Polyscene: Scenario Planning as Diversity Planning

Mustafa F. Abdelwahed; Joan Espasa; Felipe Meneguzzi

cs.AIOct 20, 2026

Polyscene: Scenario Planning as Diversity Planning

Authors: Mustafa F. Abdelwahed, Joan Espasa, Felipe Meneguzzi

Organizations: School of Computer Science, University of St Andrews, St Andrews · School of Natural and Computing Sciences, University of Aberdeen, Aberdeen

Abstract

Scenario planning is a foresight practice in which analysts construct a small set of plausible futures and prepare an organisation for each. Each scenario is a narrative that must be plausible, internally consistent, relevant and challenging, and the set as a whole must be differentiated. The practice remains manual: a workshop (i.e., a facilitated meeting in which a small group of experts writes the narratives) considers only the factors it has time for, carries the biases of the experts who run it, and cannot be redone whenever conditions change. In this paper, we automate scenario planning by casting it as a diversity planning problem, in which two scenarios differ when they end in different narratives on the analyst's drivers. The result is a set of scenarios, one for each narrative that the topic can reach within a budget. Our implementation, Polyscene, is a generic engine that generates scenarios from a topic. We use geopolitical conflicts and shipping-route disruptions as case studies. For the first, we build the planning problem's initial state from the GDELT Project so it reflects current relationships among the countries involved. Polyscene then generates scenarios showing how relationships among those countries change. With a board of six countries as of 1 January 2021 and a budget of five events, the conflict set puts five of the 15 country pairs at war, and each of the three wars that break out among these countries by 1 January 2026 ends a scenario. For the second, we build the initial state from ship traffic, a supply-pressure index, and the news, and Polyscene generates scenarios showing what a closed or diverted route could do to a manufacturer's parts stock. With four routes per board and the same date and budget, each set includes three outcomes, and the Suez Canal set includes the March 2021 blockage and the December 2023 Red Sea container diversion.

Figures & tables

Operator	Name	Holds at position $i$ when
$\mathord{\circ}\theta$	next	$\theta$ holds at $i+1$ ; false at the last position
$\theta_{1}\mathbin{\mathcal{U}}\theta_{2}$	until [ 29 ]	$\theta_{2}$ holds at some position $j\geq i$ , and $\theta_{1}$ holds at every position from $i$ up to, but excluding, $j$
$\mathord{\ominus}\theta$	yesterday [ 30 ]	$i\geq 1$ and $\theta$ holds at $i-1$ ; false at position $0$
$\theta_{1}\mathbin{\mathcal{S}}\theta_{2}$	since	$\theta_{2}$ holds at some position $j\leq i$ , and $\theta_{1}$ holds at every position after $j$ up to and including $i$
$\Diamond\theta\equiv\top\mathbin{\mathcal{U}}\theta$	sometime	$\theta$ holds at $i$ or at a later position
$\Box\theta\equiv\neg\Diamond\neg\theta$	always	$\theta$ holds at every position from $i$ to the last

Table 1: The temporal operators of the trajectory constraints of Definition 1 .

Condition	Judges	A narrative breaks it when	Source
(1) intentional	an actor’s action	the action raises no objective of its actor, replies to no action aimed at its actor, and starts no causal chain that ends in an action that does either	intentionality [ 31 ] , read through utilities [ 32 ]
(2) connected	an actor’s action after the first	no causal link reaches it from an earlier action, and it replies to none and undoes none	causal links [ 45 ]
(3) connected	a happening	no causal link leads from it to a later action of an actor	happenings [ 31 ]
(4) loop-free	an actor’s action	the state before the action occurred earlier in the narrative	loop-free plans [ 46 ]

Table 2: The plausibility conditions and their sources.

