploratory factor analysis, the BIC result is framed as selecting the smallest candidate factor order among the globally KL-optimal covariance classes (Nguyen et al., 9 Apr 2026). A plausible implication is that the planning literature and the information-criterion literature share a common architecture: define the correct selection target under misspecification or uncertainty, quantify the fluctuation scale of near-equivalent candidates, and then use a selection rule whose discrimination power is calibrated to that scale.

2. Possible-world consistency in hierarchical planning

U-Plan provides one of the clearest explicit mechanisms for plan selection by consistency across informational alternatives. The planning problem is defined by an uncertain and incomplete initial description, so the system constructs multiple possible initial states, or P-states, each of which is a complete description of one possible world using propositional statements (Mansell, 2013). Each P-state is represented hierarchically at multiple abstraction levels,

$ps(a)=\{l_1(a), \ldots, l_n(a)\},$

and evidence over world descriptions is represented using Dempster-Shafer support and plausibility intervals

$[\mathrm{Spt}(A_j), \mathrm{Pls}(A_j)],$

with

$$\mathrm{Spt}(A_j)=\sum_{A_i \subseteq A_j} m(A_i), \qquad \mathrm{Pls}(A_j)=1-\mathrm{Spt}(-A_j).$$

Selection begins by choosing the P-state with the greatest support at the highest level of abstraction, then descending to the child with greatest support until a fully specified P-state is obtained (Mansell, 2013). The rationale given is that planning first for the world with the greatest evidential support increases the chance that the resulting plan will be reusable for other worlds. This already makes plan selection information-relative: the first plan is not merely a solution to one world, but a candidate for broader reapplication within the uncertainty set.

The key consistency test is reapplication. After a plan is built for the most-supported P-state, U-Plan attempts to apply that same plan to other P-states. A plan is reapplicable if all nonredundant reduction operators in the plan have their preconditions satisfied in the new P-state and applying the operators in sequence reaches the goal (Mansell, 2013). Three outcomes are then distinguished: full reuse, partial reuse, and rejection. Full reuse means the entire plan remains valid in another possible world. Partial reuse means that the valid prefix is preserved and planning resumes from the failure point. Rejection occurs when the operator cannot be made to work and planfail/backtracking is used.

This mechanism operationalizes consistency as cross-world operator validity. The system does not require exact world knowledge before planning; instead, it tests whether a plan remains coherent as the world description varies over P-states consistent with the initial evidence. The abstraction hierarchy strengthens this logic. Strategic decisions are made at the highest abstraction level, where world properties are described as more stable and less sensitive to uncertainty, while tactical details are refined lower in the hierarchy (Mansell, 2013). That design implies that early plan commitments are chosen where consistency across worlds is most likely, and lower-level divergence is deferred until additional detail becomes necessary.

The super-plan extends the same principle to plan families. Once plans exist for relevant P-states, common initial actions are merged, and branches are introduced only where plans diverge. At branch points, a knowledge acquisition operator may be inserted to acquire the information needed to choose the correct next action. If information is not available, branch choice is guided by Dempster-Shafer mass: $m(p_i)=\sum_{Y \in C_0} m(Y),$ where $C_0$ is the set of initial P-states that use $p_i$ as their plan (Mansell, 2013). Thus plan selection is evidence-driven at both the whole-plan and branch levels, while information gathering is delayed until it becomes decision-relevant.

3. Compatibility, attack, and consistency in probabilistic argumentation

A second line of work treats plans as structured probabilistic arguments and selects them by abstract argumentation semantics. In this framework, a procedural goal is an achievement goal for which there is a set of plans that can accomplish it, but several such goals may be incompatible because their plans interfere (Morveli-Espinoza et al., 2020). The paper distinguishes three forms of incompatibility from Castelfranchi and Paglieri: terminal incompatibility, instrumental or resource incompatibility, and superfluity. These are not merely descriptive labels; they induce different attack relations among plan-arguments.

Plans are represented by probabilistic plan rules of the form

$(\psi \mid \phi)[l,u],$

where the body $\phi$ is a conjunction of beliefs, goals, and actions, and the head $\psi=g\in G$ is a goal (Morveli-Espinoza et al., 2020). Arguments are finite trees whose leaves are elementary probabilistic arguments, and uncertainty is propagated bottom-up by probabilistic modus ponens: $$\{(\psi\mid\phi)[l,u],(\phi\mid T)[l',u']\}\vdash_P (\psi\mid T)[l*l',\,1-l'+u*l'].$$ The root interval $[\mathrm{Spt}(A_j), \mathrm{Pls}(A_j)],$0 then summarizes the uncertainty inherited from the premises.

