acceptodds
Under review as a conference paper at ICLR 2027

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

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%.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
LLM agents, materials discovery, multi-objective optimization, sample-efficient search, transition-level feedback

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_pQLKOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TRACE, a framework for LLM-agent-based multi-objective materials discovery that records evaluated parent–edit–child transitions and uses their observed property changes to guide subsequent local edits. Edits are ranked by their predicted ability to reduce the current candidate’s remaining constraint violations without damaging already satisfied objectives, while global LLM proposals maintain broader exploration. On 14 LLEMABench tasks with two quantized open-source backbones, TRACE improves hit rate over LLEMA on most tasks under approximately matched evaluation counts, supported by ablation, budget-allocation, and transition-level analyses.

Critical strengths

S1. Treating evaluated parent–edit–child transitions as feedback is intuitive and well motivated for multi-objective materials search.

S2. Under the same backbone, tasks, and oracle, with approximately matched property-evaluation counts, TRACE improves hit rate over LLEMA on most tasks.

S3. The empirical analysis covers multiple aspects of the method, including component ablations, exploration-budget allocation, transition-level diagnostics, and sensitivity to noisy property feedback.

S4. The figures are clear and informative. Figures 1 and 2 effectively illustrate the motivation and framework, while Figure 3 clearly presents discovery efficiency under a fixed attempt budget.

Critical weaknesses

W1. The controlled LLEMA–TRACE comparisons use only two 8-bit quantized backbones. Table 7 reports TRACE results with GPT-5.4 but no corresponding LLEMA results, leaving it unclear whether the relative gains persist with stronger proposers.

W2. Local edits preserve the parent lattice and coordinates, but it is unclear whether child structures are relaxed before evaluation. The paper also does not report how many hits come from reference data versus surrogate predictions, making their physical reliability difficult to assess.

W3. Although TRACE improves average performance under Mistral, some tasks show substantial regressions, such as Wide-Bandgap Semiconductors. Table 1 reports no standard deviations despite five runs, making the consistency of these differences difficult to assess.

m1. The ablation and budget-allocation studies are limited to Solid-State Electrolytes with a single backbone.

m2. The manuscript does not report the numerical weights in Eqs. 8, 12, and 14 or how they were selected.

Questions for authors

Q1. Could you report LLEMA results with GPT-5.4 or another stronger proposer, to show whether the relative gains of TRACE persist? (see W1)

Q2. Are child structures relaxed before property evaluation? What fraction of hits comes from reference data versus surrogate predictions? (see W2)

Q3. Could you report standard deviations for Table 1, and comment on why TRACE underperforms LLEMA on some tasks under Mistral? (see W3)

Q4. What values are used for the weights in Eqs. 8, 12, and 14, and how were they selected? (see m2)

5 comments

Response to Q1/W1:

We thank the reviewer for requesting a matched comparison with a stronger proposer. To provide the requested controlled comparison, we evaluated both LLEMA and TRACE with GPT-5.6 on four matched tasks under the same task definitions, oracle pipeline, and budget settings used in the main comparison. Each entry reports one controlled run.

TaskLLEMA H.R. (%)TRACE H.R. (%)LLEMA Stab. (%)TRACE Stab. (%)
Wide-Bandgap Semiconductors85.7195.6248.2180.62
SAW/BAW Acoustic Substrates86.1685.445.0320.25
Solid-State Electrolytes81.4388.0512.8638.99
Acousto-optic Hybrids16.2227.924.053.25

TRACE improves both H.R. and Stab. on Wide-Bandgap Semiconductors and Solid-State Electrolytes. On SAW/BAW Acoustic Substrates, H.R. remains nearly unchanged while Stab. increases from 5.03% to 20.25%; on Acousto-optic Hybrids, H.R. increases from 16.22% to 27.92%, while Stab. changes from 4.05% to 3.25%. This task-level comparison provides evidence that TRACE's relative gains are not confined to the two quantized backbones, while the affected metric and gain magnitude remain task dependent. The comparison is included in Appendix D.7 of the revised manuscript.

