LLMs as Acquisition Policies for Finite-Pool Materials Optimization: A Controlled Study

Dino-Rober Demir; Florian Le Bronnec; Rio Yokota

cs.AIAug 20, 2026

LLMs as Acquisition Policies for Finite-Pool Materials Optimization: A Controlled Study

Authors: Dino-Rober Demir, Florian Le Bronnec, Rio Yokota

Organizations: RIKEN Center for Computational Science, Tokyo, Japan

Abstract

Discovering materials with desirable properties often requires searching large candidate spaces while experimental or computational evaluations remain costly. Active learning addresses this challenge by using previous observations to select which candidate to evaluate next, typically through probabilistic surrogate models. We investigate whether open-weight large language models (LLMs) can serve as standalone acquisition policies in this setting. We evaluate five LLMs across four retrospective finite-pool materials optimization tasks under different candidate-presentation strategies and compare them with random selection and conventional Gaussian-process methods. LLM policies generally reach the global optimum in fewer iterations than random selection, indicating that they provide a useful acquisition signal without task-specific training. Their performance relative to Gaussian-process methods is mixed: conventional acquisition performs better on most tasks, while LLMs match or outperform it in some settings. Performance varies substantially across tasks, models, initializations, and candidate presentations, with no LLM approach performing best across all tasks. Overall, open-weight LLMs show potential as acquisition policies for finite-pool materials search, although their reliability remains sensitive to the task and to how candidates and scientific context are presented.

Figures & tables

Task	Candidates	Features	Target
Kerr Rotation	921	3	Kerr rotation (mrad)
Coercivity	921	3	coercivity (mT)
Matbench Steels	312	14	yield strength (MPa)
Electrostrain	81	15	electrostrain (%)

Table 1: Retrospective optimization tasks.

Figure 1: Best-so-far optimization trajectories using Gemma 4 31B for the LLM policies. At iteration

t

, each curve shows the mean percentage gap

100(y^{\star}-b_{t})/|y^{\star}|

across the 25 shared initialization seeds. Shading denotes pointwise normal-approximation 95% confidence intervals, computed as

\bar{g}_{t}\pm 1.96s_{t}/\sqrt{25}

. We show Gemma 4 31B because it ranks among the strongest LLMs across all four tasks.

Whole-pool selection
Model	Kerr Rotation	Coercivity	Matbench Steels	Electrostrain
Random	462.48 $\pm$ 262.24	448.09 $\pm$ 269.63	159.31 $\pm$ 91.28	39.89 $\pm$ 23.00
GP-EI	8.40 $\pm$ 2.78	91.80 $\pm$ 62.77	42.72 $\pm$ 27.24	18.48 $\pm$ 14.53
Gemma4	31.28 $\pm$ 17.24	209.88 $\pm$ 134.05	44.16 $\pm$ 23.36	14.36 $\pm$ 7.71
Qw27B	27.08 $\pm$ 13.59	195.20 $\pm$ 114.39	48.24 $\pm$ 23.43	19.44 $\pm$ 10.40
Qw35B	45.72 $\pm$ 20.28	289.88 $\pm$ 127.44	50.36 $\pm$ 24.89	27.64 $\pm$ 18.15

Table 2: Iterations to the global optimum over 25 initializations (mean

\pm

standard deviation), except for random selection which is initialized over 1000 seeds. Lower is better. Baselines are shown once because they do not depend on an LLM presentation protocol.

Figure 2: Batch-size sensitivity of tournament selection on Matbench Steels and Kerr Rotation for runs of 25 seeds. Curves report the mean number of iterations to the optimum across seeds, with uncertainty shown shaded as the mean

\pm

\mathrm{std}/\sqrt{n_{\mathrm{seeds}}}

Panel of Figure 3 — Figure 3: Position bias for candidates selected by DeepSeek-V4-Flash and Gemma 4 31B on Matbench Steels . Dashed lines in panel (a) indicate the uniform-selection expectation, which may not be horizontal because the pool shrinks across iterations. Panel (b) reports $D_{\mathrm{KL}}(q\,\|\,u)$ for all evaluated models.

Figure 4: Optimization performance on Matbench Steels , with and without materials context across 25 initializations. Bars report the mean iteration at which the optimum is reached, with normal-approximation 95% confidence intervals, computed as mean

\pm

1.96\,\mathrm{std}/\sqrt{n_{\mathrm{seeds}}}

. We also report the mean iterations to optimum of random selection with a dotted line.

Figure 5: Explorative behavior of two selection protocols on Matbench Steels for seed 4 with Gemma 4 31B. Dots represent points of the pool projected onto the first two principal components, with the PCA being scaled over every pool entry. On this seed, GP-EI and batch-tournament reach the optimum after 60 and 56 iterations, respectively.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 6 — Figure 6: Batch-tournament candidate selection for an example with $N=16$ unobserved candidates and batch size $b=4$ .

