PaperlayerView exact revision on arXiv

Source-linked AI summary

Agentic-MatriBO: Capability-Compiled Agentic Bayesian Optimization for Materials and Chemical Design

Yu Zeng, Yuchuan Chen, Ziyue Zhang, Zijian Wei, Yanyi Su, Guojiang Zhao, Hongshuai Wang, Zhifeng Gao

arXiv:2610.17832v3cs.LGcs.AI

TL;DR

Materials and chemical optimization must allocate limited evaluations, while combining BO components requires compatibility, evidence, and execution contracts. Agentic-MatriBO compiles capabilities for policy-proposed recipes and candidate plans, then checks and executes them. Across four offline finite candidate pools, it achieves the best mean endpoint quality in each task’s objective direction.

  • Problem

    Limited-budget materials and chemical optimization requires sequential choices, but composing BO components entails compatibility, evidence, and execution requirements.

    Source: §1 · pp. 1–2

  • Method

    Agentic-MatriBO registers BO components and selection conditions as capabilities, lets a language policy propose recipes and candidate plans, and uses a compiler and deterministic execution layer to check and enact proposals.

    Source: §1 · p. 2

  • Results

    Across four offline finite candidate pools and five seeds per condition, Agentic-MatriBO has the best mean endpoint quality in each task’s objective direction, while cumulative-quality gains depend on the task.

    Source: §8 · p. 18

  • Takeaways & Limitations

    The framework makes capability selection conditions, composition admissibility, and execution semantics explicit for evidence-conditioned BO adaptation.

    Source: §1 · p. 2

    Evaluation is limited to four offline finite candidate pools, and offline search cannot validate materials mechanisms or replace prospective wet-lab experiments.

    Source: §7 · p. 17

Abstract

from arXiv · show

Materials and chemical optimization must choose the next synthesis, characterization, or high-fidelity simulation from sequential observations under a limited budget. Large language models can use task knowledge to propose adjustments, but composing surrogates, acquisition functions, and search strategies is constrained by compatibility, parameter dependencies, and evidence requirements. We present Agentic-MatriBO, a capability-compiled agentic Bayesian optimization (BO) framework that represents BO components and selection conditions as composable capabilities. A language policy proposes BO recipes and candidate plans; a compiler checks proposals against the registered capability catalog; and a deterministic execution layer validates and commits candidates. Oracle observations are stored in research memory and inform later rounds. We compare Agentic-MatriBO with eight baselines on four offline finite candidate pools---COF, PCE10, Buchwald--Hartwig (BH), and Suzuki--Miyaura---using five seeds per condition; results are reported descriptively. Across the four tasks, Agentic-MatriBO has the best mean endpoint quality after orienting each objective so that higher is better. On COF, it reaches the tabulated optimum in all five seeds both in the end-to-end comparison and after shared initialization, where it also retains an advantage in cumulative quality. Its normalized area under the best-so-far quality curve during the BO stage (nAUC-BO) is substantially higher on COF, near parity with the best baseline on PCE10 and BH, and lower on Suzuki. COF ablations show that removing the capability compiler or evidence gate coincides with lower cumulative quality and fewer exact-optimum hits. The framework provides an inspectable interface for evidence-conditioned BO adaptation.

1. Introduction

Materials and chemical optimization must allocate limited evaluations across candidate experiments, while composing BO components requires explicit compatibility, evidence, and execution interfaces. Agentic-MatriBO addresses this with capability-compiled, two-level adaptation that separates policy proposals from deterministic candidate commitment.

