Diffusion Scenario Tree (DST)

• Diffusion Scenario Tree (DST) is a family of tree-structured conditional models where nodes encode diffusion states and edges represent admissible transitions.
• In stochastic optimization, DST converts diffusion-based forecasts into multistage scenario trees that enforce non-anticipativity and enhance decision-making.
• In generative modeling and inference, DST underpins tree search over denoising trajectories, offering scalable improvements in planning and alignment.

Diffusion Scenario Tree (DST) denotes a tree-structured representation of diffusion scenarios, but the cited literature does not use the term as a single standardized theorem-like object. In current usage, DST can refer to a framework that converts diffusion-based probabilistic forecasts into multistage scenario trees for stochastic optimization, a scenario-tree interpretation of destroy-then-generate graphical models, or a tree of denoising trajectories used for inference-time planning and alignment. Closely related earlier work studies how diffusion unfolds on tree-shaped networks in a competitive, game-theoretic setting, where the tree itself determines the structure of strategic interaction (Zarifis et al., 18 Sep 2025, Noël, 28 May 2026, Zhao et al., 12 Oct 2025, Jain et al., 25 Jun 2025, Janssen et al., 2014).

1. Scope of the term and shared structural features

Across the cited works, DST-style constructions share a common abstraction: nodes encode conditional states of a diffusion process, and edges encode admissible transitions between those states. In diffusion-based stochastic optimization, each node corresponds to a possible future state and carries a probability mass inherited from sampled forecasts. In the destroy-then-generate formulation, nodes represent messages or samples at different destruction levels such as $A$, $B$, $C$, $\dots$, $Z$, together with reconstructed variables $A'$, $B'$, $\dots$. In inference-time tree search, nodes are partially denoised graph states or noisy diffusion states, and edges are macro denoising transitions or single reverse-diffusion transitions. Full root-to-leaf paths are therefore scenarios: future trajectories in forecasting, reconstruction paths in generative modeling, or denoising trajectories in inference-time guidance (Zarifis et al., 18 Sep 2025, Noël, 28 May 2026, Zhao et al., 12 Oct 2025, Jain et al., 25 Jun 2025).

A recurrent misconception is to equate DST with an ordinary forward Markov chain and its reverse. The destroy-then-generate account explicitly broadens the picture to deterministic or stochastic destroying processes, branching and scenario trees, non-noise destroyers such as masking or shuffling, and non-Markovian parameterizations. In the stochastic-optimization setting, the tree is not merely a visualization of uncertainty; it is the object that encodes which scenarios share a common past, enabling non-anticipativity constraints. In inference-time guidance, the tree is a planning structure that supports long-horizon lookahead instead of purely local stepwise correction. This suggests that DST is best understood as a family of tree-structured conditional models rather than a single algorithm.

2. Diffusion-based scenario trees for stochastic optimization

A fully explicit use of the term appears in "Diffusion-Based Scenario Tree Generation for Multivariate Time Series Prediction and Multistage Stochastic Optimization" (Zarifis et al., 18 Sep 2025). There, DST is defined as a general framework for constructing scenario trees for multivariate prediction tasks using a diffusion-based probabilistic forecasting model, specifically TimeGrad. The forecasting model estimates the conditional future distribution

$$p_\theta(\mathbf{x}_{k_0:N} \mid \mathbf{x}_{1:k_0-1}) \approx q(\mathbf{x}_{k_0:N} \mid \mathbf{x}_{1:k_0-1}),$$

with $\mathbf{x}_k \in \mathbb{R}^D$, and in the DST setting the stochastic component of the observation is sampled from

$B$0

Tree construction is recursive and queue-based. At each node, the forecaster samples $B$1 future trajectories over a stage horizon $B$2, reshapes them into

$B$3

clusters them with K-means into $B$4 groups, uses each centroid $B$5 as a representative branch, and assigns branch probability

$B$6

The cumulative probability of a child node is

$B$7

To control combinatorial growth, only the top-$B$8 children by cumulative probability are retained and then renormalized. The retained centroid is appended to the branch history,

$B$9

so each branch is expanded using its own realized history rather than being repeatedly resampled from the root.

The downstream control problem is modeled as a POMDP

$C$0

with observation decomposed into deterministic and stochastic parts,

$C$1

The deterministic part evolves according to known dynamics,

$C$2

Given a scenario tree $C$3, the multistage stochastic MPC problem is solved over stage-indexed node probabilities $C$4 subject to state propagation along predecessors and the non-anticipativity constraint

$C$5

This is the operational reason the tree matters: all nodes sharing the same information set must take the same action, so the resulting policy is implementable under partial observability.

