Diffusion Tree Sampling for Inference Time Adaptation of Crystal Generative Models
Summary
The paper proposes Diffusion Tree Sampling (DTS), a Monte Carlo tree search over the denoising trajectory for inference-time adaptation of diffusion-based crystal generative models, extending it with guided proposals from conditional models and multi-objective Pareto-front sampling without retraining the base models.
View Cached Full Text
Cached at: 11/16/26, 08:42 AM
# Diffusion Tree Sampling for Inference Time Adaptation of Crystal Generative Models
Source: [https://arxiv.org/html/2611.17462](https://arxiv.org/html/2611.17462)
Daniel T\. Levy\*Vineet Jain Tara Akhound\-Sadegh††thanks:Equal contribution\. Correspondence: daniel\.levy6@mail\.mcgill\.ca\. Code is available at:[https://github\.com/danielTLevy/MaterialDTS](https://github.com/danielTLevy/MaterialDTS)Affiliation:Sékou\-Oumar KabaSiamak RavanbakhshAffiliation:McGill UniversityAffiliation:Mila \- Quebec Artificial Intelligence Institute
###### Abstract
Diffusion\-based generative models have become strong priors over plausible crystal structures, but turning that prior into promising candidates requires steering toward several conflicting properties at once\. Fine\-tuning is expensive and produces a new model for every new objective, while differentiable guidance is not applicable to many real\-world tasks\. We instead adapt these generative models at*inference time*using Diffusion Tree Sampling \(DTS\), a Monte Carlo tree search procedure over the denoising trajectory\. We make two technical contributions on top of DTS that are particularly relevant for AI\-for\-science applications\. First, we show how to use existing property\-conditional generative models \(e\.g\., models fine\-tuned for classifier\-free guidance\) as guided proposals inside DTS, without distorting the target distribution\. Second, we extend DTS to multi\-objective sampling, so that a single search produces samples spanning an entire Pareto front\. We instantiate both adaptations on de novo crystal generation: improving the stability of crystals generated by different models, using pretrained conditional models to generate crystals belonging to rare space groups or with high bulk modulus, and generating crystals along a Pareto front of high dielectric constant and high bandgap materials\. Across these tasks, DTS\-based inference\-time adaptation successfully improves model outputs without modifying the models themselves\.
## 1Introduction
Foundation models are flourishing in AI for science due to the superior generalization of large models trained on massive datasets\([Kaplan et al\., 2020](https://arxiv.org/html/2611.17462#bib.bib10);[Bommasani et al\., 2021](https://arxiv.org/html/2611.17462#bib.bib9)\)\. In materials discovery, this has motivated powerful diffusion\-based generative models that capture the distribution of plausible crystal structures: MatterGen\([Zeni et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib1)\)and DiffCSP\([Jiao et al\., 2023](https://arxiv.org/html/2611.17462#bib.bib12)\)jointly model atom types, lattice, and coordinates, and can sample chemically plausible crystals\. Yet, despite strong sample quality, these models have rarely translated into confirmed new discoveries\.
A key barrier is that unconditional generation over all possible crystals is generally not directly useful\. Candidate materials should not just be sampled from the manifold of all materials, but rather obtained through exploration of structures that are stable, sit in a particular property window, and trade off several conflicting properties \(functionality, cost, element availability\)\. Existing strategies for diffusion models typically modify the base model: fine\-tuning property\-conditional variants\([Zeni et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib1)\), reward\-conditioned RL\([Chen et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib14)\), or active learning loops\([Li et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib27)\)\. Each requires labeled data, training compute, and produces a new model per design problem, which is at odds with the foundation\-model paradigm\.
Inference\-time adaptation offers a different path: leave the pretrained model fixed, and steer its sampling toward the user’s specific goals\([Dhariwal and Nichol, 2021](https://arxiv.org/html/2611.17462#bib.bib7);[Ho and Salimans, 2022](https://arxiv.org/html/2611.17462#bib.bib8);[Yeh et al\., 2024](https://arxiv.org/html/2611.17462#bib.bib6);[Tang et al\., 2024](https://arxiv.org/html/2611.17462#bib.bib5);[Jain et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib15);[Uehara et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib4)\)\. This is well suited to materials, where the validation of candidate stability with density functional theory \(DFT\) calculations and the subsequent attempt to synthesize them is orders of magnitude more expensive than generation\. Therefore, allocating inference compute to candidate quality is far more valuable than producing high volumes of low\-quality structures\. Among inference\-time methods, Diffusion Tree Sampling \(DTS\)\([Jain et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib15)\)casts denoising as a tree of partial trajectories, allocating compute adaptively to promising branches via Monte Carlo Tree Search \(MCTS\)\.
Two features of materials candidate generation are not addressed by the original DTS formulation, and motivate this work\. First, in many cases, there*are*property\-conditional models – e\.g\., MatterGen has variants for individual properties – which are dramatically more effective than the unconditional model at finding samples with a given property\. Naively swapping in a conditional sampler inside DTS tilts towards the reward twice, resulting in a different distribution from the target\. Second, real design problems are rarely single\-objective: properties trade off, and the appropriate target is the*Pareto front*of the conflicting objectives\([Miettinen, 1999](https://arxiv.org/html/2611.17462#bib.bib3);[Deb et al\., 2016](https://arxiv.org/html/2611.17462#bib.bib2);[Riebesell et al\., 2024](https://arxiv.org/html/2611.17462#bib.bib11)\)\.
Figure 1:Overview of the method\.\(a\): DTS treats intermediate states of denoising diffusion as nodes in a tree, and steers sampling towards a reward\-tilted distribution\. Each node is attached to an estimate of expected reward \(denoted here by node size\)\.\(b\): We may use a different modelqcq\_\{c\}that gives better denoising proposals, but we need to correct its influence by reweighting nodes by the proposal’s ratio with the base model\.\(c\): In the multi\-objective setting, we wish to sample from a distribution obtained by integrating reward\-tilted distributions over different preferences between objectives\.\(d\): We use a single DTS tree to sample from this multi\-objective target, conditioning each pass through the tree on a different preference\.##### Contributions\.
\(1\)*Conditional generative models inside DTS\.*We show how to use property\-conditional diffusion models \(e\.g\. fine\-tuned variants of MatterGen\) as guided proposals within DTS, while keeping the reward\-tilted unconditional base model as the target\. This uses an importance\-corrected backup and a corrected selection rule\.\(2\)*Multi\-objective DTS\.*We formalize multi\-objective sampling from a diffusion model as sampling from a mixture of reward\-tilted distributions, each corresponding to a different preference over objectives\. We extend DTS to this setting by maintaining per\-objective soft value estimates and selecting expansions through sampled preferences, so that a single search samples the Pareto front while sharing computation across preferences and across different conditioning\.\(3\)*Application to materials discovery\.*We use the method to improve de novo crystal generation across multiple tasks: we show improved results in generating stable crystals, we demonstrate the ability to generate crystals with specific features using pretrained conditional models, and we demonstrate how to sample crystals along a Pareto front of conflicting objectives, all without having to modify any model weights or use reward gradients\.
## 2Background
### 2\.1Diffusion models for materials discovery
Conditional materials generation aims to propose previously unknown crystalline compounds with desirable physical properties\. The space of possible structures is combinatorially large, spanning composition, stoichiometry, lattice geometry, and atomic coordinates; only a small fraction is thermodynamically stable, and a smaller fraction still exhibits any given target property\. We focus on*de novo generation*, which jointly proposes composition, lattice, and coordinates\. Stability of candidates can be validated computationally by DFT, which is orders of magnitude more expensive than generation\. Classical pipelines combine heuristic generation with first\-principles relaxation\([Oganov and Glass, 2006](https://arxiv.org/html/2611.17462#bib.bib16);[Pickard and Needs, 2011](https://arxiv.org/html/2611.17462#bib.bib17)\)\. Some methods search using basin hopping\([Wales and Doye, 1997](https://arxiv.org/html/2611.17462#bib.bib19);[Wales and Scheraga, 1999](https://arxiv.org/html/2611.17462#bib.bib20)\)– a structure with a natural analogue in the tree\-search method we use \(see Appendix[B](https://arxiv.org/html/2611.17462#A2)\)\.
A recent line of work trains generative models directly on databases of known crystals; diffusion models are a strong fit since they jointly model the discrete \(atom types\) and continuous \(lattice and coordinates\) components of a crystal\. We focus on two models\.DiffCSP\([Jiao et al\., 2023](https://arxiv.org/html/2611.17462#bib.bib12)\)parameterizes a joint diffusion process over fractional coordinates, lattice, and atom types, and supports time\-conditional classifier\-based guidance\.MatterGen\([Zeni et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib1)\)is trained on a substantially larger dataset and provides adapter modules for efficient fine\-tuning into property\-conditional models, sampled with classifier\-free guidance\([Ho and Salimans, 2022](https://arxiv.org/html/2611.17462#bib.bib8)\)\. Both suffer from a high rate of false positives: only a small fraction of generated structures are predicted to be stable, and a smaller fraction yet satisfies property constraints\. Existing strategies modify the base model via classifier\-based guidance\([Jiao et al\., 2023](https://arxiv.org/html/2611.17462#bib.bib12)\), fine\-tuned conditional variants\([Zeni et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib1)\), RL\([Chen et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib14)\), and active learning\([Li et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib27)\)\. Our work is complementary: it steers a fixed pretrained generative model at inference time, and extends this steering to settings where conditional variants exist and where multiple objectives must be balanced\.
### 2\.2Inference\-time steering and Diffusion Tree Sampling
Letpθp\_\{\\theta\}denote a pretrained diffusion model with reverse transitionspθ\(𝐱t−1∣𝐱t\)p\_\{\\theta\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\\mid\{\\mathbf\{x\}\}\_\{t\}\), inducing a prior𝐱0∼pθ\(𝐱0\)\{\\mathbf\{x\}\}\_\{0\}\\sim p\_\{\\theta\}\(\{\\mathbf\{x\}\}\_\{0\}\)\. Inference\-time steering introduces a rewardr\(𝐱0\)r\(\{\\mathbf\{x\}\}\_\{0\}\)available only after full denoising, and aims to sample from the*reward\-tilted*targetπ⋆\(𝐱0\)=1Zpθ\(𝐱0\)exp\(λr\(𝐱0\)\)\\pi^\{\\star\}\(\{\\mathbf\{x\}\}\_\{0\}\)=\\tfrac\{1\}\{Z\}p\_\{\\theta\}\(\{\\mathbf\{x\}\}\_\{0\}\)\\exp\(\\lambda r\(\\rvx\_0\)\), which is the KL\-regularized optimum of expected reward relative to the base model with inverse temperatureλ\\lambda\. A useful characterization of expected reward at a noisy state𝐱t\{\\mathbf\{x\}\}\_\{t\}is the*soft value function*vt\(𝐱t\)=1λlog𝔼\[exp\(λr\(𝐱0\)\)∣𝐱t\]v\_\{t\}\(\{\\mathbf\{x\}\}\_\{t\}\)=\\tfrac\{1\}\{\\lambda\}\\log\\mathbb\{E\}\[\\exp\(\\lambda r\(\\rvx\_0\)\)\\mid\{\\mathbf\{x\}\}\_\{t\}\]withv0\(𝐱0\)=r\(𝐱0\)v\_\{0\}\(\{\\mathbf\{x\}\}\_\{0\}\)=r\(\{\\mathbf\{x\}\}\_\{0\}\), which satisfies the soft Bellman recursion
vt\(𝐱t\)=1λlog𝔼pθ\(𝐱t−1∣𝐱t\)\[exp\(λvt−1\(𝐱t−1\)\)\]\.v\_\{t\}\(\{\\mathbf\{x\}\}\_\{t\}\)\\;=\\;\\frac\{1\}\{\\lambda\}\\log\\mathbb\{E\}\_\{p\_\{\\theta\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\\mid\{\\mathbf\{x\}\}\_\{t\}\)\}\\\!\\left\[\\exp\(\\lambda v\_\{t\-1\}\(\\rvx\_\{t\-1\}\)\)\\right\]\.\(1\)The optimal aligned reverse transition isπt⋆\(𝐱t−1∣𝐱t\)∝pθ\(𝐱t−1∣𝐱t\)exp\(λvt−1\(𝐱t−1\)\)\\pi^\{\\star\}\_\{t\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\\mid\{\\mathbf\{x\}\}\_\{t\}\)\\propto p\_\{\\theta\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\\mid\{\\mathbf\{x\}\}\_\{t\}\)\\exp\(\\lambda v\_\{t\-1\}\(\\rvx\_\{t\-1\}\)\)\. Most inference\-time steering methods approximatevt−1v\_\{t\-1\}via gradient guidance, SMC, or local search over denoising proposals\([Uehara et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib4)\)\.
A central difficulty is*value estimation at high noise*: common approximations become unreliable early in the reverse process\.*Diffusion Tree Sampling*\(DTS\)\([Jain et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib15)\)addresses this by casting denoising as a finite\-horizon tree whose nodes are states𝐱t\{\\mathbf\{x\}\}\_\{t\}with value estimatesv^t\(𝐱t\)\\hat\{v\}\_\{t\}\(\{\\mathbf\{x\}\}\_\{t\}\)and whose edges are reverse transitions\. The algorithm repeatedly builds a tree by starting at a root node𝐱T\+1\{\\mathbf\{x\}\}\_\{T\+1\}and performing \(1\)*selection*of children∝exp\(λv^\)\\propto\\exp\(\\lambda\\hat v\)until reaching a node that should expand, \(2\)*expansion*by sampling a new child frompθp\_\{\\theta\}, \(3\)*rollout*of the new child to a terminal𝐱0\{\\mathbf\{x\}\}\_\{0\}, assigningv^\(𝐱0\)=r\(𝐱0\)\\hat\{v\}\(\{\\mathbf\{x\}\}\_\{0\}\)=r\(\{\\mathbf\{x\}\}\_\{0\}\)and \(4\)*backup*via the soft Bellman log\-sum\-exp\. Reusing these backed\-up values yields an anytime procedure whose empirical leaf distribution converges toπ⋆\\pi^\{\\star\}; a greedy variant replaces sampling with argmax selection\.
## 3Adapting Diffusion Tree Sampling to Crystal Generative Models
Standard DTS assumes \(i\) a single scalar reward, and \(ii\) that the reverse transitions used during expansion and rollout are exactly the unconditional base model\. Both assumptions are limiting in the materials setting\. In this section we adapt DTS to relax them:[Section3\.1](https://arxiv.org/html/2611.17462#S3.SS1)shows how to use existing property\-conditional generative models as guided proposals inside DTS without changing the target distribution;[Section3\.2](https://arxiv.org/html/2611.17462#S3.SS2)extends the search to multi\-objective sampling\. The full algorithm is given in Appendix[A](https://arxiv.org/html/2611.17462#A1)\.
### 3\.1Adaptation 1: leveraging conditional models inside DTS
When available, a*property\-guided*diffusion model – e\.g\., a conditional model trained for classifier\-free guidance \(CFG\) – can be much more effective at*locally*steering denoising steps toward narrow high\-reward regions than a purely reward\-based search on top of the unconditional model\. For example, MatterGen provides property\-conditional variants for individual properties \(bandgap, space group, etc\.\), and we would like to use them as proposals inside DTS without losing the well\-defined target distribution induced by the unconditional base model\.
We denote the unconditional reverse transition bypθ\(𝐱t−1∣𝐱t\)p\_\{\\theta\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\\mid\{\\mathbf\{x\}\}\_\{t\}\)and a guided/conditional transition bypθ\(𝐱t−1∣𝐱t,c\)p\_\{\\theta\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\\mid\{\\mathbf\{x\}\}\_\{t\},c\), whereccencodes a target \(e\.g\. a desired property\)\. In our setting, we assume that providing conditioning informationcchelps with optimizing some scalar rewardr\(⋅\)r\(\\cdot\)\(e\.g\. conditioning on a target property\)\.
##### The double\-tilt problem\.
A naive attempt would be to replace the base model inside DTS with the conditional model and then tilt byexp\(λr\(𝐱0\)\)\\exp\(\\lambda r\(\\rvx\_0\)\)\. This would targetπnaive\(𝐱0\)∝pθ\(𝐱0∣c\)exp\(λr\(𝐱0\)\)\\pi\_\{\\text\{naive\}\}\(\{\\mathbf\{x\}\}\_\{0\}\)\\propto p\_\{\\theta\}\(\{\\mathbf\{x\}\}\_\{0\}\\mid c\)\\,\\exp\(\\lambda r\(\\rvx\_0\)\), which would tilt twice: the conditional model already biases mass toward the properties encoded bycc, and the additional Boltzmann tilt amplifies the same preference, yielding a different target than the originalπ⋆\\pi^\{\\star\}\. An example of this issue is shown in Appendix[C](https://arxiv.org/html/2611.17462#A3), using a Gaussian\-mixture toy experiment\.
##### Importance\-corrected backup\.
Our solution is to keep the*unconditional*model as the target and use the conditional model only as a proposal, correcting via importance sampling \(IS\)\. Concretely, during expansion and rollout from𝐱t\{\\mathbf\{x\}\}\_\{t\}we sample𝐱t−1\{\\mathbf\{x\}\}\_\{t\-1\}from the guided proposalqc\(𝐱t−1∣𝐱t\)≔pθ\(𝐱t−1∣𝐱t,c\)q\_\{c\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\\mid\{\\mathbf\{x\}\}\_\{t\}\)\\coloneqq p\_\{\\theta\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\\mid\{\\mathbf\{x\}\}\_\{t\},c\)instead ofpθ\(𝐱t−1∣𝐱t\)p\_\{\\theta\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\\mid\{\\mathbf\{x\}\}\_\{t\}\), and attach to the new node the*importance weight*
η\(𝐱t−1\)≔pθ\(𝐱t−1∣𝐱t\)qc\(𝐱t−1∣𝐱t\),\\eta\(\{\\mathbf\{x\}\}\_\{t\-1\}\)\\coloneqq\\frac\{p\_\{\\theta\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\\mid\{\\mathbf\{x\}\}\_\{t\}\)\}\{q\_\{c\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\\mid\{\\mathbf\{x\}\}\_\{t\}\)\},\(2\)the likelihood ratio of base to proposal\. Models trained for CFG are convenient here because they provide both transitions\. The soft value of[Equation1](https://arxiv.org/html/2611.17462#S2.E1)is then an expectation underqcq\_\{c\}weighted byη\\eta,
vt\(𝐱t\)=1λlog𝔼𝐱t−1∼qc\(⋅∣𝐱t\)\[η\(𝐱t−1\)exp\(λvt−1\(𝐱t−1\)\)\],v\_\{t\}\(\{\\mathbf\{x\}\}\_\{t\}\)=\\frac\{1\}\{\\lambda\}\\log\\mathbb\{E\}\_\{\{\\mathbf\{x\}\}\_\{t\-1\}\\sim q\_\{c\}\(\\cdot\\mid\{\\mathbf\{x\}\}\_\{t\}\)\}\\\!\\big\[\\eta\(\{\\mathbf\{x\}\}\_\{t\-1\}\)\\exp\\\!\\big\(\\lambda v\_\{t\-1\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\)\\big\)\\big\],\(3\)so backups remain consistent with the unconditional base dynamics\.
