TRACE: Transition-Aware Residual Control for Multi-Objective Materials Discovery

Kang Zhou; Yujia Tong; Yong Tao; Jingling Yuan

cs.AIAug 23, 2026

TRACE: Transition-Aware Residual Control for Multi-Objective Materials Discovery

Authors: Kang Zhou, Yujia Tong, Yong Tao, Jingling Yuan

Organizations: Wuhan University of Technology

Abstract

Multi-objective materials discovery with LLM agents is often limited not only by how many candidates can be proposed, but by how effectively each costly property evaluation informs the next search step. Existing agents mainly store evaluated candidates and their scores, so they know which materials succeeded but not which executable edits caused useful property changes. This makes local refinement difficult when objectives compete and an edit that improves one property may damage another. We propose TRACE, a transition-aware residual control framework that treats evaluated edits as the basic unit of feedback. TRACE records each local refinement as a parent-edit-child transition with observed property deltas, aggregates transition evidence to estimate reusable edit effects, and ranks future edits by their predicted ability to reduce the current candidate's remaining constraint violations while avoiding damage to already satisfied objectives. In a controlled same-backbone comparison, TRACE improves over LLEMA, the state-of-the-art LLM-agent baseline, raising macro-average hit rate from 18.13% to 25.96%.

Figures & tables

Figure 1: Motivation. Candidate-level feedback records which candidates perform well, whereas transition-aware feedback captures how executable edits change material properties and uses this evidence to guide subsequent search.

Figure 2: Overview of TRACE. Global LLM proposals support broad exploration, while local Delta edits refine evaluated parents according to their current constraint residuals. Evaluated parent–edit–child transitions update transition memory, whose accumulated property changes guide subsequent edit selection.

Figure 3: Search-budget efficiency across four representative tasks. Curves and shaded regions show the mean and variability of cumulative valid discoveries over five runs; insets report exact-formula redundancy.

Variant	H.R. $\uparrow$	Stab. $\uparrow$	G.R. $\uparrow$
Candidate-level	$37.29\pm 15.10$	$24.90\pm 8.55$	$85.13\pm 4.87$
Random edits	$32.94\pm 13.28$	$24.35\pm 9.10$	$95.75\pm 1.62$
History-guided	$40.15\pm 10.30$	$28.28\pm 9.54$	$\mathbf{97.88\pm 1.69}$
TRACE	$\mathbf{50.19\pm 14.45}$	$\mathbf{38.51\pm 8.69}$	$96.63\pm 1.44$

Table 2: Component ablation of TRACE on Solid-State Electrolytes. Values are mean

\pm

sample standard deviation over five runs.

Figure 4: Sensitivity to the LLM:Delta budget ratio on Solid-State Electrolytes. Lines report mean H.R.; bars report generated structures and unique valid formulas.

Task	Runs	$N_{\mathrm{edit}}$	$\Delta$ -Dir. Acc. (%)	Selection Adv. (%)	Valid Yield (%)	Valid Formula Uniq. (%)
Wide-Bandgap Semiconductors	5	218	76.49	59.66	62.75	81.75
High-k Dielectrics	5	216	61.12	64.16	9.62	100.00
Solid-State Electrolytes	5	253	69.09	54.31	67.06	89.47
Perovskite Oxides	5	195	70.99	57.51	38.63	85.71

Table 3: Transition-level analysis of 882 post-warm-up transitions across four tasks and five runs per task, reporting

\Delta

-direction accuracy, offline-replay selection advantage, and pooled formula uniqueness.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

	G.R. $\uparrow$		Conditional H.R. $\uparrow$		Scheduled yield $\uparrow$
Task	LLEMA	TRACE	LLEMA	TRACE	LLEMA	TRACE
Wide-Bandgap Semiconductors	80.5	98.2	22.6	45.7	18.2	44.9
High-k Dielectrics	78.2	99.4	7.8	29.4	6.1	29.2
Hard, Stiff Ceramics	82.2	95.1	11.8	46.2	9.7	43.9
Perovskite Oxides	77.3	97.0	8.4	21.9	6.5	21.2

Appendix

Table 4: Search-efficiency decomposition after 1,000 scheduled attempts. Values are percentages summarized over five long-horizon runs per task–method pair; metrics follow Appendix B.2 .