The reported application is energy arbitrage in New York State’s day-ahead electricity market, with a Battery Energy Storage System, hourly electricity prices, and a reward equal to trading revenue minus battery degradation costs. DST SMPC achieves an average reward of 85.81, slightly better than Monte Carlo SMPC’s 85.56. It outperforms LSTM-based scenario-tree SMPC by 18.0%, VAR-based scenario-tree SMPC by 25.4%, and the best DQN baseline by 218.8%. It also improves on the authors’ earlier diffusion-informed standard MPC approach, indicating that converting the same diffusion forecaster into a scenario tree is more effective than using it in a non-tree MPC setup. The paper notes that branching frequency and observation frequency can differ; in the illustrated experiment, branching is daily while observations are hourly.

3. Destroy-then-generate graphical models

A different usage appears in "Destruction is a General Strategy to Learn Generation; Diffusion's Strength is to Take it Seriously; Exploration is the Future" (Noël, 28 May 2026). That work does not present DST as a boxed formal definition; instead, it introduces a novel kind of probabilistic graphical model called generative commutative diagrams, with DST as the scenario-tree style instantiation of that idea. The central viewpoint is that a data sample $C$6 is transformed by a destroying process into partially destroyed states $C$7, $C$8, $C$9, $\dots$0, and that a learned model reverses the process by guessing the missing information step by step. The paper’s vocabulary includes mashing, shuffling, guessing, and singleton / known distribution.

The basic reconstruction objective is conditional:

$\dots$1

with an unconditional version

$\dots$2

The paper frames this as an information-flow problem rather than as a forward-noising/reverse-denoising story. Its information-theoretic discussion uses

$\dots$3

and for a chain $\dots$4 it writes

$\dots$5

and

$\dots$6

These inequalities express the asymmetry between destruction and generation. For DDPM-like models, the paper also presents a non-Markovian parameterization in which intermediate states share a latent clean source $\dots$7 and a noise source $\dots$8:

$\dots$9

with

$Z$0

The graphical conventions are part of the formal message. Deterministic arrows $Z$1 denote true functions, harpoons $Z$2 denote stochastic channels, and squiggly lines denote same-marginal or equivalence-in-distribution relations. Converging arrows mean information is destroyed; branching arrows mean information is generated. The framework is explicitly not restricted to Gaussian noise and includes examples such as cold diffusion, mask diffusion LLMs, uniform substitution, and masking order generation. It is also tied to a broader argument about exploration and data scarcity: the same clean sample can be transformed into many partially destroyed variants, producing many training scenarios from a single datum. A plausible implication is that DST, in this usage, is less a forecasting object than a graphical calculus for specifying what information is withheld, what remains available as conditioning, and what the model is required to guess.

4. Tree search over denoising trajectories in controllable graph generation

"Controllable Graph Generation with Diffusion Models via Inference-Time Tree Search Guidance" introduces TreeDiff, a Monte Carlo Tree Search guided dual-space diffusion framework for controllable graph generation (Zhao et al., 12 Oct 2025). TreeDiff treats denoising as a sequential decision process over partially denoised graphs. Each node in the tree is a partially denoised graph state, and each edge is a macro transition in the denoising process. The MCTS loop follows four phases: Selection, Expansion, Simulation / Evaluation, and Backpropagation. Inference repeatedly runs a small number of MCTS rounds from the current root, selects the best child, commits to it as the new root, reduces the remaining timestep, and continues until reaching $Z$3. This is the sense in which TreeDiff is explicitly described as close to a “diffusion scenario tree” method: it constructs a search tree over alternative future denoising trajectories rather than applying only local guidance at the current step.

The practical viability of TreeDiff rests on three design components. The macro-step expansion strategy groups multiple denoising updates into a single transition, reducing tree depth and enabling long-horizon exploration. The dual-space denoising mechanism couples latent-space denoising with lightweight discrete correction in graph space, with the stated purpose of obtaining both scalability and structural fidelity. The dual-space verifier predicts long-term rewards from partially denoised graphs, enabling early value estimation without full rollouts. The auxiliary modules $Z$4, $Z$5, $Z$6, and $Z$7 are trained separately by trajectory distillation and then frozen. The verifier uses a dual-branch architecture: graph transformer for structure, MLP for latent state, and cross-attention for fusion.