##### Importance correction in selection\.
Selection among existing children must be corrected for the same reason, since the children were drawn fromqcq\_\{c\}\. We therefore select a child𝐳∈𝒞\(𝐱t\)\{\\mathbf\{z\}\}\\in\\mathcal\{C\}\(\{\\mathbf\{x\}\}\_\{t\}\)with probability
∝η\(𝐳\)exp\(λv^\(𝐳\)\),\\propto\\;\\eta\(\{\\mathbf\{z\}\}\)\\,\\exp\\\!\\big\(\\lambda\\,\\hat\{v\}\(\{\\mathbf\{z\}\}\)\\big\),\(4\)which is self\-normalized importance sampling of the aligned base transitionpθ\(𝐱t−1∣𝐱t\)exp\(λvt−1\(𝐱t−1\)\)p\_\{\\theta\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\\mid\{\\mathbf\{x\}\}\_\{t\}\)\\exp\(\\lambda v\_\{t\-1\}\(\\rvx\_\{t\-1\}\)\)\. Withoutη\\eta, selection would be biased towardqcq\_\{c\}and the tree would converge to a target that depends on the proposal\.
### 3\.2Adaptation 2: Multi\-Objective Diffusion Tree Sampling \(MO\-DTS\)
Many materials problems involve*multiple*competing goals\. Let𝐫\(𝐱\)≔\[r1\(𝐱\),…,rm\(𝐱\)\]\{\\mathbf\{r\}\}\(\{\\mathbf\{x\}\}\)\\coloneqq\[r\_\{1\}\(\{\\mathbf\{x\}\}\),\\dots,r\_\{m\}\(\{\\mathbf\{x\}\}\)\]be a vector ofmmobjectives to be maximized; a point𝐱⋆\{\\mathbf\{x\}\}^\{\\star\}*dominates*𝐱\{\\mathbf\{x\}\}if it is at least as good on all objectives and strictly better on one, and the set of non\-dominated points is the*Pareto front*\. Multi\-objective optimization can be tackled by taking a*preference*over objectivesw∈Δm−1w\\in\\Delta^\{m\-1\}that encodes a trade\-off, and a*scalarization*, which is a maps:ℝm×Δm−1→ℝs:\{\\mathbb\{R\}\}^\{m\}\\crossproduct\\Delta^\{m\-1\}\\to\{\\mathbb\{R\}\}that turns a reward vector into a scalar according toww\. The simplest choice is linear scalarizations\(𝐚∣w\)≔w⊤𝐚s\(\{\\mathbf\{a\}\}\\mid w\)\\coloneqq w^\{\\top\}\{\\mathbf\{a\}\}, although other choices are possible \(see Appendix[E](https://arxiv.org/html/2611.17462#A5)\)\. For a fixedwwthis defines the scalar rewardr\(𝐱∣w\)≔s\(𝐫\(𝐱\)∣w\)r\(\{\\mathbf\{x\}\}\\mid w\)\\coloneqq s\(\{\\mathbf\{r\}\}\(\{\\mathbf\{x\}\}\)\\mid w\), and the multi\-objective problem is reduced to a set of single\-objective problems:max𝐱r\(𝐱∣w\)\\max\_\{\{\\mathbf\{x\}\}\}r\(\{\\mathbf\{x\}\}\\mid w\)\.
##### Sampling from the Pareto front\.
Given objectives𝐫\(𝐱\)∈ℝm\{\\mathbf\{r\}\}\(\{\\mathbf\{x\}\}\)\\in\{\\mathbb\{R\}\}^\{m\}and a scalarizedr\(𝐱∣w\)r\(\{\\mathbf\{x\}\}\\mid w\), we define the target of*sampling from the Pareto front*as first sampling a preferencew∼p\(w\)w\\sim p\(w\)wherep\(w\)p\(w\)is a user\-chosen distribution over preferences \(e\.g\., Dirichlet on the simplex\), then sampling𝐱0\{\\mathbf\{x\}\}\_\{0\}from the corresponding reward\-tilted base distribution:
π⋆\(𝐱0∣w\)=pθ\(𝐱0\)exp\(λr\(𝐱0∣w\)\)/Z\(w\),Z\(w\)=𝔼𝐱0∼pθ\[exp\(λr\(𝐱0∣w\)\)\]\.\\pi^\{\\star\}\(\{\\mathbf\{x\}\}\_\{0\}\\mid w\)=p\_\{\\theta\}\(\{\\mathbf\{x\}\}\_\{0\}\)\\exp\(\\lambda r\(\\rvx\_0\\mid w\)\)/Z\(w\),\\qquad Z\(w\)=\\mathbb\{E\}\_\{\{\\mathbf\{x\}\}\_\{0\}\\sim p\_\{\\theta\}\}\\left\[\\exp\(\\lambda r\(\\rvx\_0\\mid w\)\)\\right\]\.Marginalizing outwwyields a mixture of reward\-tilted distributions,
π⋆\(𝐱0\)=pθ\(𝐱0\)𝔼w∼p\(w\)\[exp\(λr\(𝐱0∣w\)\)/Z\(w\)\],\\pi^\{\\star\}\(\{\\mathbf\{x\}\}\_\{0\}\)=p\_\{\\theta\}\(\{\\mathbf\{x\}\}\_\{0\}\)\\,\\mathbb\{E\}\_\{w\\sim p\(w\)\}\\left\[\\exp\(\\lambda r\(\\rvx\_0\\mid w\)\)/Z\(w\)\\right\],\(5\)which assigns high probability to candidates that are strong under*some*trade\-off, while remaining plausible under the base model\.
##### A single shared tree\.
When lifting a single\-objective method to the multi\-objective setting, treating each scalarized objectiver\(⋅∣w\)r\(\\cdot\\mid w\)in isolation is highly inefficient: a key design requirement is to share computation across preferences\. It is non\-trivial to adapt arbitrary alignment methods \(Best\-of\-NN, SMC\([Singhal et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib28)\)\) under this requirement\. DTS\([Jain et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib15)\)can be extended rather directly to approximately sample from[Equation5](https://arxiv.org/html/2611.17462#S3.E5)while sharing computation: maintain a separate value estimate per objective during backup, and use a*scalarized value*based on these estimates during selection\.
Each terminal rollout produces a vector reward𝐫\(𝐱0\)\{\\mathbf\{r\}\}\(\{\\mathbf\{x\}\}\_\{0\}\), and we maintain at each node a value vector𝐯t\(𝐱t\)=\[vt,1\(𝐱t\),…,vt,m\(𝐱t\)\]\{\\mathbf\{v\}\}\_\{t\}\(\{\\mathbf\{x\}\}\_\{t\}\)=\[v\_\{t,1\}\(\{\\mathbf\{x\}\}\_\{t\}\),\\dots,v\_\{t,m\}\(\{\\mathbf\{x\}\}\_\{t\}\)\], with eachvt,iv\_\{t,i\}updated via the soft backup of[Equation1](https://arxiv.org/html/2611.17462#S2.E1)applied to objectiveii\. Each tree traversal first samplesw∼p\(w\)w\\sim p\(w\), then computes a scalarized value by applying the scalarization to the node’s value vector,vt\(𝐱t\|w\)≔s\(𝐯\(𝐱t\)∣w\)v\_\{t\}\(\{\\mathbf\{x\}\}\_\{t\}\|w\)\\coloneqq s\(\{\\mathbf\{v\}\}\(\{\\mathbf\{x\}\}\_\{t\}\)\\mid w\)\.
##### MO\-DTS and conditioning\.
When available, we can use a conditional diffusion model where the condition is a function of the preference\. That is, for each preferenceww, our diffusion model samples fromqw\(𝐱t−1∣𝐱t\):=pθ\(𝐱t−1∣𝐱t,c\(w\)\)q\_\{w\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\\mid\{\\mathbf\{x\}\}\_\{t\}\):=p\_\{\\theta\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\\mid\{\\mathbf\{x\}\}\_\{t\},c\(w\)\)wherec\(w\)c\(w\)is a mapping from preferences to conditions that would help maximizer\(⋅\|w\)r\(\\cdot\|w\)\. In practice, this can mean conditioning the diffusion model on the features of samples in the training set that maximizer\(⋅\|w\)r\(\\cdot\|w\), as they represent reasonable targets\.
At each pass through the tree,wwcan be sampled anew\. The sampledwwnot only affects the choice of children while moving down the tree from the root, but also affects how new leaf nodes are generated\. However, due to the importance correction of[Section3\.1](https://arxiv.org/html/2611.17462#S3.SS1), the target distribution does not depend on the choice ofc\(w\)c\(w\), and values𝐯t\(𝐱t\)\{\\mathbf\{v\}\}\_\{t\}\(\{\\mathbf\{x\}\}\_\{t\}\)do not depend onww, so the whole tree remains consistent\.
##### Per\-objective values as an upper\-confidence approximation\.
Scalarizing at selection time is an approximation: the exact soft value of the scalarized reward,vt⋆\(𝐱t∣w\)≔1λlog𝔼\[exp\(λr\(𝐱0∣w\)\)∣𝐱t\]v\_\{t\}^\{\\star\}\(\{\\mathbf\{x\}\}\_\{t\}\\mid w\)\\coloneqq\\tfrac\{1\}\{\\lambda\}\\log\\mathbb\{E\}\[\\exp\(\\lambda\\,r\(\\rvx\_0\\mid w\)\)\\mid\{\\mathbf\{x\}\}\_\{t\}\], would require a separate backup for everyww, whereas MO\-DTS backs up each objective once and usesvt\(𝐱t∣w\)=s\(𝐯t\(𝐱t\)∣w\)v\_\{t\}\(\{\\mathbf\{x\}\}\_\{t\}\\mid w\)=s\(\{\\mathbf\{v\}\}\_\{t\}\(\{\\mathbf\{x\}\}\_\{t\}\)\\mid w\)\. In Appendix[F](https://arxiv.org/html/2611.17462#A6)we show that the surrogate is optimistic: it upper\-boundsvt⋆v\_\{t\}^\{\\star\}for linear scalarizations at everyλ\\lambda, and coincides withvt⋆v\_\{t\}^\{\\star\}att=0t=0\. The gap is governed by how far the objectives are from co\-varying under the denoising distribution, so selection is most optimistic about nodes whose objectives trade off; MO\-DTS therefore provides a search heuristic for[Equation5](https://arxiv.org/html/2611.17462#S3.E5)rather than an exact sampler for everyww\.
### 3\.3Coarse branching and parallelization
Following[Jain et al\. \(2025\)](https://arxiv.org/html/2611.17462#bib.bib15), we restrict branching to a set of*branching times*𝒯⊆\{1,…,T\+1\}\\mathcal\{T\}\\subseteq\\\{1,\\dots,T\+1\\\}\. A node𝐱t\{\\mathbf\{x\}\}\_\{t\}is*expandable*ift∈𝒯t\\in\\mathcal\{T\}and progressive widening permits,\|𝒞\(𝐱t\)\|<CN\(𝐱t\)α\|\\mathcal\{C\}\(\{\\mathbf\{x\}\}\_\{t\}\)\|<C\\,N\(\{\\mathbf\{x\}\}\_\{t\}\)^\{\\alpha\}, where𝒞\\mathcal\{C\}is the children of a node,NNis the number of visits so far, andCCandα\\alphaare hyperparameters\. Nodes at other times are created only during rollouts and have a single child, so selection passes through them deterministically\. In practice, we set the branching times to be regularly spaced everyLLtimesteps\. At a matched number of function evaluations \(NFEs\), DTS produces more samples due to trajectory reuse\. The search also admits a batched implementation:BBtraversals advance synchronously from the root node, and all rollout steps at the samettare batched into one call to the sampler, so each pass through the tree generatesBBleaf nodes\. Throughput is thus at least as high as generatingBBindependent trajectories as in best\-of\-NN\.
## 4Application to Materials Discovery
We instantiate the framework on four crystal\-generation problems chosen to demonstrate both technical contributions across two generative models\. First, we demonstrate how inference\-time steering with regular DTS improves models’ ability to generate stable crystals\. Next, symmetry\-conditioned generation targeting rare space groups \([Section4\.2](https://arxiv.org/html/2611.17462#S4.SS2)\) demonstrates the benefit of using conditional models and exercises the importance\-corrected guided proposals of[Section3\.1](https://arxiv.org/html/2611.17462#S3.SS1)\. A multi\-objective experiment targeting the bandgap and dielectric constant tradeoff demonstrates the utility of building a single tree across different preferences in[Section4\.3](https://arxiv.org/html/2611.17462#S4.SS3)\. We examine the effect of a mismatched proposal distribution and the use of importance\-correction in[Section4\.4](https://arxiv.org/html/2611.17462#S4.SS4), targeting high bulk modulus materials\.
### 4\.1De novo stability search
A central challenge in de novo crystal generation is producing*thermodynamically stable*structures, quantified by the*energy above the convex hull*\(EhullE\_\{\\text\{hull\}\}\): for a given composition, the convex hull is the lower envelope of formation energies across all known competing phases\.Ehull=0E\_\{\\text\{hull\}\}\{=\}0is stable; small positiveEhullE\_\{\\text\{hull\}\}\(typically below0\.10\.1eV/atom\) is*metastable*\. Current generative models produce stable samples only a small fraction of the time, bottlenecking their usefulness\([Betala et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib45)\)\.EhullE\_\{\\text\{hull\}\}depends on the composition through the \(non\-differentiable\) hull construction, ruling out gradient\-based guidance\.
We define a single\-objective reward based onEhullE\_\{\\text\{hull\}\}as predicted by MatterSim\([Yang et al\., 2024](https://arxiv.org/html/2611.17462#bib.bib29)\), with the Alex\-MP hull from[Zeni et al\. \(2025\)](https://arxiv.org/html/2611.17462#bib.bib1)\. This setup reduces to DTS with scalar reward and serves as a controlled assessment of tree\-search\-based steering vs\. best\-of\-NNsampling from the same generative model\. We evaluate using the common summary metric of the rate of*Stable, Unique, Novel*\(SUN\) structures\([Zeni et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib1)\), where uniqueness and novelty with respect to the training set are determined using Pymatgen’sStructureMatcher\([Ong et al\., 2013](https://arxiv.org/html/2611.17462#bib.bib21)\)\.
##### Results with MatterGen\.
We compare three configurations at matched NFE budgets: standard sampling from MatterGen \(best\-of\-NN\), DTS with null rewardr\(𝐱0\)=0r\(\{\\mathbf\{x\}\}\_\{0\}\)=0, and DTS with the rewardr\(𝐱0\)=−Ehull\(𝐱0\)r\(\{\\mathbf\{x\}\}\_\{0\}\)=\-E\_\{\\text\{hull\}\}\(\{\\mathbf\{x\}\}\_\{0\}\)\. The null\-reward control isolates the effect of tree search itself \(trajectory reuse\) from the stability reward; we also vary the progressive widening parameterα\\alpha, which controls tree width\. Results averaged over three seeds are shown in[Table1](https://arxiv.org/html/2611.17462#S4.T1)\. Tree search improves over best\-of\-NNin absolute SUN count at the same NFE budget\. The energy reward consistently lowers averageEhullE\_\{\\text\{hull\}\}and improves overall stability rate\. We also report the number of unique chemical systems among the resulting samples to show that diversity is retained\.
Table 1:Single\-objective stability search with MatterGen at matched NFE budget, averaged over three seeds\. “Null” is DTS withr\(𝐱0\)=0r\(\{\\mathbf\{x\}\}\_\{0\}\)=0;α\\alphais the progressive widening parameter\.α\\alphaReward\# Structures\# SUN% StableSUN Unique SystemsAvg Ehull\{\}\_\{\\text\{hull\}\}/atom0\.1Null5050±\\pm291311±\\pm3966\.5±\\pm1\.1912\.3±\\pm33\.70\.091±\\pm0\.004−Ehull\-E\_\{\\text\{hull\}\}5226±\\pm421432±\\pm7375\.7±\\pm0\.4966±\\pm120\.073±\\pm0\.0000\.3Null3500±\\pm01095±\\pm5565\.3±\\pm1\.5904\.3±\\pm29\.40\.1±\\pm0\.004−Ehull\-E\_\{\\text\{hull\}\}3500±\\pm01209±\\pm3170\.2±\\pm0\.9971\.3±\\pm29\.00\.085±\\pm0\.0020\.5Null3000±\\pm0974±\\pm2766\.6±\\pm1\.4894\.7±\\pm27\.40\.090±\\pm0\.003−Ehull\-E\_\{\\text\{hull\}\}3000±\\pm01087±\\pm2368\.6±\\pm1\.3979\.7±\\pm15\.20\.087±\\pm0\.002MatterGen Baseline2500868±\\pm1168\.4±\\pm0\.1848\.3±\\pm10\.90\.086±\\pm0\.000
##### Results with DiffCSP and inference\-time scaling\.
DiffCSP is roughly20×20\\timesfaster per sample than MatterGen, which lets us study how DTS scales with the NFE budget from10510^\{5\}to10710^\{7\}total NFEs \([Figure2](https://arxiv.org/html/2611.17462#S4.F2)\)\. Stability rate increases with NFE budget under theEhullE\_\{\\text\{hull\}\}reward, while the null\-reward control and best\-of\-NNbaselines remain roughly flat – a clear and growing inference\-time scaling benefit\. Uniqueness and novelty decline slightly as compute increases, reflecting concentration of rollouts on high\-reward regions\.