Given a narrative $\pi_{\mathit{}}=\langle a_{1},\ldots,a_{m}\rangle$ with the states $s_{0}=I$ and $s_{i}=\gamma(s_{i-1},a_{i})$ , an actor $g\in\Gamma$ , a fluent $v\in V$ , a value $x\in\operatorname{range}(v)$ , positions $i,j,k\in\{1,\ldots,m\}$ , and a bound $h$ on the number of actions of a narrative; $\mathrm{vars}(\operatorname{pre}_{a})$ denotes the fluents that occur in the precondition of $a$ , $H\subseteq A$ the happenings, $S$ the states of $\Xi$ , $\hat{s}=\bigwedge_{v\in V}v=s[v]$ the formula that holds in the state $s$ alone, and $c$ ranges over the values that an objective takes.
First-order definition	LTL f implementation	Holds when
$\mathrm{acts}(a_{i},g):=\ \mathrm{act}(a_{i})=g$	$\mathrm{Acts}_{g}:=\bigvee\{a\in A\mid\mathrm{act}(a)=g\}$	the action $a_{i}$ is the actor $g$ ’s
$\mathrm{actorActs}(a_{i}):=\ \exists g\in\Gamma:\ \mathrm{acts}(a_{i},g)$	$\mathrm{ActorActs}:=\bigvee_{g\in\Gamma}\mathrm{Acts}_{g}$	the action $a_{i}$ has an actor
$\mathrm{hap}(a_{i}):=\ \neg\mathrm{actorActs}(a_{i})$	$\mathrm{Hap}:=\bigvee H$	the action $a_{i}$ is a happening
$\mathrm{reads}(a_{i},v):=\ v\in\mathrm{vars}(\operatorname{pre}_{a_{i}})$	$\mathrm{Reads}_{v}:=\bigvee\{a\in A\mid v\in\mathrm{vars}(\operatorname{pre}_{a})\}$	the action $a_{i}$ reads $v$
$\mathrm{repeated}(s_{i}):=\ \exists j<i:\ s_{j}=s_{i}$	$\mathrm{Rep}:=\bigvee_{s\in S}\big(\hat{s}\wedge\mathord{\ominus}\mathord{\blacklozenge}\hat{s}\big)$	the state $s_{i}$ occurred at an earlier position

Table 3: The relations over narrative positions required by Definition 9 . Implementations are evaluated at positions under the grammar of Definition 1 , where

v=x

is a formula over

V

, and each action atom is in

A

Step	Element of $\mathbb{T}$	Geopolitical conflicts ( Section 4.1 )	Supply chains ( Section 4.2 )
1	$\Xi$ : domain	65 action schemas over 6 countries, 1 per CAMEO event type; fluents for the pair ladder, 9 ties, discontent and the economy.	17 action schemas over the lines, the maker, the rivals, the supplier and the authorities, 6 of them happenings; fluents for closure, attack, diversion, shortage, orders and stops, and 4 numeric levels; flying parts in, switching source and escorting ships require a fact the analyst states on the board (air-transportable parts, a qualified second source, spare warships), false on every board here.
	actors $\Gamma$	the 6 countries; every action is a country’s, and its target is the country it is directed at	the lines, the maker, the rivals, the supplier and the authorities; the 6 happenings have no actor
2	$C$	the partnership of each pair and the economy of each country	the maker’s stock
	$D$	whether each of the 15 pairs is at war	whether the route is closed, whether the lines have diverted, and whether the maker has stopped a line
	objectives $\mathcal{O}$	the week’s motives and the maintained table; 1 preference per motive, weighted by the motive’s weight	lines want ships moving ( $2$ ) and rates up ( $1$ per level); buyers want no missed delivery ( $3$ ), their lines running ( $2$ ) and stock up ( $1$ per unit); the supplier wants fillable orders ( $1$ ); the authorities want routes open ( $2$ ) and safe ( $2$ )
	$\Phi$	escalation, stranger, reversal and capitulation	none; the invariants are in the preconditions

Table 4: The 2 topics of this section.

		Outcomes at budget				Budget $5$
Topic	Board (1 January 2021)	$4$	$5$	$6$	$7$	Time	Memory
Conflicts	6 countries	$5$	$8$	$9$	$11$	$8.4$ s	$350$ GB
Supply chain	Suez Canal	$3$	$3$	$3$	$3$	$0.07$ s	$36$ MB
	Panama Canal	$3$	$3$	$3$	$3$	$0.07$ s	$36$ MB
	Baltimore	$2$	$3$	$3$	$3$	$0.07$ s	$37$ MB
	Strait of Hormuz	$3$	$3$	$3$	$3$	$0.07$ s	$36$ MB

Table 5: The number of outcomes at budgets

4

5

6

and

7

, and the clingo runtime and peak memory at budget

5

, as means of 10 single-thread runs on an Apple M1 laptop (8 cores, 16 GB, clingo 5.8.2).