Selection depends on how argument strength is measured. The paper proposes a logical strength vector

$[\mathrm{Spt}(A_j), \mathrm{Pls}(A_j)],$1

with

$[\mathrm{Spt}(A_j), \mathrm{Pls}(A_j)],$2

Here $[\mathrm{Spt}(A_j), \mathrm{Pls}(A_j)],$3 measures interval precision, $[\mathrm{Spt}(A_j), \mathrm{Pls}(A_j)],$4 interval location, and $[\mathrm{Spt}(A_j), \mathrm{Pls}(A_j)],$5 their combination (Morveli-Espinoza et al., 2020). The paper also defines a utility-based strength for resource incompatibility,

$[\mathrm{Spt}(A_j), \mathrm{Pls}(A_j)],$6

with

$[\mathrm{Spt}(A_j), \mathrm{Pls}(A_j)],$7

Attacks become successful only when one argument is preferred over another according to these strength measures. The framework is

$[\mathrm{Spt}(A_j), \mathrm{Pls}(A_j)],$8

where $[\mathrm{Spt}(A_j), \mathrm{Pls}(A_j)],$9, and then a modified relation $$\mathrm{Spt}(A_j)=\sum_{A_i \subseteq A_j} m(A_i), \qquad \mathrm{Pls}(A_j)=1-\mathrm{Spt}(-A_j).$$0 retains only successful attacks (Morveli-Espinoza et al., 2020). Conflict-free sets $$\mathrm{Spt}(A_j)=\sum_{A_i \subseteq A_j} m(A_i), \qquad \mathrm{Pls}(A_j)=1-\mathrm{Spt}(-A_j).$$1 satisfy

$$\mathrm{Spt}(A_j)=\sum_{A_i \subseteq A_j} m(A_i), \qquad \mathrm{Pls}(A_j)=1-\mathrm{Spt}(-A_j).$$2

Two additional operators refine selection: $$\mathrm{Spt}(A_j)=\sum_{A_i \subseteq A_j} m(A_i), \qquad \mathrm{Pls}(A_j)=1-\mathrm{Spt}(-A_j).$$3 chooses maximal conflict-free sets that achieve the greatest number of pursuable goals, and $$\mathrm{Spt}(A_j)=\sum_{A_i \subseteq A_j} m(A_i), \qquad \mathrm{Pls}(A_j)=1-\mathrm{Spt}(-A_j).$$4 then selects those with maximum total goal preference. Finally, $$\mathrm{Spt}(A_j)=\sum_{A_i \subseteq A_j} m(A_i), \qquad \mathrm{Pls}(A_j)=1-\mathrm{Spt}(-A_j).$$5 projects each selected extension onto the supported compatible goals.

In this literature, information consistency is formalized as argument-theoretic coherence. Theorem 1 establishes direct consistency: in every conflict-free extension, the sets of beliefs, actions, and goals extracted from the extension are consistent, and the goal set contains no superfluous conflicting goals (Morveli-Espinoza et al., 2020). Closure is also required: $$\mathrm{Spt}(A_j)=\sum_{A_i \subseteq A_j} m(A_i), \qquad \mathrm{Pls}(A_j)=1-\mathrm{Spt}(-A_j).$$6 Together with a result invoked from Caminada et al., this yields indirect consistency of the closure of each extension and of the global output. The selected plans are therefore not just preferred; they are justified by a semantics that excludes mutually attacking or redundant goal-plan structures. A plausible implication is that this approach generalizes cross-world consistency into a more abstract notion of inferential compatibility.

4. Robust admissibility under imprecise probabilities

Quasi-Bayesian planning develops a different notion of information consistency, grounded in imprecise probabilities and robustness rather than conflict-freeness or possible-world reuse. Beliefs are represented by a nonempty convex set of probability distributions, a credal set, rather than a single distribution (Cozman et al., 2013). Lower and upper probabilities are

$$\mathrm{Spt}(A_j)=\sum_{A_i \subseteq A_j} m(A_i), \qquad \mathrm{Pls}(A_j)=1-\mathrm{Spt}(-A_j).$$7

Preference between plans is defined through expected loss over the entire credal set: one plan is at least as preferred as another only if its expected loss is no worse for every distribution in the set.