We also compare against three inference\-time steering baselines: SVDD\-PM\([Li et al\., 2024](https://arxiv.org/html/2611.17462#bib.bib43)\)\(which is equivalent to TreeG\-SC\([Guo et al\., 2026](https://arxiv.org/html/2611.17462#bib.bib44)\)with A=1\), TreeG\-SD\([Guo et al\., 2026](https://arxiv.org/html/2611.17462#bib.bib44)\)and rollout\-based SMC\([Singhal et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib28)\)with two and four resampling events\. All methods use the same pretrained DiffCSP model,EhullE\_\{\\text\{hull\}\}reward, and approximately matched 1M and 10M NFE budgets; further details are included in Appendix[G\.1](https://arxiv.org/html/2611.17462#A7.SS1)\. The results, shown in[Table2](https://arxiv.org/html/2611.17462#S4.T2), demonstrate that DTS outperforms all baselines at generating stable, unique and novel crystals\.
Figure 2:Inference\-time scaling of DTS with DiffCSP across NFE budgets, compared with the control null\-reward, and Best\-of\-N baselines\. Standard deviations are computed across 3 seeds\.Table 2:DiffCSP results at matched 1M and 10M NFE budgets\.MethodStructuresStability \(%\)↑\\uparrowUniqueness \(%\)↑\\uparrowNovelty \(%\)↑\\uparrowSUN \(%\)↑\\uparrowMeanEhullE\_\{\\text\{hull\}\}↓\\downarrowMatched 1M NFE budgetBest\-of\-NN1,00030\.7±0\.930\.7\\pm 0\.999\.6±0\.199\.6\\pm 0\.189\.4±0\.389\.4\\pm 0\.321\.0±1\.321\.0\\pm 1\.30\.314±0\.0310\.314\\pm 0\.031SMC \(2 resamples\)55430\.7±0\.530\.7\\pm 0\.584\.9±1\.484\.9\\pm 1\.488\.8±1\.088\.8\\pm 1\.016\.3±0\.616\.3\\pm 0\.60\.315±0\.0050\.315\\pm 0\.005SMC \(4 resamples\)33229\.7±3\.229\.7\\pm 3\.287\.3±0\.987\.3\\pm 0\.989\.3±4\.589\.3\\pm 4\.516\.7±6\.216\.7\\pm 6\.20\.296±0\.0200\.296\\pm 0\.020SVDD\-PM / TreeG\-SC72534\.2±1\.834\.2\\pm 1\.899\.4±0\.299\.4\\pm 0\.288\.3±1\.188\.3\\pm 1\.123\.2±2\.223\.2\\pm 2\.20\.724±0\.3880\.724\\pm 0\.388TreeG\-SD1,00033\.5±4\.833\.5\\pm 4\.899\.4±0\.199\.4\\pm 0\.189\.2±1\.289\.2\\pm 1\.223\.2±3\.423\.2\\pm 3\.40\.317±0\.0340\.317\\pm 0\.034DTS \(Null Reward\)1,35032\.0±1\.532\.0\\pm 1\.599\.3±0\.199\.3\\pm 0\.189\.6±0\.889\.6\\pm 0\.822\.5±2\.122\.5\\pm 2\.10\.399±0\.1490\.399\\pm 0\.149DTS1,35036\.8±4\.9\\mathbf\{36\.8\\pm 4\.9\}99\.2±0\.199\.2\\pm 0\.187\.5±1\.487\.5\\pm 1\.425\.1±3\.6\\mathbf\{25\.1\\pm 3\.6\}0\.267±0\.019\\mathbf\{0\.267\\pm 0\.019\}Matched 10M NFE budgetBest\-of\-NN10,00031\.1±0\.331\.1\\pm 0\.396\.9±0\.196\.9\\pm 0\.189\.2±0\.2\\mathbf\{89\.2\\pm 0\.2\}20\.3±0\.120\.3\\pm 0\.10\.435±0\.0350\.435\\pm 0\.035SMC \(2 resamples\)5,54931\.3±0\.131\.3\\pm 0\.184\.5±1\.184\.5\\pm 1\.188\.6±1\.488\.6\\pm 1\.416\.8±0\.416\.8\\pm 0\.40\.361±0\.0130\.361\\pm 0\.013SMC \(4 resamples\)3,32831\.4±2\.031\.4\\pm 2\.084\.1±0\.584\.1\\pm 0\.588\.5±1\.388\.5\\pm 1\.316\.4±0\.716\.4\\pm 0\.70\.320±0\.028\\mathbf\{0\.320\\pm 0\.028\}SVDD\-PM7,24633\.3±0\.933\.3\\pm 0\.997\.4±0\.1\\mathbf\{97\.4\\pm 0\.1\}88\.2±0\.088\.2\\pm 0\.022\.0±0\.822\.0\\pm 0\.80\.574±0\.1930\.574\\pm 0\.193TreeG\-SD10,00031\.6±0\.531\.6\\pm 0\.596\.9±0\.396\.9\\pm 0\.389\.1±0\.389\.1\\pm 0\.320\.8±0\.020\.8\\pm 0\.00\.425±0\.0330\.425\\pm 0\.033DTS \(Null Reward\)15,35031\.1±0\.831\.1\\pm 0\.895\.5±0\.295\.5\\pm 0\.289\.1±0\.589\.1\\pm 0\.519\.7±0\.119\.7\\pm 0\.10\.337±0\.0360\.337\\pm 0\.036DTS15,45041\.6±6\.8\\mathbf\{41\.6\\pm 6\.8\}90\.4±3\.290\.4\\pm 3\.284\.5±3\.484\.5\\pm 3\.424\.1±2\.2\\mathbf\{24\.1\\pm 2\.2\}0\.331±0\.1580\.331\\pm 0\.158
### 4\.2Symmetry\-Conditioned Generation
We test whether DTS can amplify a property\-conditional model on a rare, non\-differentiable target: matching a specific space group\. Both DTS and the best\-of\-NNbaseline use the same space\-group\-conditional MatterGen variant; the reward is the indicatorr\(𝐱0\)=𝟙\[SG\(𝐱0\)=g⋆\]r\(\{\\mathbf\{x\}\}\_\{0\}\)=\\mathbbm\{1\}\[\\mathrm\{SG\}\(\{\\mathbf\{x\}\}\_\{0\}\)=g^\{\\star\}\], withSG\(⋅\)\\mathrm\{SG\}\(\\cdot\)computed using pymatgen\([Ong et al\., 2013](https://arxiv.org/html/2611.17462#bib.bib21)\)\. We focus here on space group 204 \(Im3¯Im\\bar\{3\}\), which appears in the training set with frequency∼\\sim0\.2%; results for additional rare target groups are in Appendix[G\.5](https://arxiv.org/html/2611.17462#A7.SS5)\.
[Figure3](https://arxiv.org/html/2611.17462#S4.F3)shows cumulative SUN∩\\capSG 204 counts versus NFE budget\. DTS produces roughly an order of magnitude more on\-target SUN structures than best\-of\-NN, with the gap widening as compute increases\. The unconditional MatterGen, with or without DTS, generates zero structures in SG 204 across the same budget\. This contrast highlights the role of each component: the conditional proposal is*necessary*for DTS to access the targeted mode at all \(the unconditional base places negligible mass there\), but is not*sufficient*to produce on\-target samples at useful rates\. Search compounds the proposal by reusing partial trajectories that lead into the target region\.
Figure 3:Cumulative count of SUN structures matching target space group 204 \(Im3¯Im\\bar\{3\}\) versus NFE budget\. Both methods use the same space\-group\-conditional MatterGen variant\.
### 4\.3Multi\-objective: Dielectric Materials
We consider a multi\-objective setting from[Riebesell et al\. \(2024\)](https://arxiv.org/html/2611.17462#bib.bib11): simultaneously maximizing bandgap and dielectric constant\. These properties are approximately inversely related\([Penn, 1962](https://arxiv.org/html/2611.17462#bib.bib23);[Robertson, 2005](https://arxiv.org/html/2611.17462#bib.bib24);[Onishi and Fu, 2024](https://arxiv.org/html/2611.17462#bib.bib22)\), inducing a non\-convex Pareto front\. This task also demonstrates the conditional\-model adaptation: MatterGen can be fine\-tuned to condition jointly on \(bandgap, dielectric constant\), used here as a guided proposal inside DTS\.
##### Setup\.
Both properties are predicted with ALIGNN\([Choudhary and DeCost, 2021](https://arxiv.org/html/2611.17462#bib.bib18)\)\. We use a MatterGen base model fine\-tuned on labels from this predictor, and run multi\-objective DTS at 2\.5M NFEs\. As a baseline, we condition MatterGen on five target points evenly spaced along the dataset Pareto front and take the union of generated samples at matched NFE budget\. We evaluate after filtering to SUN structures, reporting Pareto\-front size, hypervolume \(HV\),R2R\_\{2\}\([Hansen and Jaszkiewicz, 1998](https://arxiv.org/html/2611.17462#bib.bib31)\), and generational distance plus \(GD\+\)\([Ishibuchi et al\., 2015](https://arxiv.org/html/2611.17462#bib.bib30)\)against the dataset Pareto front\. We also include threshold counts: number of structures whose bandgap×\\timesdielectric constant exceeds fixed thresholds \(30, 60, 120, 240\) – these values correspond to linear scalarization in log space\.
##### Results\.
[Figure4](https://arxiv.org/html/2611.17462#S4.F4)plots SUN samples and[Table3](https://arxiv.org/html/2611.17462#S4.T3)summarizes metrics\. MO\-DTS outperforms the baseline at matched NFE budget for most metrics\.
Figure 4:Dielectric constant vs\. bandgap for SUN samples at 2\.5M NFEs \(∼\\sim2500 total samples per method\)\. Diagonal lines indicate constant values of bandgap×\\timesdielectric, and X markers represent conditioning targets\.Table 3:Summary metrics for the bandgap vs\. dielectric constant experiment\.\|PF\|\|PF\|= Pareto front size\. Entries are mean±\\pmstd across seeds\. Threshold columns count structures where bandgap×\\timesdielectric≥\\geqthreshold\.MethodNN\|PF\|\|PF\|N≥30N\_\{\\geq 30\}N≥60N\_\{\\geq 60\}N≥120N\_\{\\geq 120\}N≥240N\_\{\\geq 240\}HV↑\\uparrowGD\+↓\\downarrowR2↓\\downarrowTraining set27,136204,1801,9205261090\.3090\.0000\.149MatterGen \(cond\.\)665±\\pm014±\\pm1262±\\pm3138±\\pm628±\\pm24±\\pm00\.118±\\pm0\.0110\.098±\\pm0\.0240\.292±\\pm0\.002MO\-DTS645±\\pm4016±\\pm1402±\\pm54234±\\pm2767±\\pm1217±\\pm40\.173±\\pm0\.0570\.054±\\pm0\.0150\.338±\\pm0\.040
### 4\.4Density/bulk modulus: partially aligned reward
The previous experiments used either unconditional models or models fine\-tuned for the target reward\. Conditional DTS is more general, since importance sampling decouples the proposal from the target\. To show this, we repurpose an existing MatterGen model conditioned on density as a proposal, and set the target to be generating crystals with a high bulk modulus \(300 GPa\)\. In the training set the two correlate \(Pearsonr=0\.512r=0\.512\): crystals with density≥15\\geq 15g/cm3are rare \(<2%\{<\}2\\%\) but about20×20\\timesmore likely than average to have bulk modulus≥300\\geq 300GPa, so conditioning on a density of 15 g/cm3is a useful but imperfect proxy\. The mismatch is that training crystals near the bulk modulus target span a broad density range, so without correction the proposal should raise yield while distorting the density of the resulting crystals; IS should remove the distortion\.
We compare four arms \([Table4](https://arxiv.org/html/2611.17462#S4.T4)\): DTS with an unconditional model; the density\-conditioned proposal alone \(best\-of\-NN\); conditional DTS without IS; and conditional DTS with IS\. All DTS arms use rewardr\(𝐱\)=−\|K\(𝐱\)−300\|/400r\(\{\\mathbf\{x\}\}\)=\-\\lvert K\(\{\\mathbf\{x\}\}\)\-300\\rvert/400withKKthe bulk modulus predicted by ALIGNN\([Choudhary and DeCost, 2021](https://arxiv.org/html/2611.17462#bib.bib18)\)in GPa,λ=20\\lambda=20,α=0\.3\\alpha=0\.3, and guidance strengthγ=1\\gamma=1for the conditional proposal; each arm runs10610^\{6\}NFEs over 3 seeds, with metrics computed on SUN crystals only\. We report the number of SUN crystals within 30 GPa of the target, the mean absolute error to the target, and the density distortion: the Wasserstein distanceW1W\_\{1\}between the density distribution of generated crystals within 30 GPa of the target and that of training crystals in the same range\.
Table 4:Results for bulk modulus experiment, ablating conditioning and importance sampling correction \(IS\)\.ModelDTSISSUN crystalswithin 30 GPa↑\\uparrowBulk MAEto 300 GPa↓\\downarrowNear\-target densityW1W\_\{1\}to training↓\\downarrowUnconditional✓✗237108\.232\.80Density\-Conditioned✗✗61131\.024\.72Density\-Conditioned✓✗30262\.824\.31Density\-Conditioned✓✓26070\.332\.15The mismatched proposal is useful only together with search: on its own it yields four times fewer near\-target SUN crystals than unconditional DTS, since it shifts samples toward dense crystals but does not select for bulk modulus\. Conditional DTS without IS outperforms both, but the proposal creates biased samples: the resulting near\-target crystals are much denser than training crystals in the same range\. Importance correction removes this bias while keeping most of the gain in yield\. An existing conditional model can therefore serve as a proposal for a new reward, with the output governed by the reward rather than by the proposal\.
## 5Related Work
##### Inference\-time steering of diffusion models\.
Score\-modifying approaches include classifier\-based\([Dhariwal and Nichol, 2021](https://arxiv.org/html/2611.17462#bib.bib7)\)and classifier\-free\([Ho and Salimans, 2022](https://arxiv.org/html/2611.17462#bib.bib8)\)guidance; both approximate reward\-tilted sampling but are biased and require differentiable rewards\. Particle\-based methods treat denoising as SMC\([Singhal et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib28)\)\. Methods such as SVDD\([Li et al\., 2024](https://arxiv.org/html/2611.17462#bib.bib43)\)and TreeG\([Guo et al\., 2026](https://arxiv.org/html/2611.17462#bib.bib44)\)perform a local search: they branch each denoising step into several candidates and keep the best by a value estimate\. The closest line to us organizes intermediate states into a tree: Diffusion Tree Sampling\([Jain et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib15)\)formalizes this for continuous diffusion as MCTS with soft\-Bellman backups and is the basis for our method\. PepTune\([Tang et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib13)\)develops Monte Carlo Tree Guidance for discrete diffusion in multi\-objective peptide design with a Pareto\-dominance reward\. Our approach operates on continuous diffusion with scalarization\-based per\-objective backups, and composes with conditional models via importance correction\. A separate strand fine\-tunes for specific objectives via conditional training\([Zeni et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib1)\)or RL\([Black et al\., 2023](https://arxiv.org/html/2611.17462#bib.bib41);[Chen et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib14)\); our work leaves the model fixed and composes with such fine\-tuned variants when available\.
##### Multi\-objective generation\.
Multi\-objective generation has a long history in evolutionary algorithms and Bayesian optimization\([Daulton et al\., 2020](https://arxiv.org/html/2611.17462#bib.bib39)\); multi\-objective GFlowNets\([Jain et al\., 2023](https://arxiv.org/html/2611.17462#bib.bib26);[Zhu et al\., 2023](https://arxiv.org/html/2611.17462#bib.bib40)\)sample from a Pareto\-tilted distribution by conditioning on preferences during training\. Diffusion\-based methods modify the score or velocity field during a single denoising pass \(e\.g\. PROUD\([Yao et al\., 2024](https://arxiv.org/html/2611.17462#bib.bib25)\), ParetoFlow\([Yuan et al\., 2024](https://arxiv.org/html/2611.17462#bib.bib32)\), and Preference\-Guided Diffusion\([Annadani et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib33)\)\), use diffusion as proposals\([Kong et al\., 2024](https://arxiv.org/html/2611.17462#bib.bib34)\), or wrap diffusion in outer BO loops\([Gruver et al\., 2023](https://arxiv.org/html/2611.17462#bib.bib35)\)\. Our method differs by using tree\-based search over pretrained continuous diffusion models, using per\-objective values and backups, yielding the target of[Equation5](https://arxiv.org/html/2611.17462#S3.E5)and admitting importance\-corrected use of conditional models\.
## 6Discussion
We adapted Diffusion Tree Sampling to crystal generative models in two ways: importance\-corrected guided proposals, which let existing property\-conditional models steer the search without changing its target, and a multi\-objective extension, in which per\-objective value backups and a preference sampled per traversal let a single tree search over the Pareto front of competing properties\. DTS alone improved stability and scaled with compute, beating best\-of\-NN, but could not reach rare regions in the space of possible materials\. This access can be provided by using conditional proposals, but at the cost of tying the output to whatever the proposal was tuned for\. This dependence can be solved via importance correction, enabling cases where the proposal and reward are mismatched, or where we want a single tree to capture multiple different preferences\.
The main caveat is reward fidelity\. All reported rewards use learned surrogates, so search can amplify systematic predictor errors; DFT validation and synthesis remain essential next steps\. Importance correction is also most reliable when guided and base models share a backbone; separately trained proposals may require stronger weight control\.
More broadly, this method fits the foundation\-model setting: a pretrained generative prior, adapted at inference time to objectives that it has not been specifically trained for\. This setup benefits from inference\-time search, where more compute is spent per candidate to produce higher\-quality samples\. Materials discovery is a natural fit because generation is cheap next to DFT validation and synthesis, but the same technique could also be applied to other AI for science settings, such as protein and small\-molecule design\.
## Acknowledgment
This project is in part funded by CIFAR AI chairs, NSERC Discovery grant and IVADO\. D\.L\.’s research is additionally supported by FRQNT\. Computational resources are provided by Mila and the Digital Research Alliance of Canada\.
## References
- Y\. Annadani, S\. Belakaria, S\. Ermon, S\. Bauer, and B\. E\. EngelhardtPreference\-guided diffusion for multi\-objective offline optimization\.arXiv preprint arXiv:2503\.17299\.Cited by:[§5](https://arxiv.org/html/2611.17462#S5.SS0.SSS0.Px2.p1.1)\.