Wars at the end	Scenario	Happened
Iran–Israel	Israel masses forces against Iran and strikes it	Iran–Israel: April and October 2024, 13 to 24 June 2025
Iran–US	Iran masses forces against the United States and strikes it	Iran–US: 22 to 24 June 2025
Iran–Ukraine	Iran masses forces against Ukraine and strikes it, then masses forces against the United States	no
Iran–Ukraine, Iran–US	Iran masses forces against Ukraine and strikes it; Iran and the United States mass forces against each other; the United States strikes Iran	Iran–US: 22 to 24 June 2025
Iran–Israel, Iran–US	Iran masses forces against Israel and strikes it; Iran and the United States mass forces against each other; the United States strikes Iran	Iran–Israel: April and October 2024, 13 to 24 June 2025; Iran–US: 22 to 24 June 2025
Russia–US	Russia imposes targeted sanctions on the United States; the United States appeals to Iran for support against Russia; Russia threatens the United States; the United States masses forces against Russia and strikes it	no

Table 6: Scenarios use a

5

budget as of 1 January 2021. 1 per reachable outcome of

D

. Each outcome is the set of pairs at war at the end. The last column shows which wars in the outcome occurred between 1 January 2021 and 1 January 2026, based on dated news reports.

Outcome			Scenario	Happened
closed	diverted	paused
true	true	false	demand for the parts surges; the canal is blocked; the lines reroute round the Cape of Good Hope; the maker inflates its order; the lines add a surcharge	yes: 23–29 Mar 2021, the Ever Given
false	true	false	demand for the parts surges; the canal is blocked; the lines reroute round the Cape; the maker inflates its order; the authorities clear the canal	yes: from mid-Dec 2023 to the end of the window
true	false	false	demand for the parts surges; the rivals inflate their orders; the canal is blocked; the maker inflates its order; the lines add a surcharge	no

Table 7: Scenarios for the Suez board, listed by reachable

D

outcomes. The outcome

\langle\mathit{false},\mathit{false},\mathit{false}\rangle

is excluded. The last column indicates whether the outcome occurred between 1 January 2021 and 1 January 2026, based on news reports and company announcements.

Model-Based Planning

Model-Based Planning uses explicit or learned models of how actions change states to simulate futures, compare plans, and choose actions.

See all 208 papers

Jul 3, 2026

A Bayesian Framework for Evaluating Scenario Compatibility in Generative Population Synthesis

Scenario-based transportation analysis specifies future assumptions through aggregate population targets, whereas generative population synthesis models produce detailed individual-level realizations. When scenario targets are imposed on generative models, current practice relies on deterministic marginal calibration, implicitly assuming that the targets are compatible with the model's learned structural support. However, whether scenario-level constraints lie within the generative support--and how strongly they distort structural uncertainty--remains largely unexamined. We propose an ensemble-based Bayesian updating framework to quantify scenario compatibility in conditional population synthesis. A population-aware conditional variational autoencoder is developed to learn a distribution over plausible population structures while preserving aggregate fidelity. An ensemble of realizations sampled from the learned prior provides an empirical approximation of structural uncertainty. Scenario targets are treated as probabilistic evidence over aggregate statistics, and posterior weights are obtained through Bayesian updating across the ensemble. Scenario compatibility is quantified using effective sample size (ESS), which measures posterior concentration and the compression of structural uncertainty induced by conditioning. Experiments demonstrate that scenario impact depends not only on target magnitude but also on alignment with the learned joint structure, and reveal structural failure modes when targets fall outside prior ensemble support. The proposed framework provides a probabilistic diagnostic model for evaluating scenario feasibility and structural consistency before downstream projection and transportation planning.

Zhenlin Qin, Leizhen Wang, Yancheng Ling +1

Department of Civil and Architectural Engineering, KTH Royal Institute of Technology, Stockholm, Sweden. · Department of Data Science and Artificial Intelligence, Monash University, Melbourne, Australia. · Digital Futures, KTH Royal Institute of Technology, Stockholm, Sweden.