1. Introduction

  • BO component composition requires compatibility, evidence, and execution interfaces rather than selecting one rule in isolation.LLM proposals also require candidate-validity checks and numerical modeling.Source: §1 · pp. 1–2
  • Agentic-MatriBO organizes recipe configuration and per-round candidate planning as two levels of evidence-conditioned capability compilation.A language policy proposes adjustments, while numerical surrogates and acquisitions rank candidates and deterministic checks govern commitment.Source: §1 · p. 2
  • The framework registers component parameter domains, compatibility relations, and runtime conditions, then checks proposals before numerical search, validation, and commitment.Oracle observations and reflections enter research memory for later rounds.Source: §1 · p. 2
  • The evaluation separates starting-point effects from subsequent search through four-task comparisons, COF shared initialization, and grouped ablations.These comparisons examine how optimization gains vary by task and search stage.Source: §1 · p. 2

2. Problem Definition

The problem definition casts BO as sequential evaluation under a finite budget, with task-specific candidates and an adjustable online recipe. It distinguishes proposals from commitments so that policy adjustments and target evaluations have separate permissions.

2.1 Standard Bayesian Optimization and Fixed Policies

  • Standard BO fits a probabilistic surrogate to completed observations and uses an acquisition function to propose the next candidate under a fixed policy interface.The policy and interface are fixed, but the configuration may vary with data.Source: §2.1 · pp. 2–3

2.2 Materials-Task Contracts and a Variable Optimizer

  • The materials-task contract records candidate identifiers and features, objective, feasibility constraint, evaluation budget, and task metadata.Candidates are selected by identifier, while features support numerical computation; initialization calls count toward the budget, but probes and retrieval do not.Source: §2.2 · p. 3

2.3 Separation of Authority and the Commitment Boundary

  • A proposal supplies an adjustment for checking, whereas a commitment authorizes one target evaluation and an observation records its returned result.The execution layer is deterministic given its proposal, candidate order, scores, runtime state, and numerical environment, though the full campaign includes LLM calls and random sampling.Source: §2.3 · p. 3
  • Table 1 covers roles, permissions, and commitment boundaries.Its caption identifies the scope of the table.Source: Table 1 · p. 3

3. Related Work

Prior work adapts numerical BO components and introduces language information at multiple points in the optimization process. Agentic-MatriBO instead frames component selection and composition as capability compilation with explicit execution semantics.

3.1 Numerical Modeling and Online Adaptation in Materials Optimization

  • Numerical BO adapts acquisition portfolios, local search regions, and surrogate configurations, while materials benchmarks and closed-loop studies examine search and recommendation behavior.BoTorch provides an implementation basis for composing numerical components.Source: §3.1 · p. 4

3.2 Different Entry Points for Language Information into BO

  • Language methods enter BO through task descriptions, candidate guidance, model representations, acquisition strategies, preferences, or low-cost predictions.Recent agentic approaches also reconfigure numerical components through structured interfaces or adapt surrogates and acquisition strategies separately.Source: §3.2 · p. 4
  • Agentic-MatriBO focuses on capability-selection conditions, composition admissibility, and execution semantics, and can incorporate components whose interfaces and contracts are satisfied.The paper positions this as a different design level from prior numerical components and language-guidance mechanisms.Source: §3.2 · p. 4

4. Method: Capability-Compiled Agentic-MatriBO

Agentic-MatriBO separates policy proposals, candidate commitment, and retrospective memory through capability contracts and deterministic execution checks. It coordinates recipe and candidate-plan adaptation, conditions admissibility on task and evidence, and states protocol properties alongside current implementation gaps.

4.1 Method Overview and Execution Flow

  • Policy proposals cannot authorize oracle calls: execution checks compatibility and feasibility, commits a legal candidate, and only then obtains the oracle observation.The loop records observations and evidence judgments for later retrieval, with reviews at proposal, candidate, and post-observation stages.Source: §4.1 · pp. 4–5, Figure 1 · p. 5

4.2 Atomic Capability Space

  • Atomic capabilities expose their parameter domains, compatibility conditions, runtime requirements, evidence prerequisites, and failure signals to policy and execution.The registered space covers surrogate, acquisition, search, and setup-stage initialization capabilities.Source: §4.2 · p. 5

4.3 Evidence-Conditioned Capability Compiler

  • The compiler filters capabilities by task and current state, classifies evidence prerequisites, and checks proposed compositions against the strict admissible recipe space.Support-material binding currently checks only for a nonempty list, not whether cited material exists, is visible, relevant, or sufficiently supportive.Source: §4.3 · p. 6