The benchmarks span both 2D and 3D molecular generation, under unconditional and conditional settings. For 2D conditional generation, the paper uses BACE, BBBP, and HIV, with metrics including diversity, similarity, MAE, and accuracy, and conditions on SAS and SCS in some tasks. For 3D conditional generation, it uses QM9 with targets $Z$8, $Z$9, $A'$0, $A'$1, $A'$2, and $A'$3, evaluated in in-distribution and out-of-distribution settings. For 2D unconditional generation, it uses QM9 and ZINC250k with validity, FCD, NSPDK, and scaffold similarity. For 3D unconditional generation, it uses QM9 and Drugs with validity/uniqueness, atom stability, and molecule stability. The reported findings are that TreeDiff gets the best average rank in 2D conditional generation, is best overall in 3D conditional QM9 under both ID and OOD settings, achieves the lowest FCD on 2D unconditional QM9 and ZINC250k with near-perfect validity, and attains the best overall results on 3D unconditional QM9. The paper also emphasizes favorable inference-time scaling: performance improves rapidly up to about depth $A'$4, and then plateaus as compute rises. Implementation details include tree depth $A'$5, expansion width $A'$6, adaptive step variance $A'$7, guidance scale $A'$8, latent horizon $A'$9, structural refinement $B'$0, PUCT selection, Graph Transformer for 2D, SE(3)-Transformer for 3D, and a GPM-based verifier.

5. Diffusion Tree Sampling and reward-tilted posterior inference

A second tree-search line appears in "Diffusion Tree Sampling: Scalable inference-time alignment of diffusion models" (Jain et al., 25 Jun 2025). The paper considers a pretrained reverse diffusion chain

$B'$1

and defines a reward-tilted target distribution

$B'$2

This is equivalent to the KL-regularized objective

$B'$3

The paper’s central argument is that standard steering methods are limited by inaccurate value estimation at high noise levels. To address this, it interprets reverse diffusion as a finite-horizon tree in which each node is a noisy state $B'$4, each edge is one reverse diffusion transition, and each full reverse trajectory is a scenario.

The soft value function is

$B'$5

with soft Bellman recursion

$B'$6

A tree node stores $B'$7, timestep $B'$8, a value estimate $B'$9, a visit count $\dots$0, and children $\dots$1; a dummy root $\dots$2 transitions to the prior. One iteration of DTS has four phases: Selection, Expansion, Rollout, and Backup. Selection samples children from a Boltzmann distribution over value estimates; expansion uses progressive widening,

$\dots$3

rollout continues with the pretrained diffusion kernel until $\dots$4; and backup applies the soft Bellman update

$\dots$5

The sampling version is DTS; the greedy search variant is DTS$\dots$6, which uses greedy or UCT-style traversal for high-reward search rather than posterior sampling.

The paper proves an asymptotic consistency claim: as the number of tree iterations $\dots$7, the empirical distribution of terminal states converges to the target posterior $\dots$8. Empirically, DTS is evaluated on 2D toy problems, MNIST, CIFAR-10, text-to-image generation, and language completion. On MNIST and CIFAR-10 class-conditional generation, DTS matches the FID of the best-performing baseline with up to 10× less compute. In text-to-image generation and language completion, DTS$\dots$9 matches best-of-$$p_\theta(\mathbf{x}_{k_0:N} \mid \mathbf{x}_{1:k_0-1}) \approx q(\mathbf{x}_{k_0:N} \mid \mathbf{x}_{1:k_0-1}),$$0 with up to 5× less compute. The paper reports that the method is anytime: additional compute refines the tree and future samples reuse prior rollouts. It also notes practical costs and schedules: branching only at selected timesteps, progressive widening with $$p_\theta(\mathbf{x}_{k_0:N} \mid \mathbf{x}_{1:k_0-1}) \approx q(\mathbf{x}_{k_0:N} \mid \mathbf{x}_{1:k_0-1}),$$1 for 2D and images and $$p_\theta(\mathbf{x}_{k_0:N} \mid \mathbf{x}_{1:k_0-1}) \approx q(\mathbf{x}_{k_0:N} \mid \mathbf{x}_{1:k_0-1}),$$2 for text, batched diffusion-model calls by timestep, and higher memory cost than particle methods because many intermediate states are retained in the tree.