- Artset al\.\(2023\)M\. Arts, V\. Garcia Satorras, C\. Huang, D\. Zugner, M\. Federici, C\. Clementi, F\. Noé, R\. Pinsler, and R\. van den BergTwo for one: diffusion models and force fields for coarse\-grained molecular dynamics\.Journal of Chemical Theory and Computation19\(18\),pp\. 6151–6159\.Cited by:[Appendix B](https://arxiv.org/html/2611.17462#A2.p2.1)\.
- Betalaet al\.\(2025\)S\. Betala, S\. P\. Gleason, A\. Ramlaoui, A\. Xu, G\. Channing, D\. Levy, C\. Fourrier, N\. Kazeev, C\. K\. Joshi, S\. Kaba,et al\.Lemat\-genbench: a unified evaluation framework for crystal generative models\.arXiv preprint arXiv:2512\.04562\.Cited by:[§4\.1](https://arxiv.org/html/2611.17462#S4.SS1.p1.1)\.
- Blacket al\.\(2023\)K\. Black, M\. Janner, Y\. Du, I\. Kostrikov, and S\. LevineTraining diffusion models with reinforcement learning\.arXiv preprint arXiv:2305\.13301\.Cited by:[§5](https://arxiv.org/html/2611.17462#S5.SS0.SSS0.Px1.p1.1)\.
- Bommasaniet al\.\(2021\)R\. Bommasani, D\. A\. Hudson, E\. Adeli, R\. Altman, S\. Arora, S\. von Arx, M\. S\. Bernstein, J\. Bohg, A\. Bosselut, E\. Brunskill,et al\.On the opportunities and risks of foundation models\.arXiv preprint arXiv:2108\.07258\.Cited by:[§1](https://arxiv.org/html/2611.17462#S1.p1.1)\.
- Bowman Jr \(1976\)V\. J\. Bowman JrOn the relationship of the tchebycheff norm and the efficient frontier of multiple\-criteria objectives\.InMultiple Criteria Decision Making: Proceedings of a Conference Jouy\-en\-Josas, France May 21–23, 1975,pp\. 76–86\.Cited by:[Appendix E](https://arxiv.org/html/2611.17462#A5.p1.1)\.
- Chenet al\.\(2025\)J\. Chen, J\. Guo, and P\. SchwallerMatInvent: reinforcement learning for 3d crystal diffusion generation\.InAI for Accelerated Materials Design\-ICLR 2025,Cited by:[§1](https://arxiv.org/html/2611.17462#S1.p2.1),[§2\.1](https://arxiv.org/html/2611.17462#S2.SS1.p2.1),[§5](https://arxiv.org/html/2611.17462#S5.SS0.SSS0.Px1.p1.1)\.
- Choudhary and DeCost \(2021\)K\. Choudhary and B\. DeCostAtomistic line graph neural network for improved materials property predictions\.npj Computational Materials7\(1\),pp\. 185\.Cited by:[§G\.4](https://arxiv.org/html/2611.17462#A7.SS4.p1.1),[§4\.3](https://arxiv.org/html/2611.17462#S4.SS3.SSS0.Px1.p1.1),[§4\.4](https://arxiv.org/html/2611.17462#S4.SS4.p2.1)\.
- Daultonet al\.\(2020\)S\. Daulton, M\. Balandat, and E\. BakshyDifferentiable expected hypervolume improvement for parallel multi\-objective bayesian optimization\.Advances in neural information processing systems33,pp\. 9851–9864\.Cited by:[§5](https://arxiv.org/html/2611.17462#S5.SS0.SSS0.Px2.p1.1)\.
- Debet al\.\(2016\)K\. Deb, K\. Sindhya, and J\. HakanenMulti\-objective optimization\.InDecision sciences,pp\. 161–200\.Cited by:[§1](https://arxiv.org/html/2611.17462#S1.p4.1)\.
- Dhariwal and Nichol \(2021\)P\. Dhariwal and A\. NicholDiffusion models beat gans on image synthesis\.Advances in neural information processing systems34,pp\. 8780–8794\.Cited by:[§1](https://arxiv.org/html/2611.17462#S1.p3.1),[§5](https://arxiv.org/html/2611.17462#S5.SS0.SSS0.Px1.p1.1)\.
- Gruveret al\.\(2023\)N\. Gruver, S\. Stanton, N\. Frey, T\. G\. Rudner, I\. Hotzel, J\. Lafrance\-Vanasse, A\. Rajpal, K\. Cho, and A\. G\. WilsonProtein design with guided discrete diffusion\.Advances in neural information processing systems36,pp\. 12489–12517\.Cited by:[§5](https://arxiv.org/html/2611.17462#S5.SS0.SSS0.Px2.p1.1)\.
- Guoet al\.\(2026\)Y\. Guo, Y\. Yang, H\. Yuan, and M\. WangTraining\-free guidance beyond differentiability: scalable path steering with tree search in diffusion and flow models\.Advances in Neural Information Processing Systems38,pp\. 73343–73384\.Cited by:[§G\.1](https://arxiv.org/html/2611.17462#A7.SS1.p2.1),[§4\.1](https://arxiv.org/html/2611.17462#S4.SS1.SSS0.Px2.p2.1),[§5](https://arxiv.org/html/2611.17462#S5.SS0.SSS0.Px1.p1.1)\.
- Hansen and Jaszkiewicz \(1998\)M\. P\. Hansen and A\. JaszkiewiczEvaluating the quality of approximations to the non\-dominated set\.Cited by:[§4\.3](https://arxiv.org/html/2611.17462#S4.SS3.SSS0.Px1.p1.1)\.
- Ho and Salimans \(2022\)J\. Ho and T\. SalimansClassifier\-free diffusion guidance\.arXiv preprint arXiv:2207\.12598\.Cited by:[§1](https://arxiv.org/html/2611.17462#S1.p3.1),[§2\.1](https://arxiv.org/html/2611.17462#S2.SS1.p2.1),[§5](https://arxiv.org/html/2611.17462#S5.SS0.SSS0.Px1.p1.1)\.
- Ishibuchiet al\.\(2015\)H\. Ishibuchi, H\. Masuda, Y\. Tanigaki, and Y\. NojimaModified distance calculation in generational distance and inverted generational distance\.InInternational conference on evolutionary multi\-criterion optimization,pp\. 110–125\.Cited by:[§4\.3](https://arxiv.org/html/2611.17462#S4.SS3.SSS0.Px1.p1.1)\.
- Jainet al\.\(2023\)M\. Jain, S\. C\. Raparthy, A\. Hernández\-Garcıa, J\. Rector\-Brooks, Y\. Bengio, S\. Miret, and E\. BengioMulti\-objective gflownets\.InInternational conference on machine learning,pp\. 14631–14653\.Cited by:[§5](https://arxiv.org/html/2611.17462#S5.SS0.SSS0.Px2.p1.1)\.
- Jainet al\.\(2025\)V\. Jain, K\. Sareen, M\. Pedramfar, and S\. RavanbakhshDiffusion tree sampling: scalable inference\-time alignment of diffusion models\.arXiv preprint arXiv:2506\.20701\.Cited by:[§1](https://arxiv.org/html/2611.17462#S1.p3.1),[§2\.2](https://arxiv.org/html/2611.17462#S2.SS2.p2.1),[§3\.2](https://arxiv.org/html/2611.17462#S3.SS2.SSS0.Px2.p1.1),[§3\.3](https://arxiv.org/html/2611.17462#S3.SS3.p1.1),[§5](https://arxiv.org/html/2611.17462#S5.SS0.SSS0.Px1.p1.1),[Algorithm 1](https://arxiv.org/html/2611.17462#alg1)\.
- Jiaoet al\.\(2023\)R\. Jiao, W\. Huang, P\. Lin, J\. Han, P\. Chen, Y\. Lu, and Y\. LiuCrystal structure prediction by joint equivariant diffusion\.Advances in Neural Information Processing Systems36,pp\. 17464–17497\.Cited by:[§1](https://arxiv.org/html/2611.17462#S1.p1.1),[§2\.1](https://arxiv.org/html/2611.17462#S2.SS1.p2.1)\.
- Kaplanet al\.\(2020\)J\. Kaplan, S\. McCandlish, T\. Henighan, T\. B\. Brown, B\. Chess, R\. Child, S\. Gray, A\. Radford, J\. Wu, and D\. AmodeiScaling laws for neural language models\.arXiv preprint arXiv:2001\.08361\.Cited by:[§1](https://arxiv.org/html/2611.17462#S1.p1.1)\.
- Konget al\.\(2024\)L\. Kong, Y\. Du, W\. Mu, K\. Neklyudov, V\. De Bortoli, D\. Wu, H\. Wang, A\. Ferber, Y\. Ma, C\. P\. Gomes,et al\.Diffusion models as constrained samplers for optimization with unknown constraints\.arXiv preprint arXiv:2402\.18012\.Cited by:[§5](https://arxiv.org/html/2611.17462#S5.SS0.SSS0.Px2.p1.1)\.
- Liet al\.\(2024\)X\. Li, Y\. Zhao, C\. Wang, G\. Scalia, G\. Eraslan, S\. Nair, T\. Biancalani, S\. Ji, A\. Regev, S\. Levine,et al\.Derivative\-free guidance in continuous and discrete diffusion models with soft value\-based decoding\.arXiv preprint arXiv:2408\.08252\.Cited by:[§G\.1](https://arxiv.org/html/2611.17462#A7.SS1.p2.1),[§4\.1](https://arxiv.org/html/2611.17462#S4.SS1.SSS0.Px2.p2.1),[§5](https://arxiv.org/html/2611.17462#S5.SS0.SSS0.Px1.p1.1)\.
- Li and Scheraga \(1987\)Z\. Li and H\. A\. ScheragaMonte carlo\-minimization approach to the multiple\-minima problem in protein folding\.\.Proceedings of the National Academy of Sciences84\(19\),pp\. 6611–6615\.Cited by:[Appendix B](https://arxiv.org/html/2611.17462#A2.p1.1)\.
- Liet al\.\(2025\)Z\. Li, S\. Liu, B\. Ye, D\. J\. Srolovitz, and T\. WenActive learning for conditional inverse design with crystal generation and foundation atomic models\.arXiv preprint arXiv:2502\.16984\.Cited by:[§1](https://arxiv.org/html/2611.17462#S1.p2.1),[§2\.1](https://arxiv.org/html/2611.17462#S2.SS1.p2.1)\.
- Miettinen \(1999\)K\. MiettinenNonlinear multiobjective optimization\.Vol\.12,Springer Science & Business Media\.Cited by:[§1](https://arxiv.org/html/2611.17462#S1.p4.1)\.
- Oganov and Glass \(2006\)A\. R\. Oganov and C\. W\. GlassCrystal structure prediction using ab initio evolutionary techniques: principles and applications\.The Journal of chemical physics124\(24\)\.Cited by:[§2\.1](https://arxiv.org/html/2611.17462#S2.SS1.p1.1)\.
- Onget al\.\(2013\)S\. P\. Ong, W\. D\. Richards, A\. Jain, G\. Hautier, M\. Kocher, S\. Cholia, D\. Gunter, V\. L\. Chevrier, K\. A\. Persson, and G\. CederPython materials genomics \(pymatgen\): a robust, open\-source python library for materials analysis\.Computational Materials Science68,pp\. 314–319\.Cited by:[§4\.1](https://arxiv.org/html/2611.17462#S4.SS1.p2.1),[§4\.2](https://arxiv.org/html/2611.17462#S4.SS2.p1.1)\.
- Onishi and Fu \(2024\)Y\. Onishi and L\. FuUniversal relation between energy gap and dielectric constant\.Physical Review B110\(15\),pp\. 155107\.Cited by:[§4\.3](https://arxiv.org/html/2611.17462#S4.SS3.p1.1)\.
- Penn \(1962\)D\. R\. PennWave\-number\-dependent dielectric function of semiconductors\.Physical review128\(5\),pp\. 2093\.Cited by:[§4\.3](https://arxiv.org/html/2611.17462#S4.SS3.p1.1)\.
- Pickard and Needs \(2011\)C\. J\. Pickard and R\. NeedsAb initio random structure searching\.Journal of Physics: Condensed Matter23\(5\),pp\. 053201\.Cited by:[§2\.1](https://arxiv.org/html/2611.17462#S2.SS1.p1.1)\.
- Riebesellet al\.\(2024\)J\. Riebesell, T\. Surta, R\. Goodall, M\. Gaultois, and A\. A\. LeePushing the pareto front of band gap and permittivity: ml\-guided search for dielectric materials\.arXiv preprint arXiv:2401\.05848\.Cited by:[§1](https://arxiv.org/html/2611.17462#S1.p4.1),[§4\.3](https://arxiv.org/html/2611.17462#S4.SS3.p1.1)\.
- Robertson \(2005\)J\. RobertsonHigh dielectric constant gate oxides for metal oxide si transistors\.Reports on progress in Physics69\(2\),pp\. 327\.Cited by:[§4\.3](https://arxiv.org/html/2611.17462#S4.SS3.p1.1)\.
- Singhalet al\.\(2025\)R\. Singhal, Z\. Horvitz, R\. Teehan, M\. Ren, Z\. Yu, K\. McKeown, and R\. RanganathA general framework for inference\-time scaling and steering of diffusion models\.arXiv preprint arXiv:2501\.06848\.Cited by:[§G\.1](https://arxiv.org/html/2611.17462#A7.SS1.p2.1),[§3\.2](https://arxiv.org/html/2611.17462#S3.SS2.SSS0.Px2.p1.1),[§4\.1](https://arxiv.org/html/2611.17462#S4.SS1.SSS0.Px2.p2.1),[§5](https://arxiv.org/html/2611.17462#S5.SS0.SSS0.Px1.p1.1)\.
- Tanget al\.\(2025\)S\. Tang, Y\. Zhang, and P\. ChatterjeePeptune: de novo generation of therapeutic peptides with multi\-objective\-guided discrete diffusion\.ArXiv,pp\. arXiv–2412\.Cited by:[§5](https://arxiv.org/html/2611.17462#S5.SS0.SSS0.Px1.p1.1)\.
- Tanget al\.\(2024\)Z\. Tang, J\. Peng, J\. Tang, M\. Hong, F\. Wang, and T\. ChangInference\-time alignment of diffusion models with direct noise optimization\.arXiv preprint arXiv:2405\.18881\.Cited by:[§1](https://arxiv.org/html/2611.17462#S1.p3.1)\.
- Ueharaet al\.\(2025\)M\. Uehara, Y\. Zhao, C\. Wang, X\. Li, A\. Regev, S\. Levine, and T\. BiancalaniInference\-time alignment in diffusion models with reward\-guided generation: tutorial and review\.arXiv preprint arXiv:2501\.09685\.Cited by:[§1](https://arxiv.org/html/2611.17462#S1.p3.1),[§2\.2](https://arxiv.org/html/2611.17462#S2.SS2.p1.2)\.
- Wales and Doye \(1997\)D\. J\. Wales and J\. P\. DoyeGlobal optimization by basin\-hopping and the lowest energy structures of lennard\-jones clusters containing up to 110 atoms\.The Journal of Physical Chemistry A101\(28\),pp\. 5111–5116\.Cited by:[Appendix B](https://arxiv.org/html/2611.17462#A2.p1.1),[§2\.1](https://arxiv.org/html/2611.17462#S2.SS1.p1.1)\.
- Wales and Scheraga \(1999\)D\. J\. Wales and H\. A\. ScheragaGlobal optimization of clusters, crystals, and biomolecules\.Science285\(5432\),pp\. 1368–1372\.Cited by:[Appendix B](https://arxiv.org/html/2611.17462#A2.p1.1),[§2\.1](https://arxiv.org/html/2611.17462#S2.SS1.p1.1)\.
- Yanget al\.\(2024\)H\. Yang, C\. Hu, Y\. Zhou, X\. Liu, Y\. Shi, J\. Li, G\. Li, Z\. Chen, S\. Chen, C\. Zeni,et al\.Mattersim: a deep learning atomistic model across elements, temperatures and pressures\.arXiv preprint arXiv:2405\.04967\.Cited by:[§4\.1](https://arxiv.org/html/2611.17462#S4.SS1.p2.1)\.
- Yaoet al\.\(2024\)Y\. Yao, Y\. Pan, J\. Li, I\. Tsang, and X\. YaoPROUD: pareto\-guided diffusion model for multi\-objective generation\.Machine Learning113\(9\),pp\. 6511–6538\.Cited by:[§5](https://arxiv.org/html/2611.17462#S5.SS0.SSS0.Px2.p1.1)\.
- Yehet al\.\(2024\)P\. Yeh, K\. Lee, and J\. ChenTraining\-free diffusion model alignment with sampling demons\.arXiv preprint arXiv:2410\.05760\.Cited by:[§1](https://arxiv.org/html/2611.17462#S1.p3.1)\.
- Yuanet al\.\(2024\)Y\. Yuan, C\. Chen, C\. Pal, and X\. LiuParetoflow: guided flows in multi\-objective optimization\.arXiv preprint arXiv:2412\.03718\.Cited by:[§5](https://arxiv.org/html/2611.17462#S5.SS0.SSS0.Px2.p1.1)\.
- Zaidiet al\.\(2022\)S\. Zaidi, M\. Schaarschmidt, J\. Martens, H\. Kim, Y\. W\. Teh, A\. Sanchez\-Gonzalez, P\. Battaglia, R\. Pascanu, and J\. GodwinPre\-training via denoising for molecular property prediction\.arXiv preprint arXiv:2206\.00133\.Cited by:[Appendix B](https://arxiv.org/html/2611.17462#A2.p2.1)\.
- Zeniet al\.\(2025\)C\. Zeni, R\. Pinsler, D\. Zügner, A\. Fowler, M\. Horton, X\. Fu, Z\. Wang, A\. Shysheya, J\. Crabbé, S\. Ueda,et al\.A generative model for inorganic materials design\.Nature639\(8055\),pp\. 624–632\.Cited by:[§1](https://arxiv.org/html/2611.17462#S1.p1.1),[§1](https://arxiv.org/html/2611.17462#S1.p2.1),[§2\.1](https://arxiv.org/html/2611.17462#S2.SS1.p2.1),[§4\.1](https://arxiv.org/html/2611.17462#S4.SS1.p2.1),[§5](https://arxiv.org/html/2611.17462#S5.SS0.SSS0.Px1.p1.1)\.