4.4 Two-Level Compositional Adaptation and Initialization

  • Two-level adaptation proposes both a BO recipe and a candidate plan, while diagnostics and retrieved history inform decisions without adding target observations.Plans specify ranking, shortlist size, source weights, and acquisition quotas; diagnostic cycles are capped per round, and STOP returns a report.Source: §4.4 · pp. 6–7

4.5 Multi-Source Candidates and Deterministic Fusion

  • Candidate plans combine acquisition, max-variance, farthest, and random-coverage sources, then rank legal candidates using numerical utility and, when available, valid review scores.Source weights control source-slot allocation and mixed numerical utility; review-score fusion is separate and advisory.Source: §4.5 · p. 7

4.6 Research Hypotheses, Public Reflection, and Memory Feedback

  • After each oracle result, the system links the proposal, observation, and prior expectation, records a support, refute, or inconclusive judgment, and feeds hypotheses and evidence gaps into later rounds.Hypotheses about domain effects or interactions specify comparison arms, controls, and minimum evidence.Source: §4.6 · p. 8

4.7 Protocol Properties and Conditions for Their Validity

  • The protocol states conditional invariants for declared admission, observation and budget accounting, and equivalence to classical one-step BO under specified conditions.A bounded scoring pool or tolerance-based ranks can otherwise change the selected point.Source: §4.7 · p. 8

4.8 Checking Coverage of the Current Implementation

  • Implementation coverage remains incomplete for supporting-material binding, delayed-action revalidation, and complete ID–feature consistency checks.The current implementation mainly checks newly switched components on the CHANGE path.Source: §4.8 · p. 9

5. Experiments

Across four finite candidate pools, the experiments compare Agentic-MatriBO with eight baselines under five-seed campaigns. Agentic-MatriBO has the best mean endpoint quality on all four tasks, while cumulative-area performance varies by task.

5.1 Tasks, budget, and evidence scope

  • Each condition used five campaigns of 10 initialization and 30 subsequent evaluations on offline tables; end-to-end baselines used their own initialization, with shared initialization also tested on COF.The comparisons included Vanilla BO, HEBO, LABO, CAKE, LMABO, BORA, Reasoning BO, and LLAMBO; grouped COF controls compared starting points, recipes, and execution variants.Source: §5.1 · p. 9, §5.2 · pp. 9–10, §5 · p. 9

5.3 Metrics and statistical interpretation

  • Results are descriptive, without significance, equivalence, or non-inferiority tests; model calls, probes, retries, and fitting overhead are not quantitatively compared.Endpoint and area metrics derive from shared trajectories, so their comparisons are not independent votes.Source: §5.3 · p. 10, §5.4 · p. 10, §5.5 · p. 12

5.5 Cross-task endpoint performance and cumulative search quality

  • Agentic-MatriBO has the best mean endpoint on all four tasks, but cumulative area is not uniformly best; nAUC-BO is lower than BORA on Suzuki.COF and PCE10 show better objective-direction-adjusted means on all four reported metrics; BH and Suzuki do so on three of four. Comparisons are descriptive.Source: §5.5 · pp. 10, 12, Figure 3 · p. 11, Table 3 · p. 11

5.6 COF search performance after shared initialization

  • With the same 10 initial candidates, Agentic-MatriBO retains the highest COF endpoint, BO-stage area, and late-stage area, hitting the tabulated optimum in all five seeds.It also exceeds CAKE on endpoint and both area measures under identical initial conditions, though individual configurations are not isolated.Source: §5.6 · pp. 12–13, Figure 4 · p. 12, Table 4 · p. 13

5.7 Grouped ablations separating starting point, recipe, and review

  • The COF warm-start variants tie on endpoint, while the random start has lower nAUC-BO and later t95; fixed GP+UCB lowers endpoint and area, and No review lowers endpoint and area despite an earlier t95.The reported warm-start values are nAUC-BO 0.970 versus 0.888 and t95 13.4 versus 23.4; No review also reduces exact-optimum hits from 5/5 to 0/5.Source: §5.7 · pp. 13–14, Figure 5 · p. 13, Table 5 · p. 14