6. Competitive diffusion on trees as a related antecedent

"Finding safe strategies for competitive diffusion on trees" studies a two-player safe game of Competitive Diffusion on tree-shaped networks (Janssen et al., 2014). Although the paper does not formalize DST under that name, it is explicitly described as closely related to what one might call a diffusion scenario tree, because the diffusion process unfolds on a tree and the tree structure is central to both analysis and algorithm design. In the model introduced by Alon et al., each player chooses an initial seed vertex. Diffusion then proceeds in rounds: a white vertex with one or more neighbors of a single colour adopts that colour; a white vertex reached simultaneously by two different colours becomes grey; and grey vertices block further propagation. If Player 1 starts at $$p_\theta(\mathbf{x}_{k_0:N} \mid \mathbf{x}_{1:k_0-1}) \approx q(\mathbf{x}_{k_0:N} \mid \mathbf{x}_{1:k_0-1}),$$3 and Player 2 at $$p_\theta(\mathbf{x}_{k_0:N} \mid \mathbf{x}_{1:k_0-1}) \approx q(\mathbf{x}_{k_0:N} \mid \mathbf{x}_{1:k_0-1}),$$4, Player 1’s payoff is the $$p_\theta(\mathbf{x}_{k_0:N} \mid \mathbf{x}_{1:k_0-1}) \approx q(\mathbf{x}_{k_0:N} \mid \mathbf{x}_{1:k_0-1}),$$5 entry of the game matrix $$p_\theta(\mathbf{x}_{k_0:N} \mid \mathbf{x}_{1:k_0-1}) \approx q(\mathbf{x}_{k_0:N} \mid \mathbf{x}_{1:k_0-1}),$$6, and for mixed strategies $$p_\theta(\mathbf{x}_{k_0:N} \mid \mathbf{x}_{1:k_0-1}) \approx q(\mathbf{x}_{k_0:N} \mid \mathbf{x}_{1:k_0-1}),$$7 and $$p_\theta(\mathbf{x}_{k_0:N} \mid \mathbf{x}_{1:k_0-1}) \approx q(\mathbf{x}_{k_0:N} \mid \mathbf{x}_{1:k_0-1}),$$8,

$$p_\theta(\mathbf{x}_{k_0:N} \mid \mathbf{x}_{1:k_0-1}) \approx q(\mathbf{x}_{k_0:N} \mid \mathbf{x}_{1:k_0-1}),$$9

The safe game asks for a maxmin strategy, with safety value

$\mathbf{x}_k \in \mathbb{R}^D$0

A key structural fact is that any pure strategy for Player 1 has guaranteed gain $\mathbf{x}_k \in \mathbb{R}^D$1, because Player 2 can mirror it.

The analysis is organized around the centroid $\mathbf{x}_k \in \mathbb{R}^D$2 and branch weights. For a spider with $\mathbf{x}_k \in \mathbb{R}^D$3 equal legs of length $\mathbf{x}_k \in \mathbb{R}^D$4, the paper defines a mixed strategy $\mathbf{x}_k \in \mathbb{R}^D$5 that places equal probability on the body and on the first $\mathbf{x}_k \in \mathbb{R}^D$6 vertices of each leg:

$\mathbf{x}_k \in \mathbb{R}^D$7

It proves the asymptotic bound

$\mathbf{x}_k \in \mathbb{R}^D$8

as $\mathbf{x}_k \in \mathbb{R}^D$9, and states that the safe value matches the disadvantaged player’s payoff in a pure Nash equilibrium. For complete $B$00-ary trees $B$01, with

$B$02

the safety value is exact:

$B$03

The optimal strategies $B$04 and $B$05 are supported only on the root and the first level.

For arbitrary centroidal trees, the paper proposes the Centroidal Safe Strategy (CSS) algorithm. The tree is rooted at the centroid, branches are classified as thick, medium, or thin, and probability is assigned only to the centroid and a few nearby vertices in each branch. Branches are ordered by a criterion $B$06, and the algorithm adds branches in decreasing order until adding another branch no longer improves the guaranteed gain. Its output satisfies

$B$07

so the worst opposing move is the centroid itself. Experiments on randomly generated centroidal trees with $B$08 and $B$09 show that CSS performs well overall and that its guaranteed gain is often close to an upper bound, but performance degrades when the tree has many large thin branches. In relation to later DST formulations, this line of work shows that tree structure can be central even when diffusion is modeled not as generative uncertainty but as adversarial competition over branching propagation paths.