Response to Q2/W2:

We thank the reviewer for requesting this clarification. Following the common LLEMABench evaluation protocol, TRACE does not introduce an additional structure-relaxation stage after a local substitution. The edited CIF is evaluated through the same oracle pipeline used for LLEMA: properties are retrieved from the Materials Project when available, and unavailable properties are completed using the task-specific surrogate models. This shared protocol provides a controlled comparison of the two search methods without introducing method-specific evaluation differences and is stated explicitly in Appendix B.1 of the revised manuscript.

To make the source composition explicit, we audited valid hits across four representative tasks, two backbones, and five runs per setting. The table below reports the pooled percentage of valid hits across the five runs in two mutually exclusive categories. “Reference-assisted” means that the Materials Project supplied at least one constrained property and therefore includes mixed cases in which remaining properties were completed by surrogates. “Surrogate-only” means that no constrained property was obtained from the reference API.

BackboneTaskLLEMA Reference-assisted (%)LLEMA Surrogate-only (%)TRACE Reference-assisted (%)TRACE Surrogate-only (%)
QwenWide-Bandgap Semiconductors89.0110.9949.2750.73
QwenSolid-State Electrolytes75.0025.0065.6634.34
QwenHigh-k Dielectrics44.9055.1028.2171.79
QwenPiezo Energy Harvesters13.8086.2011.9488.06
MistralWide-Bandgap Semiconductors93.476.5392.837.17
MistralSolid-State Electrolytes85.2514.7570.6829.32
MistralHigh-k Dielectrics45.5754.4336.9663.04
MistralPiezo Energy Harvesters12.9987.0110.6189.39

TRACE has a lower reference-assisted fraction in all eight matched comparisons. Thus, in settings where TRACE improves, the gain cannot be explained by retrieving a larger fraction of known database materials. Our claims concern search performance under the shared LLEMABench oracle rather than independent experimental validation of the generated structures.

Response to Q3/W3:

We thank the reviewer for requesting task-level uncertainty estimates and a more explicit analysis of the task-dependent behavior under Mistral. We report the sample standard deviation over five independent runs for every task–backbone–method setting in Appendix D.1. The matched Mistral H.R. and Stab. results are reproduced below so that the corresponding gains and regressions can be assessed directly.

TaskLLEMA H.R. (%)TRACE H.R. (%)LLEMA Stab. (%)TRACE Stab. (%)
Wide-Bandgap Semiconductors51.98 ± 8.4336.59 ± 5.6843.67 ± 7.5829.50 ± 3.71
SAW/BAW Acoustic Substrates27.54 ± 6.7743.21 ± 6.626.65 ± 2.7014.70 ± 9.31
High-k Dielectrics11.09 ± 2.6412.38 ± 4.145.28 ± 1.804.85 ± 1.70
Solid-State Electrolytes31.96 ± 6.2160.06 ± 16.9827.35 ± 3.9148.15 ± 11.33
Piezo Energy Harvesters69.98 ± 4.3466.23 ± 7.637.77 ± 3.0015.53 ± 2.53
Transparent Conductors21.62 ± 3.3127.65 ± 6.9910.01 ± 1.8911.26 ± 3.68
Insulating Dielectrics15.32 ± 2.6316.62 ± 5.248.75 ± 3.329.59 ± 4.65
Photovoltaic Absorbers21.78 ± 4.1122.96 ± 4.5413.52 ± 3.408.96 ± 1.80
Hard Coating Materials9.91 ± 4.797.28 ± 3.282.92 ± 1.442.17 ± 1.25
Hard, Stiff Ceramics52.32 ± 14.5357.42 ± 13.1519.87 ± 5.6911.59 ± 3.14
Aerospace Materials9.05 ± 2.445.92 ± 1.642.11 ± 0.971.13 ± 1.29
Acousto-optic Hybrids11.67 ± 2.7914.28 ± 3.414.36 ± 1.798.32 ± 6.61
Low-Density Structures16.63 ± 4.8419.07 ± 7.372.14 ± 1.593.30 ± 1.14
Perovskite Oxides4.60 ± 1.2812.70 ± 4.753.68 ± 0.982.45 ± 1.55

Under Mistral, TRACE increases the macro-average H.R. from 25.39% to 28.74% and has a higher mean H.R. on 10 of the 14 tasks. The run-level analysis therefore supports a majority-task and macro-average improvement while showing that the effect magnitude and variability remain task dependent.