5.8 Effects of capability compilation, evidence gating, and candidate commitment

  • Removing the capability compiler or evidence gate reduces cumulative area and exact-optimum hits; LLM commitment preserves endpoint and hit count but lowers nAUC-BO and delays t95.Bare LLM performs substantially worse and records no global-optimum hits across the five seeds.Source: Figure 5 · p. 13, §5.8 · p. 14

5.9 Observed model differences in initialization and early area

  • In five-seed COF runs, the largest observed between-model differences occur during initialization and evaluations 10–20; late-stage area means for the strongest models differ by about 0.0003.The model participates in both initialization and subsequent decisions; the separate ID audit addresses possible prior exposure to public candidate tables.Source: §5.9 · p. 14, Figure 6 · p. 15

5.10 Zero-observation ranking shows no consistent hit advantage

  • Across 16 model–task pairs, original IDs yield higher, lower, or equal hit rates in 6, 6, and 4 pairs; corrected paired tests give q=1.0, showing no consistent hit advantage.The audit does not establish equivalence: it uses ten paired replicates, and both groups retain descriptors.Source: §5.10 · pp. 15–16, Figure 7 · p. 16

6. Discussion

Agentic-MatriBO uses capability conditions to connect task evidence with algorithm composition, while results vary by task, search stage, and configuration. Research memory carries evidence across rounds, but its standalone contribution has not been quantified.

6.1 Capability compilation connects task evidence and algorithm composition

  • Capability compilation exposes component applicability conditions to the strategy and uses them to check proposal admission.In the reported 20 runs, the surrogate switched at least once in every run, while acquisition and search strategy each remained unchanged in nine runs.Source: §6.1 · p. 16

6.2 Gains depend on task, stage, and configuration

  • Endpoint quality measures the best candidate found by budget end, whereas cumulative area also reflects when high-quality candidates first appear.On PCE10 and Suzuki, endpoint improvement accompanies only a very small area gain or a slightly lower cumulative area.Source: §6.2 · p. 17
  • The comparisons cannot separate the contributions of candidate distribution, component adaptation, and domain knowledge.Shared initialization focuses on subsequent search, but warm-start results show that initial candidates can affect cumulative quality.Source: §6.2 · p. 17

6.3 Evidence levels in cross-round feedback

  • Research memory links expectations to observations and stores diagnostics, reflections, and evidence requirements for later retrieval, but its standalone gain remains unquantified.The experiments include no independent no-memory or no-Playbook control.Source: §6.3 · p. 17

7. Limitations

The evaluation is limited to offline finite pools and five seeds per condition, with incomplete coverage across tasks and implementation checks. Its scope does not establish performance in open-ended or prospective settings.

Evaluation scope and statistical analysis

  • Four offline finite candidate pools and five seeds per condition bound the evaluation; COF shared-initialization and mechanism ablations were not repeated on other tasks.Tabulated optima are candidate-set-relative, so results may not generalize to open-ended generation, continuous multi-objective settings, or asynchronous noisy experiments; offline search cannot replace prospective wet-lab experiments.Source: §7 · p. 17

Public data and model priors

  • The zero-observation audit cannot separate general chemical knowledge from memorization of public reaction conditions because both prompt groups retain descriptors.This limits interpretation of the Buchwald–Hartwig and Suzuki tasks.Source: §7 · p. 17

Implementation coverage

  • The evaluated implementation mainly checks newly switched components on the CHANGE path and does not enforce every declared protocol check.It omits support-material verification, deferred-component revalidation after CHANGE-to-KEEP transitions, full checks on heuristic paths, and the declared no-candidate termination behavior.Source: §7 · p. 17

8. Conclusion

Agentic-MatriBO combines capability-based evidence-aware compilation with compositional adaptation for numerical search. Across four offline finite pools it has the best mean endpoint quality in each task’s objective direction, while cumulative-quality gains vary by task.