- Zhuet al\.\(2023\)Y\. Zhu, J\. Wu, C\. Hu, J\. Yan, T\. Hou, J\. Wu,et al\.Sample\-efficient multi\-objective molecular optimization with gflownets\.Advances in Neural Information Processing Systems36,pp\. 79667–79684\.Cited by:[§5](https://arxiv.org/html/2611.17462#S5.SS0.SSS0.Px2.p1.1)\.
## Appendix AAlgorithm
Algorithm 1Diffusion Tree Sampling with guided proposals and multiple objectives\. Unmarked lines are DTS\[[Jain et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib15)\]; lines tagged\[IS\]add the importance\-corrected guided proposals of[Section3\.1](https://arxiv.org/html/2611.17462#S3.SS1), and lines tagged\[MO\]add multi\-objective selection and backup of[Section3\.2](https://arxiv.org/html/2611.17462#S3.SS2)\. Batched execution is described in[Section3\.3](https://arxiv.org/html/2611.17462#S3.SS3)\.1:Input:base model
pθp\_\{\\theta\}; inverse temperature
λ\\lambda; NFE budget
NNFEN\_\{\\mathrm\{NFE\}\};
2:Input:widening parameters
CC,
α\\alpha, branching timesteps
𝒯\\mathcal\{T\}
3:Input:guided proposalqcq\_\{c\}\(default
qc=pθq\_\{c\}=p\_\{\\theta\}\)\[IS\]
4:Input:objectives𝐫=\(r1,…,rm\)\{\\mathbf\{r\}\}=\(r\_\{1\},\\dots,r\_\{m\}\); scalarizationss; preferencesp\(w\)p\(w\); condition mapw↦cww\\mapsto c\_\{w\}\(default
m=1m=1,
s=ids=\\mathrm\{id\},
cw≡cc\_\{w\}\\equiv c\)\[MO\]
5:Initialize dummy root
𝐱T\+1\{\\mathbf\{x\}\}\_\{T\+1\}with
p\(𝐱T∣𝐱T\+1\)=pTp\(\{\\mathbf\{x\}\}\_\{T\}\\mid\{\\mathbf\{x\}\}\_\{T\+1\}\)=p\_\{T\};
𝐯^\(𝐱T\+1\)←𝟎\\hat\{\{\\mathbf\{v\}\}\}\(\{\\mathbf\{x\}\}\_\{T\+1\}\)\\leftarrow\\mathbf\{0\},
N\(𝐱T\+1\)←1N\(\{\\mathbf\{x\}\}\_\{T\+1\}\)\\leftarrow 1,
𝒞\(𝐱T\+1\)←∅\\mathcal\{C\}\(\{\\mathbf\{x\}\}\_\{T\+1\}\)\\leftarrow\\emptyset
6:
nfe←0\\mathrm\{nfe\}\\leftarrow 0
7:while
nfe<NNFE\\mathrm\{nfe\}<N\_\{\\mathrm\{NFE\}\}do
8:// Initialize path
9:
P←\{𝐱T\+1\}P\\leftarrow\\\{\{\\mathbf\{x\}\}\_\{T\+1\}\\\};
t←T\+1t\\leftarrow T\+1
10:
w∼p\(w\)w\\sim p\(w\);c←cwc\\leftarrow c\_\{w\}\[MO\]
11:// Selection: descend until reaching a terminal or an expandable node
12:while
t\>0t\>0do
13:if
t∈𝒯t\\in\\mathcal\{T\}and
\|𝒞\(𝐱t\)\|<CN\(𝐱t\)α\|\\mathcal\{C\}\(\{\\mathbf\{x\}\}\_\{t\}\)\|<C\\,N\(\{\\mathbf\{x\}\}\_\{t\}\)^\{\\alpha\}then
14:break
15:endif
16:
𝐱t−1←𝐳∈𝒞\(𝐱t\)\{\\mathbf\{x\}\}\_\{t\-1\}\\leftarrow\{\\mathbf\{z\}\}\\in\\mathcal\{C\}\(\{\\mathbf\{x\}\}\_\{t\}\)with probability
η\(𝐳\)exp\(λs\(𝐯^\(𝐳\)∣w\)\)∑𝐳′∈𝒞\(𝐱t\)η\(𝐳′\)exp\(λs\(𝐯^\(𝐳′\)∣w\)\)\\dfrac\{\{\\color\[rgb\]\{0,0\.3516,0\.7109\}\\eta\(\{\\mathbf\{z\}\}\)\}\\,\\exp\\\!\\big\(\\lambda\\,\{\\color\[rgb\]\{0\.6992,0\.0938,0\.168\}s\(\\hat\{\{\\mathbf\{v\}\}\}\(\{\\mathbf\{z\}\}\)\\mid w\)\}\\big\)\}\{\\sum\_\{\{\\mathbf\{z\}\}^\{\\prime\}\\in\\mathcal\{C\}\(\{\\mathbf\{x\}\}\_\{t\}\)\}\{\\color\[rgb\]\{0,0\.3516,0\.7109\}\\eta\(\{\\mathbf\{z\}\}^\{\\prime\}\)\}\\,\\exp\\\!\\big\(\\lambda\\,\{\\color\[rgb\]\{0\.6992,0\.0938,0\.168\}s\(\\hat\{\{\\mathbf\{v\}\}\}\(\{\\mathbf\{z\}\}^\{\\prime\}\)\\mid w\)\}\\big\)\}\[IS\]\[MO\]
17:
P←P∪\{𝐱t−1\}P\\leftarrow P\\cup\\\{\{\\mathbf\{x\}\}\_\{t\-1\}\\\};
t←t−1t\\leftarrow t\-1
18:endwhile
19:// Expansion and rollout: grow a new path tot=0t=0, adding every node to the tree
20:while
t\>0t\>0do
21:
𝐱t−1∼qc\(⋅∣𝐱t\)\{\\mathbf\{x\}\}\_\{t\-1\}\\sim\{\\color\[rgb\]\{0,0\.3516,0\.7109\}q\_\{c\}\}\(\\,\\cdot\\mid\{\\mathbf\{x\}\}\_\{t\}\);
nfe←nfe\+1\\mathrm\{nfe\}\\leftarrow\\mathrm\{nfe\}\+1\[IS\]
22:
𝐯^\(𝐱t−1\)←𝟎∈ℝm\\hat\{\{\\mathbf\{v\}\}\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\)\\leftarrow\\mathbf\{0\}\\in\{\\color\[rgb\]\{0\.6992,0\.0938,0\.168\}\{\\mathbb\{R\}\}^\{m\}\};
N\(𝐱t−1\)←1N\(\{\\mathbf\{x\}\}\_\{t\-1\}\)\\leftarrow 1;
𝒞\(𝐱t\)←𝒞\(𝐱t\)∪\{𝐱t−1\}\\mathcal\{C\}\(\{\\mathbf\{x\}\}\_\{t\}\)\\leftarrow\\mathcal\{C\}\(\{\\mathbf\{x\}\}\_\{t\}\)\\cup\\\{\{\\mathbf\{x\}\}\_\{t\-1\}\\\}\[MO\]
23:
η\(𝐱t−1\)←pθ\(𝐱t−1∣𝐱t\)/qc\(𝐱t−1∣𝐱t\)\\eta\(\{\\mathbf\{x\}\}\_\{t\-1\}\)\\leftarrow p\_\{\\theta\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\\mid\{\\mathbf\{x\}\}\_\{t\}\)\\,/\\,q\_\{c\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\\mid\{\\mathbf\{x\}\}\_\{t\}\)\[IS\]
24:
P←P∪\{𝐱t−1\}P\\leftarrow P\\cup\\\{\{\\mathbf\{x\}\}\_\{t\-1\}\\\};
t←t−1t\\leftarrow t\-1
25:endwhile
26:// Terminal evaluation
27:
𝐯^\(𝐱0\)←𝐫\(𝐱0\)\\hat\{\{\\mathbf\{v\}\}\}\(\{\\mathbf\{x\}\}\_\{0\}\)\\leftarrow\{\\color\[rgb\]\{0\.6992,0\.0938,0\.168\}\{\\mathbf\{r\}\}\(\{\\mathbf\{x\}\}\_\{0\}\)\}\[MO\]
28:// Backup: per\-objective self\-normalized soft Bellman update
29:foreach
𝐱t∈P\{\\mathbf\{x\}\}\_\{t\}\\in Pwith
t≥1t\\geq 1, from
𝐱1\{\\mathbf\{x\}\}\_\{1\}up to
𝐱T\+1\{\\mathbf\{x\}\}\_\{T\+1\}do
30:fori=1,…,mi=1,\\dots,m:
v^i\(𝐱t\)←1λlog∑𝐳∈𝒞\(𝐱t\)η\(𝐳\)eλv^i\(𝐳\)∑𝐳∈𝒞\(𝐱t\)η\(𝐳\)\\hat\{v\}\_\{i\}\(\{\\mathbf\{x\}\}\_\{t\}\)\\leftarrow\\dfrac\{1\}\{\\lambda\}\\log\\dfrac\{\\sum\_\{\{\\mathbf\{z\}\}\\in\\mathcal\{C\}\(\{\\mathbf\{x\}\}\_\{t\}\)\}\{\\color\[rgb\]\{0,0\.3516,0\.7109\}\\eta\(\{\\mathbf\{z\}\}\)\}\\,e^\{\\lambda\\hat\{v\}\_\{i\}\(\{\\mathbf\{z\}\}\)\}\}\{\\sum\_\{\{\\mathbf\{z\}\}\\in\\mathcal\{C\}\(\{\\mathbf\{x\}\}\_\{t\}\)\}\{\\color\[rgb\]\{0,0\.3516,0\.7109\}\\eta\(\{\\mathbf\{z\}\}\)\}\}\[IS\]\[MO\]
31:
N\(𝐱t\)←N\(𝐱t\)\+1N\(\{\\mathbf\{x\}\}\_\{t\}\)\\leftarrow N\(\{\\mathbf\{x\}\}\_\{t\}\)\+1
32:endfor
33:endwhile
34:returnterminal nodes
\{𝐱0\}\\\{\{\\mathbf\{x\}\}\_\{0\}\\\}of the tree
## Appendix BConnection to Basin Hopping and Monte Carlo Minimization
DTS bears a structural resemblance to a family of classical global optimization methods used in computational chemistry, most notably Monte Carlo Minimization\[[Li and Scheraga, 1987](https://arxiv.org/html/2611.17462#bib.bib36)\]and its successor, basin hopping\[[Wales and Doye, 1997](https://arxiv.org/html/2611.17462#bib.bib19),[Wales and Scheraga, 1999](https://arxiv.org/html/2611.17462#bib.bib20)\]\. These algorithms iterate a perturb–relax–accept cycle: a candidate configuration is randomly displaced, mapped to the bottom of the resulting basin via a deterministic local minimizer, and accepted or rejected by a Metropolis criterion at temperatureTTon the minimized energies\.
Different components of DTS have natural counterparts in this framework\. Tree expansion of a nodextx\_\{t\}plays the role of perturbation, with the noise levelttcontrolling an effective step size; the unrolled reverse process fromttto00plays the role of local minimization, an interpretation supported by the established equivalence between denoising score matching and learning a Boltzmann force field over equilibrium structures\[[Zaidi et al\., 2022](https://arxiv.org/html/2611.17462#bib.bib37),[Arts et al\., 2023](https://arxiv.org/html/2611.17462#bib.bib38)\]; the soft Bellman backup with inverse temperatureλ\\lambdaplays the role of Metropolis acceptance, with the greedy variant of DTS recovering monotonic basin hopping\.
Several features distinguish our method from these classical methods\. The search tree retains and shares value estimates across all visited nodes rather than maintaining a single chain\. Rather than just minimizing energy via structural relaxation, using a generative model enables swapping of atom types in a structure\. Lastly, the equivalence between score matching and force fields does not apply to property\-conditional models, which steer generation towards specific properties rather than just energy minima\.
## Appendix CIllustrative experiments
### C\.1Gaussian\-mixture toy illustration of the double\-tilt problem
[Figure5](https://arxiv.org/html/2611.17462#A3.F5)illustrates the double\-tilt problem and the importance correction of[Section3\.1](https://arxiv.org/html/2611.17462#S3.SS1)on a Gaussian mixture\. The unconditional distribution is an equal\-weight four\-component mixture; with linear rewardr\(x\)=xr\(x\)=xthe analytic targetπ\(x\)∝p\(𝐱\)exp\(λf\(x\)\)\\pi\(x\)\\propto p\(\{\\mathbf\{x\}\}\)\\exp\(\\lambda f\(x\)\)concentrates on the higher\-xxmodes\. A conditional proposalp\(𝐱∣c\)p\(\{\\mathbf\{x\}\}\\mid c\)biased toward positivexxaligns better withπ\\pi, but naively running DTS with this proposal double\-tilts and distorts the distribution\. Multiplying the reward weights by the importance ratio recoversπ\\pito within a few percent at every mode\.
Figure 5:\(a\) The unconditional distributionp\(𝐱\)p\(\\mathbf\{x\}\)is an equal\-weight four\-component Gaussian mixture; \(b\) with linear rewardf\(x\)=xf\(x\)=x, the analytic targetπ\(x\)∝p\(𝐱\)exp\(λf\(x\)\)\\pi\(x\)\\propto p\(\\mathbf\{x\}\)\\,\\exp\(\\lambda f\(x\)\)concentrates mass on higher\-𝐱\\mathbf\{x\}modes; \(c\) the conditional proposalp\(𝐱\|c\)p\(\\mathbf\{x\}\|c\)is biased toward positive𝐱\\mathbf\{x\}and aligns better with the target \(dashed black\); \(d\) naively running DTS withp\(𝐱\|c\)p\(\\mathbf\{x\}\|c\)as the proposal double\-tilts the distribution: the two left modes collapse by roughly85%85\\%and the rightmost mode overshoots by21%21\\%; \(e\) multiplying the reward weights by the importance ratiop\(𝐱\)/p\(𝐱\|c\)p\(\\mathbf\{x\}\)/p\(\\mathbf\{x\}\|c\)recoversπ\\pito within a few percent at every mode\.
### C\.2Illustration on a Gaussian mixture in multi\-objective setting
[Figure6](https://arxiv.org/html/2611.17462#A3.F6)illustrated how our method handles multiple conflicting objectives by scalarizing different preferences over the objectives in a single shared tree\. The unconditional distribution is a two\-dimensional Gaussianp\(x,y\)=𝒩\(\(0,0\),I\)p\(x,y\)=\\mathcal\{N\}\(\(0,0\);I\)\. The two objectivesf1\(x,y\)=xf\_\{1\}\(x,y\)=xandf2\(x,y\)=yf\_\{2\}\(x,y\)=ywith corresponding weightsw1w\_\{1\}andw2w\_\{2\}tilt the distribution to give the targetπ\(x,y\)∝p\(x,y\)exp\(w1x\+w2y\)\\pi\(x,y\)\\propto p\(x,y\)\\,\\exp\(w\_1 x \+ w\_2 y\)\. Different preferences𝐰=\(w1,w2\)\\mathbf\{w\}=\(w\_\{1\},w\_\{2\}\)define different target distributions, and our aim is to produce samples from the Pareto front that covers all possible preferences\. For illustrative purposes, we fix three specific preferences\.
The unconditional model, even with DTS, fails to produce samples from the target distribution since the proposal model does not have high mass in the target region\. We use a conditional model that is conditioned on\(w1x\+w2y\)\(w\_\{1\}x\+w\_\{2\}y\)to produce samples closer to the target region and use it as a guided proposal within DTS\. However, naively replacing the proposal results in*double\-tilting*, so the samples lie outside the target distribution\. We fix this by using an importance sampling correction \(as described in[Section3\.1](https://arxiv.org/html/2611.17462#S3.SS1)\) that accounts for the conditional model\.
Figure 6:\(a\) The target distribution is a mixture of different preferences over the tilted distributionπ\(x,y\)∝p\(x,y\)exp\(w1x\+w2y\)\\pi\(x,y\)\\propto p\(x,y\)\\,\\exp\(w\_1 x \+ w\_2 y\); \(b\) the unconditional distributionp\(x,y\)p\(x,y\)is a two\-dimensional standard Gaussian; \(c\) using DTS with the unconditional proposal fails to sample accurately from the target distribution due to low density of the base distribution in this region; \(d\) naively running DTS with a conditional modelp\(x,y\|c\)p\(x,y\|c\)as the proposal double\-tilts the distribution; \(e\) multiplying the reward weights by the importance ratiop\(x,y\)/p\(x,y\|c\)p\(x,y\)/p\(x,y\|c\)produces samples that align better with the target distribution\.
## Appendix DPredictor–corrector and other multi\-stage samplers
The discussion in[Section3\.1](https://arxiv.org/html/2611.17462#S3.SS1)assumes a single\-step reverse transition\. In practice, modern diffusion samplers often implement a*multi\-stage*update betweenttandt−1t\-1, for improved stability and sample quality\. A common pattern is*predictor–corrector*\(PC\) sampling, where a predictor step performs a discretized reverse\-SDE/ODE update, and a corrector step applies one or more refinement moves \(e\.g\., Langevin/MCMC steps\) that approximately preserve the desired marginal at the new noise level\. More broadly, modern samplers may use higher\-order solvers, adaptive step sizes, or alternating deterministic/stochastic updates; our method remains valid as long as we can write down \(or evaluate up to a constant\) the probability of the composite transition used by the base sampler\.