Figure 7: Batch-size sensitivity of tournament selection on Coercivity and Electrostrain . Curves report the mean number of iterations to the optimum across seeds, with uncertainty shown shaded as the mean

\pm

\mathrm{std}/\sqrt{n_{\mathrm{seeds}}}

	Electrostrain		Matbench Steels		Coercivity		Kerr Rotation
Models	Whole	Batch	Whole	Batch	Whole	Batch	Whole	Batch
Gemma4	0.998	2.27	1.02	2.11	1.60	2.08	2.21	2.02
DS	1.36	2.29	1.47	2.13	2.48	2.35	2.60	2.15
Qw27B	0.929	2.10	0.852	2.04	2.18	2.06	2.17	2.04
Qw35B	1.26	2.20	0.747	2.12	1.67	2.24	2.77	2.08
Qw397B	1.04	2.20	1.36 ∗	2.08	–	1.84 †	–	2.06

Appendix

Table 3: Position-bias metrics for all model-task combinations. We report

D_{\mathrm{KL}}

for whole-pool and batch-tournament selection with

b=20

, except for the proxy configurations marked with an asterisk. Missing measurements are denoted by ’-’.

Model	Random	GP-EI	Whole-pool	Generation-and-matching
Qw27B	159.31 $\pm$ 91.28	42.72 $\pm$ 27.24	48.24 $\pm$ 23.43	123.60 $\pm$ 63.60
Gemma4	159.31 $\pm$ 91.28	42.72 $\pm$ 27.24	44.16 $\pm$ 23.36	48.80 $\pm$ 26.14

Appendix

Table 4: Generation-and-matching performance on Matbench Steels compared with random, GP-EI, and whole-pool selection. We report iterations to the global optimum over 25 initializations (mean

\pm

standard deviation), except for random selection which is initialized over 1000 seeds.

Figure 8: Comparison of evolution of fitted

\widehat{\beta}_{t}

on Matbench Steels between Gemma 4 31B, Random and GP-EI. Results are averaged across seeds for each iteration, when at least half of the seeds have not converged yet.

Large Language Model-Guided Optimization

Large Language Model-Guided Optimization uses language models to propose, steer, and refine solutions to hard search and design problems.

See all 608 papers

From Blind Guess to Informed Judgment: Teaching LLMs to Evaluate Materials by Building Knowledge-Augmented Preference Signals
May 28, 2026Yeyong Yu, Wenya Hu, Xing Wu +1LLM Evaluation Pairwise Preference Evaluation
School of Computer Engineering & Science, Shanghai University, Shanghai, 200444, China. · Center of Materials Informatics and Data Science, Materials Genome2026 Institute, Shanghai University, Shanghai, 200444, China. · Key Laboratory of Silicate Cultural Relics Conservation (Shanghai University), Ministry of Education, China. +1
PRISMat: Policy-Driven, Permutation-Invariant Autoregressive Material Generation
May 15, 2026Claire Schlesinger, Circe Hsu, Peter Schindler +1Materials Science Crystal Structure Generation
Khoury College of Computer Sciences Northeastern University Boston, MA 02115 · College of Engineering Northeastern University Boston, MA 02115
Multi-Agent LLMs for Adaptive Acquisition in Bayesian Optimization
Mar 30, 2026Andrea Carbonati, Mohammadsina Almasi, Hadis AnahidehLarge Language Model-Guided Optimization Exploration-Exploitation Tradeoff
University of Illinois Chicago, IL, USA

May 28, 2026J/K move · Enter open · S save

From Blind Guess to Informed Judgment: Teaching LLMs to Evaluate Materials by Building Knowledge-Augmented Preference Signals

Yeyong Yu, Wenya Hu, Xing Wu, Quan Qian

School of Computer Engineering & Science, Shanghai University, Shanghai, 200444, China. · Center of Materials Informatics and Data Science, Materials Genome2026 Institute, Shanghai University, Shanghai, 200444, China. · Key Laboratory of Silicate Cultural Relics Conservation (Shanghai University), Ministry of Education, China.+1

LLM Evaluation Pairwise Preference Evaluation Materials Science Large Language Model-Guided Scientific Discovery Preference Learning

Open paper AskPDF arXiv

As candidate generation and high-throughput experimentation advance, the primary bottleneck in materials discovery is shifting from property prediction to making reliable evaluations among massive candidate sets. We propose a Knowledge-Augmented Preference Signals Framework, MaterEval, that automatically produces, for the same candidate, two evaluations: an informed judgment that follows expert rules and provides supporting evidence, and a rule-removed blind guess. By pairing the two evaluations as preference data, we guide general-purpose large language models (LLMs), originally lacking materials-specific criteria, from intuitive judgment toward reliable evaluation supported by explicit evidence. To balance throughput, cost, and reliability, we further introduce a fast-slow reasoning scheme that decouples large-scale rapid screening from in-depth review on a small subset. Using high-entropy alloy (HEA) assessment as a case study, we show that, without external retrieval and relying solely on internalized capabilities, small open-source LLMs achieve substantial gains in accuracy, conclusion consistency, and evidence discrimination, approaching the performance of rule-based closed-source LLMs. These results demonstrate that expert rules can be systematically transformed into learnable preference signals, enabling a low-cost and deployable evaluation module for autonomous materials discovery loops.

Abstract

Large Language Model-Guided Optimization

Explore similar work

From Blind Guess to Informed Judgment: Teaching LLMs to Evaluate Materials by Building Knowledge-Augmented Preference Signals

PRISMat: Policy-Driven, Permutation-Invariant Autoregressive Material Generation

Multi-Agent LLMs for Adaptive Acquisition in Bayesian Optimization