This leads directly to the paper’s main selection concepts: admissibility and robustness. An E-admissible plan is optimal for at least one distribution in the credal set (Cozman et al., 2013). Robustness concerns whether the status of a plan changes under admissible perturbations of the probability model. The proposed strategy is not to search exhaustively for a uniquely optimal plan, but to find an admissible plan quickly and monitor whether it is robust. The paper explicitly recommends that the agent should not insist on resolving every admissible-plan ambiguity.

The planning-to-observe problem makes this concrete. Observations are

$$\mathrm{Spt}(A_j)=\sum_{A_i \subseteq A_j} m(A_i), \qquad \mathrm{Pls}(A_j)=1-\mathrm{Spt}(-A_j).$$8

with observation cost $$\mathrm{Spt}(A_j)=\sum_{A_i \subseteq A_j} m(A_i), \qquad \mathrm{Pls}(A_j)=1-\mathrm{Spt}(-A_j).$$9, and prior beliefs over $m(p_i)=\sum_{Y \in C_0} m(Y),$0 are represented by a convex set of Gaussian priors with uncertain mean: $m(p_i)=\sum_{Y \in C_0} m(Y),$1 If the prior width is $m(p_i)=\sum_{Y \in C_0} m(Y),$2, then after $m(p_i)=\sum_{Y \in C_0} m(Y),$3 observations with sample mean $m(p_i)=\sum_{Y \in C_0} m(Y),$4, the posterior credal set is

$m(p_i)=\sum_{Y \in C_0} m(Y),$5

and the posterior width shrinks as

$m(p_i)=\sum_{Y \in C_0} m(Y),$6

For each individual Gaussian prior, the Bayes-optimal sequential policy partitions the $m(p_i)=\sum_{Y \in C_0} m(Y),$7-plane into Continue, Stop0, and Stop1 regions (Cozman et al., 2013). The Quasi-Bayesian selection rule then tests the entire credal set against these regions. If the whole set lies in Stop0, choose $m(p_i)=\sum_{Y \in C_0} m(Y),$8; if it lies in Stop1, choose $m(p_i)=\sum_{Y \in C_0} m(Y),$9; if it lies in Continue, take another observation; if it intersects more than one region, the situation is non-robust. This is the paper’s central plan-selection criterion: accept a current plan if it is admissible and robust with respect to the credal set; otherwise continue searching or refining if time permits.

Here consistency is setwise rather than pointwise. A plan is selected when all admissible beliefs represented by the credal set support the same region-level action. Non-robustness is not treated as failure of optimization but as a signal that the current informational state does not justify a sharper decision. The paper’s broader conclusion is that robustness replaces exact optimality as the key planning criterion (Cozman et al., 2013). This suggests a strong form of information-consistency-based plan selection: the chosen plan is one whose acceptability survives the full acknowledged imprecision of the belief model.

5. KL-regularized consistency in MDP and trajectory planning

In KL-regularized MDP planning, information consistency is encoded directly in the objective function through divergence constraints. The planner maximizes discounted reward while limiting deviation from a reference policy $C_0$0 and from a Bayesian posterior over latent transition models $C_0$1 (Grau-Moya et al., 2016). The unified objective is

$C_0$2

where $C_0$3 denotes $C_0$4 for $C_0$5 and $C_0$6 for $C_0$7 (Grau-Moya et al., 2016). The corresponding generalized Bellman recursion is

$C_0$8

The selected policy is therefore not simply reward-maximizing. It is the best policy among those that remain sufficiently close to the prior action distribution and to the posterior model, with closeness measured by KL divergence (Grau-Moya et al., 2016). The paper explicitly interprets the two KL terms as information-processing constraints: one on action selection and one on model uncertainty. The plan is “information-consistent” with both priors up to an allowed information budget. In the relevant limit cases, the framework reduces to standard value iteration with a known model, Bayesian MDP planning, or robust planning.