To incorporate PC sampling into guided proposals and importance correction, we treat the overallt→t−1t\\to t\-1move as a composition through an intermediate state𝐱~t−1\\tilde\{\{\\mathbf\{x\}\}\}\_\{t\-1\}:
pθ\(𝐱t−1∣𝐱t\)=∫ppred\(𝐱~t−1∣𝐱t\)pcorr\(𝐱t−1∣𝐱~t−1\)d𝐱~t−1,p\_\{\\theta\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\\mid\{\\mathbf\{x\}\}\_\{t\}\)=\\int p\_\{\\text\{pred\}\}\(\\tilde\{\{\\mathbf\{x\}\}\}\_\{t\-1\}\\mid\{\\mathbf\{x\}\}\_\{t\}\)\\,p\_\{\\text\{corr\}\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\\mid\\tilde\{\{\\mathbf\{x\}\}\}\_\{t\-1\}\)\\,d\\tilde\{\{\\mathbf\{x\}\}\}\_\{t\-1\},with an analogous guided/conditional composite transition whereqpred,w\(⋅∣𝐱t\):=ppred\(⋅∣𝐱t,c\(w\)\)q\_\{\\text\{pred\},w\}\(\\cdot\\mid\{\\mathbf\{x\}\}\_\{t\}\):=p\_\{\\text\{pred\}\}\(\\cdot\\mid\{\\mathbf\{x\}\}\_\{t\},c\(w\)\)and similarly forqcorr,wq\_\{\\text\{corr\},w\}\. In the tree expansion/rollout we explicitly sample𝐱~t−1\\tilde\{\{\\mathbf\{x\}\}\}\_\{t\-1\}and then𝐱t−1\{\\mathbf\{x\}\}\_\{t\-1\}, and use a factorized importance ratio consistent with this sampling path:
pθ\(𝐱t−1∣𝐱t\)qw\(𝐱t−1∣𝐱t\)≈ppred\(𝐱~t−1∣𝐱t\)qpred,w\(𝐱~t−1∣𝐱t\)×pcorr\(𝐱t−1∣𝐱~t−1\)qcorr,w\(𝐱t−1∣𝐱~t−1\)\.\\frac\{p\_\{\\theta\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\\mid\{\\mathbf\{x\}\}\_\{t\}\)\}\{q\_\{w\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\\mid\{\\mathbf\{x\}\}\_\{t\}\)\}\\;\\approx\\;\\frac\{p\_\{\\text\{pred\}\}\(\\tilde\{\{\\mathbf\{x\}\}\}\_\{t\-1\}\\mid\{\\mathbf\{x\}\}\_\{t\}\)\}\{q\_\{\\text\{pred\},w\}\(\\tilde\{\{\\mathbf\{x\}\}\}\_\{t\-1\}\\mid\{\\mathbf\{x\}\}\_\{t\}\)\}\\;\\times\\;\\frac\{p\_\{\\text\{corr\}\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\\mid\\tilde\{\{\\mathbf\{x\}\}\}\_\{t\-1\}\)\}\{q\_\{\\text\{corr\},w\}\(\{\\mathbf\{x\}\}\_\{t\-1\}\\mid\\tilde\{\{\\mathbf\{x\}\}\}\_\{t\-1\}\)\}\.\(6\)This replaces the single\-step ratio in[Equation3](https://arxiv.org/html/2611.17462#S3.E3)\. Although the right\-hand side is not the marginal density ratio for𝐱t−1\{\\mathbf\{x\}\}\_\{t\-1\}alone, it is the exact path\-space importance weight for the joint augmented transition\(𝐱~t−1,𝐱t−1\)\(\\tilde\{\\mathbf\{x\}\}\_\{t\-1\},\{\\mathbf\{x\}\}\_\{t\-1\}\)actually sampled by the algorithm; the≈\\approxsymbol reflects that the marginalization over𝐱~t−1\\tilde\{\\mathbf\{x\}\}\_\{t\-1\}is performed implicitly by sampling\. In particular, if the corrector is shared between conditional and unconditional samplers \(i\.e\.,qcorr,w=pcorrq\_\{\\text\{corr\},w\}=p\_\{\\text\{corr\}\}\), the corrector ratio cancels and only the predictor ratio is needed\. When both predictor and corrector differ \(e\.g\., guided score used in both stages\),[Equation6](https://arxiv.org/html/2611.17462#A4.E6)yields the appropriate weight for maintaining consistency with the unconditional\-base target while still leveraging guided proposals during the full sampling procedure\.
## Appendix EScalarizations
As explained in[Section3\.2](https://arxiv.org/html/2611.17462#S3.SS2), a standard way to operationalize multi\-objective optimization is*scalarization*: decompose it into a family of single\-objective problems of the formmax𝐱r\(𝐱∣w\)\\max\_\{\{\\mathbf\{x\}\}\}\\,r\(\{\\mathbf\{x\}\}\\mid w\), wherew∈Δm−1≔\{w∈ℝ≥0m:∑iwi=1\}w\\in\\Delta^\{m\-1\}\\coloneqq\\\{w\\in\\mathbb\{R\}^\{m\}\_\{\\geq 0\}:\\sum\_\{i\}w\_\{i\}=1\\\}encodes preferences andr\(𝐱∣w\)≔s\(𝐫\(𝐱\)∣w\)r\(\{\\mathbf\{x\}\}\\mid w\)\\coloneqq s\(\{\\mathbf\{r\}\}\(\{\\mathbf\{x\}\}\)\\mid w\)for scalarizationss\. The simplest choice forssis the weighted sum,s\(𝐚∣w\)=∑i=1mwiais\(\{\\mathbf\{a\}\}\\mid w\)=\\sum\_\{i=1\}^\{m\}w\_\{i\}\\,a\_\{i\}\. An important alternative is \(weighted\)*Tchebycheff*scalarization\[[Bowman Jr, 1976](https://arxiv.org/html/2611.17462#bib.bib42)\], which measures the worst \(weighted\) deviation from an*ideal*vector𝐳∈ℝm\{\\mathbf\{z\}\}\\in\\mathbb\{R\}^\{m\}:
s\(𝐚∣w,𝐳\)=−maxi∈\{1,…,m\}wi\|ai−zi\|,s\(\{\\mathbf\{a\}\}\\mid w,\{\\mathbf\{z\}\}\)\\;=\\;\-\\max\_\{i\\in\\\{1,\\dots,m\\\}\}w\_\{i\}\\,\\absolutevalue\{a\_i\-z\_i\},\(7\)i\.e\., maximizings\(⋅∣w,𝒛\)s\(\\cdot\\mid w,\{\\bm\{z\}\}\)is equivalent to minimizing the worst weighted distance to the ideal\. This encourages balanced progress across objectives and can recover Pareto\-optimal points that may be missed by linear scalarization on non\-convex fronts\. Both linear and Tchebycheff scalarizations can be unified via a generalizedpp\-norm form,
s\(𝐚∣w,𝒛\)=−‖w⊙\|𝐚−𝐳\|‖p,s\\\!\\left\(\{\\mathbf\{a\}\}\\mid w,\{\\bm\{z\}\}\\right\)\\;=\\;\-\\norm\{ w \\odot\\abs\{\\rva\-\\rvz\} \}\_\{p\},\(8\)where⊙\\odotis elementwise multiplication\. Linear and Tchebycheff scalarizations are recovered forp=1p=1andp=∞p=\\inftyrespectively\. More generally, other weight\-parameterized scalarizations \(e\.g\., augmented Tchebycheff, PBI, achievement scalarizing functions\) can be viewed as different ways to sweep the Pareto front\. Our method is*indifferent*to the particular scalarization: any mappingw↦s\(⋅∣w\)w\\mapsto s\(\\cdot\\mid w\)is compatible as long as we can systematically choosewwfrom a specified set/distribution and evaluate the resulting scalar objective\.
## Appendix FProperties of the per\-objective value surrogate
Scalarizing only at selection time enables computation sharing across all preferences\. The exact soft value of the scalarized reward,vt⋆\(𝐱t∣w\):=1λlog𝔼\[exp\(λr\(𝐱0∣w\)\)∣𝐱t\]v\_\{t\}^\{\\star\}\(\{\\mathbf\{x\}\}\_\{t\}\\mid w\):=\\tfrac\{1\}\{\\lambda\}\\log\\mathbb\{E\}\[\\exp\(\\lambda\\, r\(\\rvx\_0\\mid w\)\)\\mid\{\\mathbf\{x\}\}\_\{t\}\], would instead require a separate backup tree for everyww\. The per\-objective surrogate ofvt\(𝐱t\|w\)v\_\{t\}\(\{\\mathbf\{x\}\}\_\{t\}\|w\)relates to it as follows\.
###### Proposition F\.1\.
Letvt⋆\(𝐱t∣w\)v\_\{t\}^\{\\star\}\(\{\\mathbf\{x\}\}\_\{t\}\\mid w\)be the exact scalarized soft value, and let
vt,i\(𝐱t\):=1λlog𝔼pθ\(𝐱0∣𝐱t\)\[exp\(λri\(𝐱0\)\)\],v\_\{t,i\}\(\{\\mathbf\{x\}\}\_\{t\}\)\\;:=\\;\\tfrac\{1\}\{\\lambda\}\\log\\mathbb\{E\}\_\{p\_\{\\theta\}\(\{\\mathbf\{x\}\}\_\{0\}\\mid\{\\mathbf\{x\}\}\_\{t\}\)\}\\\!\\big\[\\exp\(\\lambda\\, r\_i\(\\rvx\_0\)\)\\big\],\(9\)be the per\-objective values used by MO\-DTS, withvt\(𝐱t\|w\)≔s\(𝐯t\(𝐱t\)\|w\)v\_\{t\}\(\{\\mathbf\{x\}\}\_\{t\}\|w\)\\coloneqq s\(\{\\mathbf\{v\}\}\_\{t\}\(\{\\mathbf\{x\}\}\_\{t\}\)\|w\)its scalarization\. Assume the rewardsrir\_\{i\}are bounded\. Then:
1. 1\.For*linear*scalarization,vt⋆\(𝐱t∣w\)≤vt\(𝐱t∣w\)v\_\{t\}^\{\\star\}\(\{\\mathbf\{x\}\}\_\{t\}\\mid w\)\\leq v\_\{t\}\(\{\\mathbf\{x\}\}\_\{t\}\\mid w\)for everyλ\>0\\lambda\>0\.
2. 2\.For any continuous scalarization that is non\-decreasing in each objective,vt⋆v\_\{t\}^\{\\star\}is upper bounded byvtv\_\{t\}in the low\-temperature limit: limλ→∞vt⋆\(𝐱t∣w\)≤limλ→∞vt\(𝐱t∣w\)\.\\lim\_\{\\lambda\\to\\infty\}v\_\{t\}^\{\\star\}\(\{\\mathbf\{x\}\}\_\{t\}\\mid w\)\\;\\leq\\;\\lim\_\{\\lambda\\to\\infty\}v\_\{t\}\(\{\\mathbf\{x\}\}\_\{t\}\\mid w\)\.\(10\)
3. 3\.For any continuous scalarization, the approximation becomes exact att=0t=0: v0\(𝐱0∣w\)=v0⋆\(𝐱0∣w\),v\_\{0\}\(\{\\mathbf\{x\}\}\_\{0\}\\mid w\)=v\_\{0\}^\{\\star\}\(\{\\mathbf\{x\}\}\_\{0\}\\mid w\),\(11\)and, whenever the conditional uncertainty in𝐫\(𝐱0\)\{\\mathbf\{r\}\}\(\{\\mathbf\{x\}\}\_\{0\}\)given𝐱t\{\\mathbf\{x\}\}\_\{t\}vanishes ast→0t\\to 0, for every fixedλ\>0\\lambda\>0, \|vt\(𝐱t∣w\)−vt⋆\(𝐱t∣w\)\|→0\.\\absolutevalue\{v\_t\(\\rvx\_t\\mid w\)\-v\_t^\\star\(\\rvx\_t\\mid w\)\}\\to 0\.\(12\)
###### Proof\.
Fix𝐱t\{\\mathbf\{x\}\}\_\{t\}and let all expectations and essential suprema be underpθ\(𝐱0∣𝐱t\)p\_\{\\theta\}\(\{\\mathbf\{x\}\}\_\{0\}\\mid\{\\mathbf\{x\}\}\_\{t\}\)\. WriteRi:=ri\(𝐱0\)R\_\{i\}:=r\_\{i\}\(\{\\mathbf\{x\}\}\_\{0\}\)and𝐫:=\(R1,…,Rm\)\{\\mathbf\{r\}\}:=\(R\_\{1\},\\dots,R\_\{m\}\)\.
##### \(i\) Linear scalarization, allλ\\lambda\.
Assumes\(𝐚∣w\)=∑iwiais\(\{\\mathbf\{a\}\}\\mid w\)=\\sum\_\{i\}w\_\{i\}a\_\{i\}withwi≥0w\_\{i\}\\geq 0and∑iwi=1\\sum\_\{i\}w\_\{i\}=1, and setXi:=exp\(λRi\)X\_\{i\}:=\\exp\(\\lambda R\_i\)\. Then
𝔼\[exp\(λs\(𝐫∣w\)\)\]=𝔼\[∏iXiwi\]\.\\mathbb\{E\}\[\\exp\(\\lambda s\(\\rvr\\mid w\)\)\]\\;=\\;\\mathbb\{E\}\\\!\\Big\[\\textstyle\\prod\_\{i\}X\_\{i\}^\{w\_\{i\}\}\\Big\]\.\(13\)Hölder’s inequality, with conjugate exponents1/wi1/w\_\{i\}for the positive weights, gives
𝔼\[∏iXiwi\]≤∏i𝔼\[Xi\]wi\.\\mathbb\{E\}\\\!\\Big\[\\textstyle\\prod\_\{i\}X\_\{i\}^\{w\_\{i\}\}\\Big\]\\;\\leq\\;\\textstyle\\prod\_\{i\}\\mathbb\{E\}\[X\_\{i\}\]^\{w\_\{i\}\}\.\(14\)Taking1λlog\\tfrac\{1\}\{\\lambda\}\\logon both sides yields
vt⋆\(𝐱t∣w\)≤∑iwi1λlog𝔼\[Xi\]=∑iwivt,i\(𝐱t\)=vt\(𝐱t∣w\)\.v\_\{t\}^\{\\star\}\(\{\\mathbf\{x\}\}\_\{t\}\\mid w\)\\;\\leq\\;\\sum\_\{i\}w\_\{i\}\\,\\tfrac\{1\}\{\\lambda\}\\log\\mathbb\{E\}\[X\_\{i\}\]\\;=\\;\\sum\_\{i\}w\_\{i\}\\,v\_\{t,i\}\(\{\\mathbf\{x\}\}\_\{t\}\)\\;=\\;v\_\{t\}\(\{\\mathbf\{x\}\}\_\{t\}\\mid w\)\.\(15\)Equality in[Equation14](https://arxiv.org/html/2611.17462#A6.E14)holds iff the active tilted variables are almost surely proportional, equivalentlyRi−RjR\_\{i\}\-R\_\{j\}is almost surely constant for all pairs withwi,wj\>0w\_\{i\},w\_\{j\}\>0\.
A second\-order Taylor expansion of the cumulant generating function gives, for bounded rewards,
vt,i\(𝐱t\)\\displaystyle v\_\{t,i\}\(\{\\mathbf\{x\}\}\_\{t\}\)=𝔼\[ri∣𝐱t\]\+λ2Var\(ri∣𝐱t\)\+O\(λ2\),\\displaystyle=\\mathbb\{E\}\[r\_\{i\}\\mid\{\\mathbf\{x\}\}\_\{t\}\]\+\\tfrac\{\\lambda\}\{2\}\\,\\mathrm\{Var\}\(r\_\{i\}\\mid\{\\mathbf\{x\}\}\_\{t\}\)\+O\(\\lambda^\{2\}\),\(16\)vt⋆\(𝐱t∣w\)\\displaystyle v\_\{t\}^\{\\star\}\(\{\\mathbf\{x\}\}\_\{t\}\\mid w\)=𝔼\[r\(𝐱0∣w\)∣𝐱t\]\+λ2Var\(r\(𝐱0∣w\)∣𝐱t\)\+O\(λ2\)\.\\displaystyle=\\mathbb\{E\}\[r\(\{\\mathbf\{x\}\}\_\{0\}\\mid w\)\\mid\{\\mathbf\{x\}\}\_\{t\}\]\+\\tfrac\{\\lambda\}\{2\}\\,\\mathrm\{Var\}\(r\(\{\\mathbf\{x\}\}\_\{0\}\\mid w\)\\mid\{\\mathbf\{x\}\}\_\{t\}\)\+O\(\\lambda^\{2\}\)\.\(17\)Under linear scalarization,𝔼\[r\(𝐱0∣w\)∣𝐱t\]=∑iwi𝔼\[ri∣𝐱t\]\\mathbb\{E\}\[r\(\{\\mathbf\{x\}\}\_\{0\}\\mid w\)\\mid\{\\mathbf\{x\}\}\_\{t\}\]=\\sum\_\{i\}w\_\{i\}\\mathbb\{E\}\[r\_\{i\}\\mid\{\\mathbf\{x\}\}\_\{t\}\], so the zeroth\-order terms agree and
vt\(𝐱t∣w\)−vt⋆\(𝐱t∣w\)=λ2\(∑iwiVar\(ri∣𝐱t\)−Var\(∑iwiri∣𝐱t\)\)\+O\(λ2\),v\_\{t\}\(\{\\mathbf\{x\}\}\_\{t\}\\mid w\)\-v\_\{t\}^\{\\star\}\(\{\\mathbf\{x\}\}\_\{t\}\\mid w\)\\;=\\;\\tfrac\{\\lambda\}\{2\}\\Big\(\\sum\_\{i\}w\_\{i\}\\,\\mathrm\{Var\}\(r\_\{i\}\\mid\{\\mathbf\{x\}\}\_\{t\}\)\-\\mathrm\{Var\}\(\\textstyle\\sum\_\{i\}w\_\{i\}r\_\{i\}\\mid\{\\mathbf\{x\}\}\_\{t\}\)\\Big\)\+O\(\\lambda^\{2\}\),\(18\)whose first\-order coefficient is non\-negative by convexity of variance\. The exact non\-negativity for allλ\\lambdais the Hölder bound above\.
##### \(ii\) Monotone scalarizations, low temperature\.
We use the standard log\-sum\-exp limit: if a random variableYYis bounded above, then
limλ→∞1λlog𝔼\[exp\(λY\)\]=esssupY\.\\lim\_\{\\lambda\\to\\infty\}\\frac\{1\}\{\\lambda\}\\log\\mathbb\{E\}\[\\exp\(\\lambda Y\)\]\\;=\\;\\operatorname\*\{ess\\,sup\}Y\.\(19\)Indeed, the upper bound follows from𝔼eλY≤eλesssupY\\mathbb\{E\}e^\{\\lambda Y\}\\leq e^\{\\lambda\\operatorname\*\{ess\\,sup\}Y\}; for anyϵ\>0\\epsilon\>0, the eventY≥esssupY−ϵY\\geq\\operatorname\*\{ess\\,sup\}Y\-\\epsilonhas positive probability, giving the matching lower bound after dividing byλ\\lambdaand takingλ→∞\\lambda\\to\\infty\.