8. Conclusion

  • Across four offline finite pools, eight baselines, and five seeds per condition, Agentic-MatriBO has the best mean endpoint quality in each task’s objective direction.Cumulative-quality gains depend on the task.Source: §8 · p. 18
  • Agentic-MatriBO registers component-selection conditions as capability contracts and uses evidence-aware compilation and compositional adaptation to connect language strategy with numerical search.Under the declared protocol, the execution layer validates and commits candidates, and oracle observations update research memory for later rounds.Source: §8 · p. 18
  • After shared initialization on COF, all five seeds reached the tabulated optimum while retaining an area advantage, indicating differences in this group also arise from subsequent search.Ablations show that threshold attainment, cumulative quality, and exact hits respond differently to configuration.Source: §8 · p. 18
  • The zero-observation audit found no consistent hit advantage from original candidate IDs, and the evaluated implementation covers the protocol only partially.The implementation limitations are detailed in Section 4.8 and Appendix C.4.Source: §8 · p. 18

Declarations

The declarations report study disclosures and describe the supplemental values underlying the main figures and tables. The appendices specify the scope and interpretation of the benchmark tables, including the shared-initialization COF comparison.

Declarations

  • The authors report no competing interests or external funding, and state that the study involved no human participants, human data, animals, or animal-derived samples.All authors contributed to design, methodology, analysis, and manuscript preparation and approved the final manuscript.Source: Declarations · p. 19
  • The supplement provides values underlying the main-text figures and tables, along with process cases, metric definitions, and implementation details.Benchmark tables and source-processing details are available from the authors upon reasonable request; implementation code and per-seed results are not publicly released at this stage.Source: Declarations · pp. 19–20

A.1 Values underlying Fig. 3 and Table 3

  • Table A1 reports five-seed means underlying Fig. 3 and Table 3 for 10 initialization evaluations and 30 subsequent evaluations.Its final row in each task block reports Agentic-MatriBO’s improvement relative to the best baseline; positive percentages favor the proposed method.Source: Table A1 · p. 20, §A.1 · p. 20

A.2 Values underlying Fig. 4 and Table 4

  • Table A2 reports five-seed means underlying Fig. 4 and Table 4 for COF methods sharing the same 10 initial observations within each seed.The final row of each task block reports improvement relative to the best baseline.Source: §A.2 · pp. 20–21, Table A2 · p. 21

A.3 Ablation and Model Comparisons

The appendix reports grouped COF ablations and model comparisons, with five seeds per condition; supplied passages clarify selected comparisons and implementation details but do not provide full table values.

A.3 Ablation and Model Comparisons

  • The random-initialization warm-start group and component-group full system have identical values for five listed metrics, but are not independent repetitions.The listed metrics are Best-so-far, nAUC-BO, EF40, late-stage nAUC, and t95; the global-optimum hit is reported only for the component group.Source: §A.3 · p. 21
  • Model-sensitivity records use four invocation labels, and aggregate tables omit quantitative LLM cost and call-overhead comparisons.Policy and reviewer calls use temperature 0.0; exact call counts, access dates, and per-call token or cost records were not retained in the aggregate result tables.Source: §A.3 · p. 21
  • The COF comparison tables cover five seeds per condition and report exact-optimum hits and evaluation indices or grouped ablation metrics.Table A4 covers the complete COF model comparison; Table A3 reports grouped ablations, with t95 brackets giving the number of seeds reaching the threshold out of five.Source: Table A3 · p. 21, Table A4 · p. 22

B.1 Metric Definitions

These appendix sections define the reported quality and threshold metrics, document the finite-table sources and comparison procedures, and describe selected search-process and feedback mechanisms. They also delimit what the records can establish, including incomplete reconstruction of cross-round hypothesis updates.