To address the second part of the question, we use the route-level audit in Appendix D.4 to examine two representative regressions: the largest H.R. decrease on Wide-Bandgap Semiconductors and a distinct route-balance pattern on Piezo Energy Harvesters. On Wide-Bandgap Semiconductors, H.R. decreases from 51.98 ± 8.43% for LLEMA to 36.59 ± 5.68% for TRACE. The regression cannot be attributed to TRACE evaluating fewer candidates: the corresponding means are 156.0 ± 5.1 and 158.0 ± 1.0. The route-level audit provides a concrete diagnostic: under Mistral, post-initialization valid-candidate yield is 55.20% for LLEMA, compared with 41.52% for TRACE's global route and 43.01% for its local route. The direction reverses with Qwen, where the corresponding yields are 25.14%, 34.24%, and 60.29%, and H.R. increases from 25.23% to 44.78%. This matched reversal indicates that the regression is specific to the backbone–task combination rather than a general failure on Wide-Bandgap Semiconductors.

Piezo Energy Harvesters illustrates a different route-balance effect. Under Mistral, TRACE's local route has a higher valid-candidate yield than LLEMA (79.25% versus 69.65%), whereas its global route is weaker (57.67%). This route imbalance is consistent with a slightly lower aggregate H.R. (66.23 ± 7.63% versus 69.98 ± 4.34%), while Stab. increases from 7.77 ± 3.00% to 15.53 ± 2.53%. Together, these results indicate that TRACE's gains depend jointly on the backbone-conditioned proposal distribution and the effectiveness and balance of its global and local routes. We therefore characterize the Mistral result as a majority-task and macro-average improvement with identifiable task-specific regressions.

Reviewer_h9bxOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TRACE, a framework for materials discovery under multiple constraints and limited property-evaluation budgets. TRACE records evaluated local modifications as parent–edit–child transitions with observed property deltas. It estimates reusable edit effects by aggregating historical deltas, with a hierarchical backoff from exact-edit statistics to operator-level or global estimates when observations are sparse. For a given parent material, these estimates are used to rank edits according to their predicted effect on the remaining constraint violations. A separate near-feasible score assigns different weights to repairing violated constraints and damaging satisfied ones, and rewards predicted feasibility. Local editing is combined with global LLM proposals, which also receive transition summaries. The implemented local operator is restricted to same-group elemental substitutions that preserve the lattice and fractional coordinates.

The main evaluation covers 14 LLEMABench tasks and two LLM backbones, with five runs per task–backbone setting. Reported macro-average hit rates increase from 18.13% to 25.96% with Qwen and from 25.39% to 28.74% with Mistral. Additional experiments include component ablations, global–local allocation studies, long-horizon search, and transition-level diagnostics.

Critical strengths
  1. A meaningful problem and a coherent, lightweight approach. Reusing expensive evaluation outcomes to guide subsequent modifications is an important problem in materials discovery. TRACE provides a concrete mechanism for retaining the effects of executed edits and matching them to the current candidate's remaining requirements. Transition memory, empirical effect estimation, and constraint-aware ranking form a well-motivated search loop. The estimator is simple and does not require training a separate neural dynamics model, which is attractive when evaluation data are limited.

  2. Broad empirical coverage and useful analyses of the mechanism. The main experiments cover fourteen tasks, two backbones, and five runs per setting, with hit-rate improvements over LLEMA on the majority of tasks under both backbones. The authors also study component contributions, global–local allocation, and transition prediction and selection quality. Appendix B.2 distinguishes evaluation-conditional success from scheduled-attempt yield; Appendix D.1 decomposes the long-horizon gains into evaluation completion and conditional candidate quality; and Appendices D.2–D.4 examine accumulated evidence, task dependence, and sensitivity to noisy feedback. These analyses provide a useful foundation for understanding the approach.