Uncertainty Quantification Bayesian Inference

Ask PDF

Oct 26, 2026

Model-Aware Generation of Safety-Critical Scenes for Robust End-to-End Driving via Distilled Planner Surrogates

End-to-end driving models perform well on standard benchmarks but collide frequently in safety-critical scenarios. We present a failure-mode analysis on NeuroNCAP which shows that 95.9% of DiffusionDrive's collisions are planning failures despite reliable detection, pinpointing the bottleneck in planning rather than perception. We attribute this to a fundamental data imbalance, as safety-critical interactions are rare in real driving logs, so planners are hardly ever trained on them. We close this gap with a novel model-aware framework that generates safety-critical training scenes against a specific planner. Each planner is attacked at the bird's-eye-view (BEV) level through a differentiable surrogate distilled from its own planning head. This keeps the expensive video renderer out of the attack loop. We extract 50 scenes per planner, each pairing an attack with a safe escape trajectory that serves as the fine-tuning label. A fully automated quality-diversity search generates both the attacks and the escapes, so that their variety is not specified by hand. Only these scenes are rendered to fine-tune the planner, which reduces the computational cost significantly. Fine-tuning on these scenes roughly halves the closed-loop collision rate of two planners on NeuroNCAP (SparseDrive 28.0% to 12.2%, DiffusionDrive 37.5% to 19.2%) while largely preserving normal driving. A cross-planner transfer experiment shows that the gain comes from targeting the specific planner. Each planner improves most on the scenes generated against itself, and DiffusionDrive fine-tuned on SparseDrive's scenes retains only 6.7 of its 18.3-point collision-rate reduction.

Khaled M. Seyam, Julian Wiederer, Samar Fathalah +2

University of Stuttgart, Germany · Mercedes-Benz AG, Germany

End-to-End Autonomous Driving Autonomous Driving Scenario Generation Adversarial Scenario Generation Autonomous Driving Safety Evaluation

Ask PDF

May 20, 2026

PlanningBench: Generating Scalable and Verifiable Planning Data for Evaluating and Training Large Language Models

Planning is a fundamental capability for large language models (LLMs) because such complex tasks require models to coordinate goals, constraints, resources, and long-term consequences into executable and verifiable solutions. Existing planning benchmarks, however, usually treat planning data as fixed collections of instances rather than controllable generation targets. This limits scenario coverage, ties difficulty to surface-level proxies rather than structural sources, and offers limited support for scalable generation, automatic verification, or planning-oriented training. We introduce PlanningBench, a framework for generating scalable, diverse, and verifiable planning data for both evaluation and training. PlanningBench starts from real planning scenarios and abstracts practical workflows into a structured taxonomy of more than 30 task types, subtasks, constraint families, and difficulty factors. Guided by this taxonomy, a constraint-driven synthesis pipeline instantiates self-contained planning problems with adaptive difficulty control, quality filtering, and instance-level verification checklists. This shifts planning data construction from fixed benchmark collection to controllable generation while preserving realistic task grounding. We use PlanningBench to evaluate open-source and closed-source frontier LLMs, and find that current models still struggle to produce complete solutions under coupled constraints. Beyond evaluation, reinforcement learning on verified PlanningBench data improves performance on unseen planning benchmarks and broader instruction-following tasks. Further analysis suggests that determinate or well-specified optimal solutions provide clearer reward signals and more stable training dynamics. Overall, PlanningBench provides a controllable source of planning data for diagnosing and improving generalizable planning abilities in LLMs.

Ziliang Zhao, Zenan Xu, Shuting Wang +7

Gaoling School of Artificial Intelligence, Renmin University of China · LLM Department, Hunyuan Team, Tencent · Beijing Academy of Artificial Intelligence +1

LLM Evaluation LLM Planning Synthetic Benchmark Generation Synthetic Data Generation for Language Models

Ask PDF

Abstract

Model-Based Planning

Explore similar work

A Bayesian Framework for Evaluating Scenario Compatibility in Generative Population Synthesis

Model-Aware Generation of Safety-Critical Scenes for Robust End-to-End Driving via Distilled Planner Surrogates

PlanningBench: Generating Scalable and Verifiable Planning Data for Evaluating and Training Large Language Models