B.1 Metric Definitions

  • COF uses LABO’s frozen 605-row finite candidate table [23] for high-fidelity GCMC equilibrium adsorption selectivity; the source study reports 609 structures, with no established row-by-row mapping.Source: §B.2.3 · p. 23

B.2.1 Normalization and Discrete Integration

  • Normalized quality is oriented by task objective, and nAUC-BO and late-stage nAUC use trapezoidal means over specified evaluation intervals.nAUC-BO averages 30 adjacent segments from n=10 to 40; late-stage nAUC averages 20 segments from n=20 to 40, using full-table extrema for offline normalization.Source: §B.2.1 · p. 22

B.2.2 Threshold Reaching and Exact Hits

  • t95 is the first evaluation reaching Q_n≥0.95; non-reaching seeds receive 41, while exact-optimum hits and threshold reaching are counted separately.The evaluation count includes the 10 initialization evaluations, and the threshold-reaching seed count is reported separately from the mean t95.Source: §B.2.2 · pp. 22–23

B.2.3 Result Sources and Record Scope

  • The reported COF record uses a 605-row frozen table, while PCE10 contains 1,040 records representing 1,020 distinct compositions.COF's source study reports 609 structures, with no established mapping to the 605-row table.Source: §B.2.3 · p. 23

B.2.4 Segmented Area Comparison across Models

  • About 82%, 90%, and 91% of three models’ BO-stage area advantages over DeepSeek 4.0 Flash lie in evaluations 10–20.The models are DeepSeek 4.0 Pro, Qwen3-Max, and GPT-5.6-sol, respectively.Source: §B.2.4 · p. 23

B.2.5 Baseline Sources, Registered Versions, and Adaptation Scope

  • The appendix documents registered baseline implementations and finite-table adaptations; input-feature checks found no duplicate combinations in either reaction table.The BH and Suzuki candidate tables underwent column-name normalization and IUPAC-name mapping.Source: §B.2.5 · p. 23, Table B2 · p. 23, §B.2.6 · p. 24

C.1 Search-Process Cases

  • Search-process cases are examples rather than benchmark or mechanistic evidence, while Figure C1 records strategy and formulation events across 20 runs.The effective surrogate switched at least once in every run and at most 12 times; nine runs each kept acquisition function and search strategy unchanged throughout. The separate Branin run was deterministic, heuristic, and did not invoke an LLM; its best value was found during initialization.Source: §C.1 · pp. 24–25, §C.2 · p. 25, Figure C1 · p. 26

C.3.3 Evaluation Scope of Cross-Round Feedback

  • Reflection uses declared contrast conditions and evidence requirements, but available logs do not support reconstructing the full sequence of hypotheses, observations, revisions, and later actions.Optional proposal review can return ACCEPT or REVISE, with at most one revision; candidate review scores shortlist candidates, while reflection links actual results to public hypotheses.Source: §C.3.1 · p. 25, §C.3.2 · pp. 25, 27, §C.3.3 · p. 27

C.4 Declaration Protocol and Current Implementation

The appendix contrasts declared protocol with implementation details and notes that the workflow figures are functional schematics. Search-path projections show incumbent-update examples, not optimizer inputs.

C.4 Declaration Protocol and Current Implementation

  • The declaration-protocol comparison identifies implementation boundaries, including incomplete adversarial consistency checks and deferred components that bypass re-validation.The heuristic path does not cover all checks applied by the main path; STOP returns a report, while an empty candidate set causes selection failure.Source: Table C2 · p. 27, §C.4 · pp. 27–28

D.1 Figure Notes

  • The workflow figures organize strategy adjustment, candidate execution, observation, and memory feedback; they are mechanism schematics rather than complete accounts of implementation checks.The capability figure depicts evidence-condition compilation, typed proposals, and execution submission; paired E markers indicate computational evidence returned to compiler input.Source: §D.1 · p. 28

D.2 Two-Dimensional Projection Examples of Search Paths

  • Figure D1 marks first and final incumbents and connects updates in temporal order; its backgrounds are post hoc projections, not optimizer inputs.For BH and Suzuki, categorical-label adjacency does not represent continuous physical distance.Source: §D.2 · p. 28, Figure D1 · p. 28
Loading 2610.17832v3…