Scheduled attempts	Wide-Bandgap Semiconductors		High-k Dielectrics		Hard, Stiff Ceramics		Perovskite Oxides
Scheduled attempts	Dir. acc.	Valid rate	Dir. acc.	Valid rate	Dir. acc.	Valid rate	Dir. acc.	Valid rate
250	70.78	48.05	69.77	17.44	65.79	77.19	74.67	24.00
500	77.46	40.88	64.80	27.93	68.67	79.33	77.59	38.51
750	74.60	55.60	63.67	43.88	71.08	84.34	77.18	42.46
1,000	73.17	65.53	64.47	48.73	73.20	85.33	78.61	48.84

Appendix

Table 5: Cumulative edit diagnostics at the scheduled-attempt checkpoints used in the main search-efficiency analysis. Values are percentages; direction accuracy is computed over nonzero property changes.

Figure 5: Task-level conditions for realizing gains from transition guidance. Each point summarizes five short runs for one task. a , Offline selection advantage versus matched H.R. difference. b , Valid-child rate versus matched H.R. difference.

	All evaluated		All valid
Task	LLEMA	TRACE	LLEMA	TRACE	Pooled overlap
Wide-Bandgap Semiconductors	$18.74\!\pm\!6.68$	$1.74\!\pm\!1.13$	$24.06\!\pm\!11.31$	$1.11\!\pm\!1.11$	18.25
High-k Dielectrics	$21.21\!\pm\!4.40$	$4.07\!\pm\!2.09$	$10.30\!\pm\!9.43$	$0.00\!\pm\!0.00$	0.00
Solid-State Electrolytes	$10.68\!\pm\!3.49$	$5.29\!\pm\!2.24$	$17.14\!\pm\!6.36$	$8.00\!\pm\!3.70$	10.53
Perovskite Oxides	$24.06\!\pm\!4.22$	$5.66\!\pm\!3.49$	$20.93\!\pm\!9.59$	$3.41\!\pm\!3.37$	14.29

Appendix

Table 6: Exact-formula repetition in matched short-run LLEMA and TRACE archives. Values are percentages; within-run entries report mean

\pm

sample standard deviation over five runs, and the final column reports pooled cross-run overlap of valid TRACE edit offspring.

Figure 6: Elemental distributions produced by LLEMA and TRACE at the common 500-iteration checkpoint. Colors show within-panel relative occurrence; $N$ denotes evaluated candidates and $K$ denotes distinct observed elements.

Large Language Model-Guided Scientific Discovery

Large language model-guided scientific discovery uses AI to propose, test, and refine hypotheses, models, experiments, and designs.

See all 278 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
An LLM agent for end-to-end computational materials discovery
Aug 20, 2026Chen Yuntong, Huang Lijiang, Li Yingqi +7Materials Discovery AI Agents for Scientific Discovery
School of Mechanical Engineering, Northwestern Polytechnical University, Xi’an, 710072, Shaanxi, China. · Institute of Information Engineering, Chinese Academy of Sciences, Beijing, 100045, China. · Department of Radiology, Huazhong University of Science and Technology, Wuhan, 430022, Hubei, China. +2
ReTRACE: Feedback-Guided LLM Agent in a Chemistry Harness for Retrosynthesis Planning
Oct 17, 2026Jinxiang Xie, Yijian Zhang, Qingyuan Yi +2LLM Agents Chemical Reasoning in LLMs
State Key Laboratory of Analytical Chemistry for Life Science, Kuang Yaming Honors School, Chemistry and Biomedicine Innovation Center (ChemBIC), ChemBioMed Interdisciplinary Research Center at Nanjing University, Nanjing University, Nanjing 210023, China · School of Medicine, Guangxi University, Nanning, 530004, China

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 Scientific Discovery

Explore similar work

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

An LLM agent for end-to-end computational materials discovery

ReTRACE: Feedback-Guided LLM Agent in a Chemistry Harness for Retrosynthesis Planning