Applying[Equation19](https://arxiv.org/html/2611.17462#A6.E19)toY=s\(𝐫∣w\)Y=s\(\{\\mathbf\{r\}\}\\mid w\)gives
limλ→∞vt⋆\(𝐱t∣w\)=esssups\(𝐫∣w\)\.\\lim\_\{\\lambda\\to\\infty\}v\_\{t\}^\{\\star\}\(\{\\mathbf\{x\}\}\_\{t\}\\mid w\)\\;=\\;\\operatorname\*\{ess\\,sup\}s\(\{\\mathbf\{r\}\}\\mid w\)\.\(20\)Applying it componentwise givesvt,i\(𝐱t\)→esssupRiv\_\{t,i\}\(\{\\mathbf\{x\}\}\_\{t\}\)\\to\\operatorname\*\{ess\\,sup\}R\_\{i\}\. By continuity ofs\(⋅∣w\)s\(\\cdot\\mid w\),
limλ→∞vt\(𝐱t∣w\)=s\(\(esssupR1,…,esssupRm\)∣w\)\.\\lim\_\{\\lambda\\to\\infty\}v\_\{t\}\(\{\\mathbf\{x\}\}\_\{t\}\\mid w\)\\;=\\;s\\\!\\big\(\(\\operatorname\*\{ess\\,sup\}R\_\{1\},\\dots,\\operatorname\*\{ess\\,sup\}R\_\{m\}\)\\mid w\\big\)\.\(21\)SinceRi≤esssupRiR\_\{i\}\\leq\\operatorname\*\{ess\\,sup\}R\_\{i\}almost surely for everyii, coordinatewise monotonicity impliess\(𝐫∣w\)≤s\(\(esssupR1,…,esssupRm\)∣w\)s\(\{\\mathbf\{r\}\}\\mid w\)\\leq s\\\!\\big\(\(\\operatorname\*\{ess\\,sup\}R\_\{1\},\\dots,\\operatorname\*\{ess\\,sup\}R\_\{m\}\)\\mid w\\big\)almost surely\. Therefore
esssups\(𝐫∣w\)≤s\(\(esssupR1,…,esssupRm\)∣w\),\\operatorname\*\{ess\\,sup\}s\(\{\\mathbf\{r\}\}\\mid w\)\\;\\leq\\;s\\\!\\big\(\(\\operatorname\*\{ess\\,sup\}R\_\{1\},\\dots,\\operatorname\*\{ess\\,sup\}R\_\{m\}\)\\mid w\\big\),\(22\)which proves the claimed low\-temperature optimistic bound\.
##### \(iii\) Low\-noise limit\.
Att=0t=0, the conditional lawpθ\(𝐱0′∣𝐱0\)p\_\{\\theta\}\(\{\\mathbf\{x\}\}\_\{0\}^\{\\prime\}\\mid\{\\mathbf\{x\}\}\_\{0\}\)is a point mass at𝐱0\{\\mathbf\{x\}\}\_\{0\}, so
v0,i\(𝐱0\)=ri\(𝐱0\),v0\(𝐱0∣w\)=v0⋆\(𝐱0∣w\)=s\(𝐫\(𝐱0\)∣w\)\.v\_\{0,i\}\(\{\\mathbf\{x\}\}\_\{0\}\)=r\_\{i\}\(\{\\mathbf\{x\}\}\_\{0\}\),\\qquad v\_\{0\}\(\{\\mathbf\{x\}\}\_\{0\}\\mid w\)=v\_\{0\}^\{\\star\}\(\{\\mathbf\{x\}\}\_\{0\}\\mid w\)=s\(\{\\mathbf\{r\}\}\(\{\\mathbf\{x\}\}\_\{0\}\)\\mid w\)\.\(23\)For the limiting statement, consider any sequence of low\-noise conditional laws for which𝐫→𝐫¯\{\\mathbf\{r\}\}\\to\\bar\{\\mathbf\{r\}\}in probability and the rewards remain bounded\. By dominated convergence,
𝔼\[exp\(λRi\)\]→exp\(λr¯i\),𝔼\[exp\(λs\(𝐫∣w\)\)\]→exp\(λs\(𝐫¯∣w\)\),\\mathbb\{E\}\[\\exp\(\\lambda R\_i\)\]\\to\\exp\(\\lambda\\bar r\_i\),\\qquad\\mathbb\{E\}\[\\exp\(\\lambda s\(\\rvr\\mid w\)\)\]\\to\\exp\(\\lambda s\(\\bar\\rvr\\mid w\)\),\(24\)for every fixedλ\>0\\lambda\>0\. Thusvt,i\(𝐱t\)→r¯iv\_\{t,i\}\(\{\\mathbf\{x\}\}\_\{t\}\)\\to\\bar\{r\}\_\{i\}andvt⋆\(𝐱t∣w\)→s\(𝐫¯∣w\)v\_\{t\}^\{\\star\}\(\{\\mathbf\{x\}\}\_\{t\}\\mid w\)\\to s\(\\bar\{\\mathbf\{r\}\}\\mid w\)\. Continuity ofs\(⋅∣w\)s\(\\cdot\\mid w\)then givesvt\(𝐱t∣w\)=s\(𝐯t\(𝐱t\)∣w\)→s\(𝐫¯∣w\)v\_\{t\}\(\{\\mathbf\{x\}\}\_\{t\}\\mid w\)=s\(\{\\mathbf\{v\}\}\_\{t\}\(\{\\mathbf\{x\}\}\_\{t\}\)\\mid w\)\\to s\(\\bar\{\\mathbf\{r\}\}\\mid w\), so\|vt\(𝐱t∣w\)−vt⋆\(𝐱t∣w\)\|→0\\absolutevalue\{v\_t\(\\rvx\_t\\mid w\)\-v\_t^\\star\(\\rvx\_t\\mid w\)\}\\to 0\. ∎
##### Remark on nonlinear scalarizations\.
The low\-temperature statement above is intentionally asymptotic\. For general nonlinear scalarizations,vt⋆\(𝐱t∣w\)≤vt\(𝐱t∣w\)v\_\{t\}^\{\\star\}\(\{\\mathbf\{x\}\}\_\{t\}\\mid w\)\\leq v\_\{t\}\(\{\\mathbf\{x\}\}\_\{t\}\\mid w\)need not hold at every finiteλ\\lambda\. The monotonicity condition is also essential: negative\-distance Tchebycheff scalarization,s\(𝐚∣w,z\)=−maxiwi\|ai−zi\|s\(\{\\mathbf\{a\}\}\\mid w,z\)=\-\\max\_\{i\}w\_\{i\}\\absolutevalue\{a\_i\-z\_i\}, is monotone only on domains where each objective lies on the same side of the ideal pointziz\_\{i\}and improvement means moving towardziz\_\{i\}\. If feasible rewards can overshoot the ideal, the componentwise essential supremum may move away from the ideal, and the optimistic bound can fail\. In applications that use such scalarizations, the proposition applies either after restricting to the monotone region or after redefining the backed\-up coordinates as monotone improvement scores relative to the target\.
## Appendix GAdditional Experimental Details and Results
### G\.1DiffCSP
For the DiffCSP experiment, we sweep progressive\-widening parametersα\\alpha, C, and inverse temperatureλ\\lambdaat a 1M\-NFE budget, selecting the configuration that maximizes the stability metric\. The chosen configuration —α=0\.3\\alpha=0\.3,C=0\.5C=0\.5andλ=10\\lambda=10— is then evaluated at all reported NFE budgets\. The batch size is set to 50\. Novelty is computed against DiffCSP’s MP\-20 training set\.
We compare against SMC\[[Singhal et al\., 2025](https://arxiv.org/html/2611.17462#bib.bib28)\], SVDD\-PM\[[Li et al\., 2024](https://arxiv.org/html/2611.17462#bib.bib43)\], and TreeG\-SD\[[Guo et al\., 2026](https://arxiv.org/html/2611.17462#bib.bib44)\], which are each training\-free and gradient\-free steering methods for diffusion models\. SVDD\-PM branches each particle into1010next\-state candidates and scores their Tweediex^0\\hat\{x\}\_\{0\}estimates\. TreeG\-SD samples1010perturbed clean targets and moves toward the highest\-reward candidate\. In our DiffCSP adaptation, TreeG\-SD perturbs fractional coordinates and lattice variables while holding atom types fixed, since Gaussian perturbations of discrete type logits would produce arbitrary compositions\. Because Tweedie estimates at high noise levels generally do not represent physically meaningful crystals, we activate branching after certain noise levels:t≤900t\\leq 900for SVDD\-PM andt≤600t\\leq 600for TreeG\-SD, wherettdecreases from 1000 \(most noise\) to 0 \(no noise\)\. SMC instead uses full terminal rollouts to calculate its resampling potentials\. SMC is run with 400 particles over 1000 denoising steps, resampling every 20 steps fort<990t<990\. At each resample point, the rewards are estimated by a full rollout tot=0t=0, resulting in≈\\approx10M NFEs for rollout\-based reward estimation\. All reported NFE budgets fall within 2 % of Best\-of\-N’s nominal 10M\-NFE budget\.
### G\.2Wall\-clock time comparisons
In[Table5](https://arxiv.org/html/2611.17462#A7.T5), we compare the wall\-clock times for generating samples for all methods at≈\\approx10M NFEs\. For DiffCSP, the overhead is mostly from reward computation since the denoising itself is very fast\. In[Table6](https://arxiv.org/html/2611.17462#A7.T6), we compare the wall\-clock times for 1M NFEs using MatterGen\. MatterGen is much slower, so the reward calculation time makes less of a difference\.
Table 5:Approximate wall\-clock times on a single A100 for the 10M NFE budget for DiffCSP, with a batch size of 50\.MethodRewardTimeDTS\-EhullE\_\{\\text\{hull\}\}≈\\approx5\.9 hDTS0≈\\approx5\.2 hBest\-of\-N baseline0≈\\approx1\.5 hSMC\-EhullE\_\{\\text\{hull\}\}≈\\approx2\.2 hTable 6:Approximate wall\-clock times on a single RTX8000 for the 1M NFE budget for MatterGen, with a batch size of 50\.MethodRewardTimeDTS\-EhullE\_\{\\text\{hull\}\}≈\\approx3\.2 hDTS0≈\\approx3\.0 hBest\-of\-N baseline0≈\\approx2\.7 h
### G\.3Matched\-output count results
Due to trajectory reuse, DTS produces more candidate structures than regular best\-of\-N sampling\. As such, we present results for the stability experiment of[Section4\.1](https://arxiv.org/html/2611.17462#S4.SS1)at a fixed number of output samples, rather than presenting results over all samples created using the same NFEs\.
##### MatterGen\.
All methods are evaluated atN=2500N=2500terminal candidates, matching the baseline\. Two selection rules are shown: first\-NNby NFE \(stopping the search early\) and best\-NNby predictedEhullE\_\{\\text\{hull\}\}\.
Table 7:MatterGen results with matched output counts \(N=2500N=2500\), using first\-NNand best\-NNselection\.α\\alphaMethodSelection\# SUNStable \(%\)SUN unique systemsAvg\.EhullE\_\{\\text\{hull\}\}/atom—MatterGen baselineAll 2500868\.0868\.068\.468\.4848\.3\\mathbf\{848\.3\}0\.08580\.08580\.1DTS \(Null\-Reward\)First\-NN733\.0±26\.8733\.0\\pm 26\.866\.9±1\.366\.9\\pm 1\.3544\.0±25\.2544\.0\\pm 25\.20\.0904±0\.00380\.0904\\pm 0\.00380\.3DTS \(Null\-Reward\)First\-NN789\.3±32\.7789\.3\\pm 32\.765\.7±1\.565\.7\\pm 1\.5673\.7±17\.6673\.7\\pm 17\.60\.0911±0\.00390\.0911\\pm 0\.00390\.5DTS \(Null\-Reward\)First\-NN813\.3±24\.0813\.3\\pm 24\.066\.7±1\.466\.7\\pm 1\.4749\.7±23\.4749\.7\\pm 23\.40\.0901±0\.00350\.0901\\pm 0\.00350\.1DTS \+EhullE\_\{\\text\{hull\}\}rewardFirst\-NN734\.3±52\.2734\.3\\pm 52\.272\.7±0\.7\\mathbf\{72\.7\\pm 0\.7\}533\.0±19\.1533\.0\\pm 19\.10\.0792±0\.0007\\mathbf\{0\.0792\\pm 0\.0007\}0\.3DTS \+EhullE\_\{\\text\{hull\}\}rewardFirst\-NN866\.0±35\.5866\.0\\pm 35\.570\.6±0\.970\.6\\pm 0\.9714\.7±27\.8714\.7\\pm 27\.80\.0843±0\.00210\.0843\\pm 0\.00210\.5DTS \+EhullE\_\{\\text\{hull\}\}rewardFirst\-NN921\.0±22\.5\\mathbf\{921\.0\\pm 22\.5\}69\.3±1\.369\.3\\pm 1\.3833\.3±14\.9833\.3\\pm 14\.90\.0861±0\.00280\.0861\\pm 0\.00280\.1DTS \(Null\-Reward\)Best\-NN859\.0±30\.4859\.0\\pm 30\.4100\.0±0\.0\\mathbf\{100\.0\\pm 0\.0\}629\.3±33\.9629\.3\\pm 33\.90\.0319±0\.00060\.0319\\pm 0\.00060\.3DTS \(Null\-Reward\)Best\-NN1095\.3±55\.41095\.3\\pm 55\.491\.4±2\.291\.4\\pm 2\.2904\.3±29\.4904\.3\\pm 29\.40\.0500±0\.00070\.0500\\pm 0\.00070\.5DTS \(Null\-Reward\)Best\-NN973\.7±26\.9973\.7\\pm 26\.980\.0±1\.780\.0\\pm 1\.7894\.7±27\.4894\.7\\pm 27\.40\.0599±0\.00200\.0599\\pm 0\.00200\.1DTS \+EhullE\_\{\\text\{hull\}\}rewardBest\-NN710\.3±25\.6710\.3\\pm 25\.6100\.0±0\.0\\mathbf\{100\.0\\pm 0\.0\}506\.3±13\.1506\.3\\pm 13\.10\.0185±0\.0023\\mathbf\{0\.0185\\pm 0\.0023\}0\.3DTS \+EhullE\_\{\\text\{hull\}\}rewardBest\-NN1207\.0±29\.8\\mathbf\{1207\.0\\pm 29\.8\}98\.3±1\.398\.3\\pm 1\.3969\.7±27\.9969\.7\\pm 27\.90\.0454±0\.00150\.0454\\pm 0\.00150\.5DTS \+EhullE\_\{\\text\{hull\}\}rewardBest\-NN1087\.3±22\.81087\.3\\pm 22\.882\.4±1\.582\.4\\pm 1\.5979\.7±15\.2\\mathbf\{979\.7\\pm 15\.2\}0\.0581±0\.00240\.0581\\pm 0\.0024
These results show that given the same NFE but selecting the highest\-reward samples, DTS is able to generate far more SUN candidates\. Alternatively, for the same target number of SUN crystals, DTS is able to generate using fewer NFEs\. However, using a smaller value ofα\\alphaleads to more trajectory reuse and less diversity, which can result in fewer SUN crystals\.
##### DiffCSP\.
We evaluate a matched output count of 1,000 and 10,000 candidates at the 1M\- and 10M\-NFE budgets, respectively\. We use three protocols to select theseNNcandidates: \(i\) the firstNNstructures generated, which involves no post\-hoc energy selection; \(ii\) theNNstructures with the lowestEhullE\_\{\\text\{hull\}\}; and \(iii\) a formula\-diverse selection that first maximizes distinct chemical\-formula coverage and then ranks candidates within each formula byEhullE\_\{\\text\{hull\}\}\.