A different but related use of consistency appears in autonomous driving with ConsistencyPlanner. Here the issue is not symbolic compatibility or credal robustness, but real-time generation of safe multimodal trajectories (Zhang et al., 10 Jun 2026). The consistency model is trained so that multiple noisy versions of the same action map to the same clean sample: $C_0$9 with training objective

$p_i$0

The policy parameterization is

$p_i$1

and the executed action is

$p_i$2

The paper states that candidate diversity comes from sampling different noisy inputs, while “selection” is largely implicit in the consistency-decoded generation rather than an external reranker (Zhang et al., 10 Jun 2026). Heterogeneous feature fusion combines scene features, route features, timestep encoding, and action tokens. For example, route and timestep are fused by

$p_i$3

and action tokens are fused with scene context through cross-attention. The practical role of consistency is denoising stability and trajectory coherence across the noise schedule. This suggests a different but still relevant sense of information-consistency-based plan selection: the final plan is the denoised action that remains coherent across latent perturbations and aligned with route and scene information.

6. Representation, empirical consistency, and broader selection criteria

In LLM web agents, the relevant problem is that plan quality cannot be judged adequately by single-run success alone because execution is stochastic and brittle (Zambrano et al., 28 May 2026). PlanAhead isolates the effect of plan representation by using a static planner-executor architecture in which the planner sees the task goal and initial screenshot once, while the executor receives the resulting plan and the evolving browser state. The study compares four plan representations on the hard subset of WebArena tasks: sequential subgoals, narrative, pseudocode, and checklist. The paper’s claim is that both the plan formulation and the underlying LLM generating the plan significantly influence robustness and task success.

To measure this, the paper introduces Achievement Rate (AR) and Solved-Task Consistency (STC). If $p_i$4 is the number of tasks, $p_i$5 the number of runs, and $p_i$6 the outcome of task $p_i$7 on run $p_i$8, then

$p_i$9

AR measures whether a task is solved at least once. Let

$(\psi \mid \phi)[l,u],$0

then

$(\psi \mid \phi)[l,u],$1

STC measures how consistently tasks are solved once they are achievable at all (Zambrano et al., 28 May 2026).

This defines a specifically empirical notion of information consistency. A plan representation is preferred not only because it can produce a success, but because it supports repeatable success under stochastic execution. The paper explicitly interprets STC as a robustness assessment and notes that a plan is better if it solves tasks consistently under stochastic conditions (Zambrano et al., 28 May 2026). The reported qualitative patterns are representation-dependent: GPT-4.1-mini tends to perform best with narrative plans, Qwen-2.5-VL with checklists, and Gemini with several nonstandard formats, including strong AR with pseudocode when used as both planner and executor.

A broader selection perspective is provided by information-criterion theory, even though that literature addresses models rather than plans. PanIC studies selection by penalized empirical risk

$(\psi \mid \phi)[l,u],$2

with penalty conditions

$(\psi \mid \phi)[l,u],$3

and, for $(\psi \mid \phi)[l,u],$4,

$(\psi \mid \phi)[l,u],$5

yielding consistency under weak assumptions (Nguyen, 2023). In fixed-dimensional exploratory factor analysis, the BIC result similarly hinges on separating underfit losses of order $(\psi \mid \phi)[l,u],$6 from exact-overfit fluctuations of order $(\psi \mid \phi)[l,u],$7 almost surely, so that the minimal optimal order is selected (Nguyen et al., 9 Apr 2026). A plausible implication is that plan selection can also be understood through scale separation: informative differences must dominate noise, while the selection rule must still preserve sensitivity to genuinely superior candidates.

Taken together, these literatures indicate that information-consistency-based plan selection is not reducible to one formal definition. In possible-world planning it means reapplicability across P-states; in argumentation it means survival in a conflict-free justified extension; in Quasi-Bayesian decision-making it means robustness over a credal set; in KL-regularized control it means bounded deviation from informational priors; in consistency-model trajectory generation it means coherent denoising across latent perturbations; and in stochastic web-agent evaluation it means repeatable success as quantified by AR and STC (Mansell, 2013, Morveli-Espinoza et al., 2020, Cozman et al., 2013, Grau-Moya et al., 2016, Zhang et al., 10 Jun 2026, Zambrano et al., 28 May 2026). The unifying theme is that plan quality is assessed relative to the structure of uncertainty itself, not solely by nominal goal achievement in a single presumed world.