Critical weaknesses
  1. The specific methodological contribution and the independent benefit of the near-feasible score need stronger support. LLMatDesign already incorporates material modifications and reflections on their outcomes into subsequent decisions, while MARS learns guidance for molecular fragment-edit selection from online search experience. Neither implements the particular estimator and ranking mechanism proposed here, but these precedents suggest that the contrast with candidate-level feedback should be qualified. EoH and ReEvo provide broader context for LLM-guided evolutionary search; because they evolve heuristic programs rather than material candidates, they are not direct substitutes for TRACE. The clearest contribution of TRACE appears to be the combination of explicit numerical property-delta estimation with constraint-dependent edit ranking. This distinction should be articulated more precisely.

    From Eqs. (11) and (13),

    g^D(x,e)=Gmiss(x,e)−Bsat(x,e).\widehat g_D(x,e)=G_{\mathrm{miss}}(x,e)-B_{\mathrm{sat}}(x,e).

    Thus, total predicted residual improvement already accounts for deterioration in previously satisfied constraints, although gains on other constraints may compensate for that deterioration. The near-feasible score changes this trade-off through its weights and explicit feasibility bonus; historical utility and novelty also contribute to the overall ranking. This is a reasonable design, but the incremental benefit of these scoring choices is not directly isolated. Table 2 provides encouraging evidence for the full method, yet its progressive ablation on one task and one backbone does not separately identify the contribution of the near-feasible score relative to total-residual ranking. A targeted comparison using the same edit-effect estimates, parent-selection procedure, admissible edits, and evaluation budget would be more informative. The authors should also specify precisely what changes between the history-guided variant and full TRACE, including whether the former already serves as a historical mean-utility baseline. Extending the most informative comparison to another task with different behavior would strengthen the attribution of the gains.

  2. The empirical evidence needs more complete statistical reporting and a clearer characterization of the transition analysis and its scope. Table 1 reports averages over five runs but does not include uncertainty for the main task-level comparisons. The standard deviations in Table 2 are useful, but also indicate appreciable run-to-run variation. Together with the task-dependent improvements, this motivates reporting individual-run results or uncertainty estimates for the principal LLEMA–TRACE comparisons. This does not require pooling heterogeneous tasks for a significance test; the relevant issue is the consistency and magnitude of the differences within each setting.

    The budget distinctions in Appendix B.2 and the efficiency decomposition in Appendix D.1 are appropriate. The approximately matched property-evaluation counts in the main comparison support an oracle-limited assessment, while the long-horizon experiment measures productivity under a fixed scheduled-attempt budget. Actual evaluation-count distributions for the short-run comparisons would make the degree of matching easier to assess. These findings should retain their respective interpretations: fixed-attempt gains do not by themselves establish equal-cost gains per oracle evaluation or equal overall computation. LLM inference usage or runtime would be useful if broader computational-efficiency claims are intended, but matching such costs is not a prerequisite for the stated oracle-limited comparison. The non-LLM baselines should likewise remain contextual reference points, as the authors acknowledge.

    Table 3 and Appendix D.2 provide initial evidence that historical edit effects contain predictive information, including improved direction accuracy with greater exact-edit support. However, interpreting the selection advantage requires a more explicit account of how alternatives are chosen, whether they apply to the same parent, how their outcomes are obtained, what evidence is available when predictions are made, and how ties are handled. Since the logged transitions depend on the search policy, the authors should clarify what the offline comparison can establish about selection over the admissible edit set. Sufficient configurations and analysis records should be supplied to connect the implementation to the reported results, without requiring readers to reconstruct the protocol from code. Finally, the estimator pools effects across parent materials without explicitly conditioning on their structural environments, although the ranking itself uses the current parent's properties. Appendix D.3 and the limitations section acknowledge the resulting task dependence and potential contextual variation. The noise analysis is also useful, but does not directly establish transfer across structurally different parents. A more direct characterization by compositional or structural similarity would help identify when pooling is reliable. A simple parent-conditioned comparison could be informative, but is not necessary to establish the usefulness of a deliberately lightweight estimator.