Table 8:Matched\-output comparisons with 1,000 candidates at the 1M\-NFE budget and 10,000 candidates at the 10M\-NFE budget\.SelectionMethodStability \(%\)↑\\uparrowSUN \(%\)↑\\uparrowMeanEhullE\_\{\\text\{hull\}\}↓\\downarrowMatched 1,000 output candidatesFirst 1,000Best\-of\-NN30\.7±0\.930\.7\\pm 0\.921\.0±1\.321\.0\\pm 1\.30\.314±0\.0310\.314\\pm 0\.031DTS \(Null\-Reward\)31\.3±0\.331\.3\\pm 0\.322\.2±1\.322\.2\\pm 1\.30\.349±0\.0620\.349\\pm 0\.062DTS35\.2±5\.6\\mathbf\{35\.2\\pm 5\.6\}24\.2±4\.3\\mathbf\{24\.2\\pm 4\.3\}0\.278±0\.027\\mathbf\{0\.278\\pm 0\.027\}TreeG\-SD33\.5±4\.833\.5\\pm 4\.823\.2±3\.423\.2\\pm 3\.40\.317±0\.0340\.317\\pm 0\.034LowestEhullE\_\{\\text\{hull\}\}1,000Best\-of\-NN30\.7±0\.930\.7\\pm 0\.921\.0±1\.321\.0\\pm 1\.30\.314±0\.0310\.314\\pm 0\.031DTS \(Null\-Reward\)43\.2±2\.043\.2\\pm 2\.030\.4±2\.830\.4\\pm 2\.80\.145±0\.0020\.145\\pm 0\.002DTS49\.7±6\.6\\mathbf\{49\.7\\pm 6\.6\}33\.9±4\.8\\mathbf\{33\.9\\pm 4\.8\}0\.123±0\.018\\mathbf\{0\.123\\pm 0\.018\}TreeG\-SD33\.5±4\.833\.5\\pm 4\.823\.2±3\.423\.2\\pm 3\.40\.317±0\.0340\.317\\pm 0\.034Formula\-diverse 1,000Best\-of\-NN30\.7±0\.930\.7\\pm 0\.921\.0±1\.321\.0\\pm 1\.30\.314±0\.0310\.314\\pm 0\.031DTS \(Null\-Reward\)41\.9±1\.841\.9\\pm 1\.829\.9±2\.829\.9\\pm 2\.80\.151±0\.0020\.151\\pm 0\.002DTS48\.3±6\.4\\mathbf\{48\.3\\pm 6\.4\}33\.6±4\.7\\mathbf\{33\.6\\pm 4\.7\}0\.128±0\.018\\mathbf\{0\.128\\pm 0\.018\}TreeG\-SD33\.5±4\.833\.5\\pm 4\.823\.2±3\.423\.2\\pm 3\.40\.317±0\.0340\.317\\pm 0\.034Matched 10,000 output candidatesFirst 10,000Best\-of\-NN31\.1±0\.331\.1\\pm 0\.320\.3±0\.120\.3\\pm 0\.10\.435±0\.0350\.435\\pm 0\.035DTS \(Null\-Reward\)30\.3±0\.630\.3\\pm 0\.619\.6±0\.919\.6\\pm 0\.90\.338±0\.0200\.338\\pm 0\.020DTS40\.5±6\.6\\mathbf\{40\.5\\pm 6\.6\}24\.4±2\.1\\mathbf\{24\.4\\pm 2\.1\}0\.317±0\.125\\mathbf\{0\.317\\pm 0\.125\}TreeG\-SD31\.6±0\.531\.6\\pm 0\.520\.8±0\.020\.8\\pm 0\.00\.425±0\.0330\.425\\pm 0\.033LowestEhullE\_\{\\text\{hull\}\}10,000Best\-of\-NN31\.1±0\.331\.1\\pm 0\.320\.3±0\.120\.3\\pm 0\.10\.435±0\.0350\.435\\pm 0\.035DTS \(Null\-Reward\)47\.8±1\.347\.8\\pm 1\.330\.3±0\.230\.3\\pm 0\.20\.121±0\.0030\.121\\pm 0\.003DTS64\.2±10\.4\\mathbf\{64\.2\\pm 10\.4\}37\.3±3\.5\\mathbf\{37\.3\\pm 3\.5\}0\.080±0\.022\\mathbf\{0\.080\\pm 0\.022\}TreeG\-SD31\.6±0\.531\.6\\pm 0\.520\.8±0\.020\.8\\pm 0\.00\.425±0\.0330\.425\\pm 0\.033Formula\-diverse 10,000Best\-of\-NN31\.1±0\.331\.1\\pm 0\.320\.3±0\.120\.3\\pm 0\.10\.435±0\.0350\.435\\pm 0\.035DTS \(Null\-Reward\)40\.5±0\.940\.5\\pm 0\.929\.0±0\.329\.0\\pm 0\.30\.149±0\.0020\.149\\pm 0\.002DTS50\.0±5\.5\\mathbf\{50\.0\\pm 5\.5\}35\.7±3\.3\\mathbf\{35\.7\\pm 3\.3\}0\.120±0\.014\\mathbf\{0\.120\\pm 0\.014\}TreeG\-SD31\.6±0\.531\.6\\pm 0\.520\.8±0\.020\.8\\pm 0\.00\.425±0\.0330\.425\\pm 0\.033
SVDD\-PM and SMC are omitted only from these matched\-output comparisons because they produce fewer than 1,000 and 10,000 terminal structures at the corresponding NFE budgets\. The first\-NNprotocol shows that DTS’s improvement is not solely due to post\-hoc selection\. The other protocols demonstrate the practical benefit of trajectory reuse: at the same inference budget, DTS produces additional candidates from which a fixed\-size downstream batch can be selected, while the formula\-diverse protocol shows that the improvement persists after controlling for composition concentration\.
### G\.4Dielectric Constant and Bandgap
More results for the multi\-objective dielectric constant vs bandgap experiment are included in[Table9](https://arxiv.org/html/2611.17462#A7.T9)\. We compare across different values of progressive widening parameterα\\alpha\. The conditional version of MatterGen \(DTS\-C\) is fine\-tuned over 200 epochs on the MP\-20 dataset, where each crystal is labeled with dielectric constants and bandgaps predicted by ALIGNN\[[Choudhary and DeCost, 2021](https://arxiv.org/html/2611.17462#bib.bib18)\]and using a conditioning dropout of 20%\. We use Tchebycheff scalarization throughout this experiment\. The conditional model is conditioned on the Dielectric constant and bandgaps of samples from the Pareto front of the training set, determined using Tchebycheff scalarization\. Preferences are sampled from a Dirichlet distribution withα=1\\alpha=1\. We also compare against a variant that uses an unconditional MatterGen model as its base \(DTS\-U\)\.
Table 9:Summary metrics for all evaluated methods on the dielectric/bandgap multi\-objective task\.\|PF\|\|PF\|= Pareto front size\. For methods with multiple seeds, entries are mean±\\pmstd across seeds\. Threshold columns count structures where bandgap×\\timesdielectric≥\\geqthreshold\.MethodNN\|PF\|\|PF\|N≥30N\_\{\\geq 30\}N≥60N\_\{\\geq 60\}N≥120N\_\{\\geq 120\}N≥240N\_\{\\geq 240\}HV↑\\uparrowGD\+↓\\downarrowIGD\+↓\\downarrowR2↓\\downarrowTraining set27,136204,1801,9205261090\.3090\.0000\.0000\.149MatterGen \(cond\.\)665±\\pm014±\\pm1262±\\pm3138±\\pm628±\\pm24±\\pm00\.118±\\pm0\.0110\.098±\\pm0\.0240\.143±\\pm0\.0050\.292±\\pm0\.002DTS\-Cα\\alpha=0\.5645±\\pm4016±\\pm1402±\\pm54234±\\pm2767±\\pm1217±\\pm40\.173±\\pm0\.0570\.054±\\pm0\.0150\.180±\\pm0\.0450\.338±\\pm0\.040DTS\-Cα\\alpha=0\.3674±\\pm4415±\\pm3403±\\pm31231±\\pm2368±\\pm619±\\pm10\.236±\\pm0\.0520\.054±\\pm0\.0120\.120±\\pm0\.0430\.287±\\pm0\.037DTS\-Cα\\alpha=0\.1804±\\pm4611±\\pm1455±\\pm35242±\\pm4959±\\pm2113±\\pm50\.155±\\pm0\.0320\.030±\\pm0\.0250\.214±\\pm0\.0280\.375±\\pm0\.024DTS\-Uα\\alpha=0\.11,4331219860930\.0610\.1800\.2490\.369DTS\-Uα\\alpha=0\.31,1721618757600\.0310\.2110\.1940\.325DTS\-Uα\\alpha=0\.51,0572311637200\.0240\.2660\.2010\.330
### G\.5Symmetry\-Conditioned Generation
We ran the symmetry\-conditioned experiment of[Section4\.2](https://arxiv.org/html/2611.17462#S4.SS2)with four target space groups, chosen to span a range of rarity in the MatterGen training distribution\. All runs use the space\-group\-conditional MatterGen variant for both DTS and the best\-of\-NNbaseline; the reward is the indicatorr\(𝐱0\)=𝟙\[SG\(𝐱0\)=g⋆\]r\(\{\\mathbf\{x\}\}\_\{0\}\)=\\mathbbm\{1\}\[\\mathrm\{SG\}\(\{\\mathbf\{x\}\}\_\{0\}\)=g^\{\\star\}\], for target groupg⋆g^\{\\star\}\.
##### Target space groups\.
We selected four space groups that appear rarely in the training set\. Their frequencies span from∼\\sim10\-2to∼\\sim10\-5, allowing us to see how proposal\-distribution coverage interacts with search\.
##### Results\.
[Figure7](https://arxiv.org/html/2611.17462#A7.F7)shows cumulative SUN∩\\captarget\-SG counts versus NFE budget for each target\. Three of the four targets \(SG 40, 146, 204\) follow the pattern of[Section4\.2](https://arxiv.org/html/2611.17462#S4.SS2): DTS produces substantially more on\-target SUN structures than best\-of\-NN, with the gap widening as the budget grows\. The exception is SG 138, the rarest target at∼\\sim10\-5% of the training set: neither method produces any SUN structures matching this space group within the 500K\-NFE budget\. All runs useα=0\.5\\alpha=0\.5,C=1\.0C=1\.0,λ=100\\lambda=100\. The NFE budget is 500K per method per target\. Results are from a single seed per \(method, target\) pair\.
Figure 7:Full view symmetry\-conditioned experiment\. Color encodes method \(DTS vs\. MatterGen best\-of\-NN\); linestyle encodes metric\.
## Appendix HImportance Sampling Details
### H\.1Importance Sampling Ablations
In addition to the ablation study in[Section4\.4](https://arxiv.org/html/2611.17462#S4.SS4), we ran ablations on importance sampling \(IS\) correction for the Space Group and Dielectric/Bandgap experiments\. We note that in our method, the purpose of the correction is not to improve reward or yield\. A conditional model can provide a more efficient proposal, but it also changes the proposal distribution\. The likelihood\-ratio correction is intended to reduce this proposal\-induced bias relative to the reward\-tilted unconditional target\.
##### Target space group\.
We ablate the correction when generating crystals in space group 204 using a space\-group\-conditioned MatterGen proposal\. We evaluate over 3 seeds and10610^\{6\}NFEs\.
Table 10:IS ablation for space group 204 targeting \(3 seeds, 1M NFEs\)\.Method\# SG 204 and SUNDTS with IS88\.0±16\.488\.0\\pm 16\.4DTS without IS75\.0±24\.775\.0\\pm 24\.7
##### Dielectric constant and bandgap\.
We next ablate IS on the anti\-correlated dielectric/bandgap multi\-objective task\. All runs useλ=100\\lambda=100and progressive wideningα=0\.5\\alpha=0\.5\. Note that this evaluation uses only10610^\{6\}NFEs, and is not directly comparable to the main results of[Section4\.3](https://arxiv.org/html/2611.17462#S4.SS3), which used2\.5×1062\.5\\times 10^\{6\}NFEs\.
We compare DTS with importance sampling correction, DTS without IS, and unconditional MatterGen best\-of\-NN\.
Table 11:IS ablation on the dielectric/bandgap multi\-objective task \(1M NFEs\)\.N≥kN\_\{\\geq k\}denotes the number of structures satisfying thresholdkk\.MethodSUN StructuresHV↑\\uparrowGD\+↓\\downarrowR2↓\\downarrowN≥30N\_\{\\geq 30\}N≥60N\_\{\\geq 60\}N≥120N\_\{\\geq 120\}N≥240N\_\{\\geq 240\}Best\-of\-NN264±0264\\pm 00\.072±0\.0350\.072\\pm 0\.0350\.183±0\.0600\.183\\pm 0\.0600\.306±0\.0080\.306\\pm 0\.008103±1103\\pm 157±357\\pm 313±213\\pm 21±11\\pm 1DTS, no IS325±34325\\pm 340\.084±0\.0210\.084\\pm 0\.0210\.149±0\.0220\.149\\pm 0\.0220\.371±0\.0240\.371\\pm 0\.024154±16154\\pm 1678±878\\pm 820±420\\pm 43±13\\pm 1DTS, IS286±12286\\pm 120\.068±0\.0240\.068\\pm 0\.0240\.152±0\.0370\.152\\pm 0\.0370\.352±0\.0130\.352\\pm 0\.013146±7146\\pm 777±577\\pm 520±320\\pm 32±02\\pm 0
Removing the IS correction leaves most Pareto metrics within one seed standard deviation of the corrected run, so there is no strong degradation from dropping importance\-corrected backup/selection in this setting\.
### H\.2Importance Sampling Statistics
We report edge log\-weight distributions and normalized effective sample size,ESS/n\\mathrm\{ESS\}/n, for the space\-group \([Table12](https://arxiv.org/html/2611.17462#A8.T12)\), dielectric/bandgap \([Table13](https://arxiv.org/html/2611.17462#A8.T13)\), and density/bulk modulus experiments \([Table14](https://arxiv.org/html/2611.17462#A8.T14)\)\. Each tree edge spans 100 diffusion steps between consecutive branching points; its log\-weight is the sum of per\-step log\-ratios over that interval, and the quantiles below are computed over all edges in each interval\. At each branching node,
ESSn=\(∑i=1nηi\)2n∑i=1nηi2\\frac\{\\mathrm\{ESS\}\}\{n\}=\\frac\{\\left\(\\sum\_\{i=1\}^\{n\}\\eta\_\{i\}\\right\)^\{2\}\}\{n\\sum\_\{i=1\}^\{n\}\\eta\_\{i\}^\{2\}\}\(25\)across itsnnchildren: a value of 1 indicates equal weights, while a value of approximately1/n1/nindicates concentration on a single child\.
#### H\.2\.1Space group\.
Table 12:Importance sampling statistics for the space group 204 experiment\.TimestepsMedianESS/n\\mathrm\{ESS\}/n1000–901–100\.000\.348−0\.03\-0\.03−0\.0001\-0\.00010\.03900–8011695\.290\.763−0\.05\-0\.05−0\.0000\-0\.00000\.05800–7013045\.300\.688−0\.08\-0\.08−0\.0001\-0\.00010\.07700–6014773\.360\.501−0\.40\-0\.40−0\.0011\-0\.00110\.36600–5011882\.880\.500−15\.20\-15\.20−1\.40\-1\.401\.50500–401212\.760\.500−19\.79\-19\.79−2\.38\-2\.380\.71400–30114\.000\.458−1\.65\-1\.65−0\.07\-0\.070\.58
For most of the trajectory \(t=1000t=1000–701701\), conditional and unconditional transitions remain closely aligned: the median log\-weight is approximately 0, the central 90% lies roughly within\[−0\.08,0\.07\]\[\-0\.08,0\.07\], and the medianESS/n\\mathrm\{ESS\}/nis 0\.69–0\.76 at levels with∼5\{\\sim\}5children\. Concentration peaks betweent=600t=600and401401, where the log\-weight tail widens and the medianESS/n\\mathrm\{ESS\}/ndrops to 0\.50\. This effect is localized rather than trajectory\-wide; thet=500t=500–401401and400400–301301bins contain only 21 and 1 branching nodes, respectively, so those estimates are noisier\.
#### H\.2\.2Dielectric constant and bandgap\.
Table 13:Importance sampling statistics for the dielectric/bandgap experiment\.TimestepsMedianESS/n\\mathrm\{ESS\}/n1000–901373\.000\.508−0\.002\-0\.002−0\.000\-0\.0000\.002900–8012106\.290\.709−0\.002\-0\.002−0\.000\-0\.0000\.002800–7014813\.590\.795−0\.003\-0\.003−0\.000\-0\.0000\.003700–6012843\.330\.791−0\.090\-0\.090−0\.000\-0\.0000\.087600–5011303\.810\.519−0\.681\-0\.681−0\.021\-0\.0210\.482500–401763\.780\.547−0\.804\-0\.804−0\.028\-0\.0280\.495400–301463\.860\.657−0\.457\-0\.457−0\.007\-0\.0070\.319300–201313\.330\.623−0\.370\-0\.370−0\.005\-0\.0050\.273200–101152\.980\.717−0\.416\-0\.416−0\.006\-0\.0060\.299100–00152\.000\.746−0\.508\-0\.508−0\.007\-0\.0070\.348
For dielectric/bandgap, the medianESS/n\\mathrm\{ESS\}/nis 0\.71–0\.80 betweent=900t=900and601601, then dips to 0\.52–0\.55 betweent=600t=600and401401, before recovering to 0\.62–0\.75\.
Weights remain usable because DTS corrects locally between consecutive branching nodes rather than accumulating a single importance ratio over the full 1,000\-step trajectory\. Each decision uses only the ratio from the preceding 100\-step segment\.
Overall, the weights are neither uniformly benign nor fully degenerate: most branching levels retain moderate ESS, and both tasks show a localized mid\-trajectory interval with broader log\-weight tails and fewer effective children\.
#### H\.2\.3Density/bulk modulus\.
Table 14:Importance sampling statistics for the density/bulk modulus experiment\.TimestepsMedianESS/n\\mathrm\{ESS\}/n1000–900480\.740−0\.0145\-0\.0145−0\.0003\-0\.00030\.0133900–8001220\.816−0\.0145\-0\.0145−0\.0001\-0\.00010\.0138800–7002440\.826−0\.0174\-0\.0174−0\.0001\-0\.00010\.0167700–6003140\.658−0\.0777\-0\.0777−0\.0002\-0\.00020\.0742600–5001610\.674−0\.4631\-0\.4631−0\.0093\-0\.00930\.3683500–400210\.803−0\.4804\-0\.4804−0\.0100\-0\.01000\.3521400–30020\.729−0\.1806\-0\.1806−0\.0016\-0\.00160\.1508300–20010\.995−0\.1350\-0\.1350−0\.0013\-0\.00130\.1199200–1000–−0\.1571\-0\.1571−0\.0013\-0\.00130\.1384100–00–−0\.2253\-0\.2253−0\.0032\-0\.00320\.1930Similar Articles
Accelerating Diffusion Sampling via Speculative Draft Trees
This paper introduces speculative draft trees to accelerate diffusion model sampling by enriching candidate drafts and using greedy rejection sampling, achieving up to 8.3% speedup over baseline methods.
Temporal Difference Learning for Diffusion Models
This paper introduces a temporal difference (TD) learning objective for diffusion models that enforces cross-time consistency along the denoising trajectory. It reformulates denoising as a reinforcement learning policy evaluation problem, showing significant improvements in sample quality (FID), especially for few-step samplers.
FluxLite: Inference-Time Proposal Control for Discrete Diffusion Models
FluxLite introduces a training-free, inference-time proposal-control framework for discrete diffusion models that compensates jump-rate perturbations via a graph-divergence term in the Feynman-Kac potential, yielding two samplers (HEU and D-VCG) that substantially reduce reweighting variance and sampling error over standard SMC baselines.
Rank-Aware Speculative Sampling for Diffusion Draft Trees
论文提出 Rank-Aware Speculative Sampling(RASS),一种基于 rank-aware list coupling 的扩散草稿树验证规则,通过对候选按 proposal-target 均值位移排序并优化秩权重以最小化总变差,在保持精确采样的同时加速扩散生成,在 CIFAR-10 上相比 D-GRS 节省约 20% 目标模型评估。
Synthetic Crystal Data Generation for Diffusion Pretraining in Material Property Prediction
The paper introduces a synthetic crystal structure generation pipeline combining perturb-and-relax, MatterGen sampling, and protostructure enumeration to pretrain graph neural networks for material property prediction without DFT labels, outperforming random initialization across seven MatBench tasks.