Questions for authors
  1. What is the incremental contribution of the near-feasible score? Please clarify the exact differences between the history-guided variant and full TRACE. Can the authors compare Eq. (14) with ranking by Eq. (11), keeping the edit-effect estimator, parent selection, admissible edits, and budget fixed? Does the history-guided variant already provide the corresponding historical mean-utility control, and does the principal comparison hold on an additional task?

  2. How robust are the main comparisons under the stated budgets? Please provide individual-run results or task-level uncertainty estimates, together with actual property-evaluation counts for the approximately matched settings. Building on the existing discussion of task dependence, which conditions appear to explain improvements or regressions relative to LLEMA? If broader computational-efficiency claims are intended, please also report LLM inference usage or runtime; otherwise, keep those claims specific to the evaluation or attempt budgets tested.

  3. What does the transition-level analysis establish about actionable and transferable edit effects? Please specify the selection-advantage protocol, including the alternatives considered, their parent materials and outcomes, the evidence available at each ranking decision, and the handling of ties. Beyond the existing support-count and noise analyses, can the authors characterize when an edit's observed effects remain reliable across different parent-material environments?

14 comments

Reviewer comment (Q1 — Part 1/3):

LLMatDesign already incorporates material modifications and reflections on their outcomes into subsequent decisions, while MARS learns guidance for molecular fragment-edit selection from online search experience. Neither implements the particular estimator and ranking mechanism proposed here, but these precedents suggest that the contrast with candidate-level feedback should be qualified. [...] The clearest contribution of TRACE appears to be the combination of explicit numerical property-delta estimation with constraint-dependent edit ranking. This distinction should be articulated more precisely.

Response:

We thank the reviewer for drawing attention to these relevant precedents and for providing a useful characterization of TRACE's methodological distinction. LLMatDesign incorporates reflections on prior material modifications into subsequent decisions, MARS learns an adaptive fragment-editing proposal from samples generated during online MCMC search, and EoH and ReEvo provide broader context for experience-guided evolutionary search. Together, these works establish important precedents for reusing search experience. Within this context, TRACE organizes evaluated parent–edit–child transitions as structured edit-level evidence, estimates reusable numerical property changes, and conditions executable edit selection on the current parent's constraint residuals. We have revised the Related Work, marked in red, to position these precedents explicitly and clarify that TRACE's methodological contribution lies in integrating edit-effect estimation with residual- and state-dependent control.

Reviewer question (Q1 — Part 2/3):

What is the incremental contribution of the near-feasible score?

Response:

Eq. (11) is a derived intermediate signal: it maps the historically estimated edit effect onto the current parent's residual landscape. As the reviewer notes, Eq. (11) already accounts for both repair of violated constraints and potential damage to satisfied constraints. However, because it is derived from edit effects estimated from finite transition evidence collected across parents, it is informative but incomplete and is not used as a standalone decision rule. Eq. (14) neither introduces a new edit-effect estimate nor first introduces damage awareness. Rather, it is the near-feasible instantiation of the integrated state–residual controller: it reorganizes the same residual information through asymmetric repair–damage weighting and an explicit feasibility-crossing bonus, while retaining historical utility and novelty. Its role is therefore to adapt the decision rule to the parent's state, not to replace Eq. (11) with a fundamentally different predictor. We revised the Method section, marked in red, to make this relationship explicit.

Reviewer question (Q1 — Part 3/3):

Please clarify the exact differences between the history-guided variant and full TRACE. Can the authors compare Eq. (14) with ranking by Eq. (11), keeping the edit-effect estimator, parent selection, admissible edits, and budget fixed? Does the history-guided variant already provide the corresponding historical mean-utility control, and does the principal comparison hold on an additional task?

Response:

The proposed fixed-condition comparison isolates only the near-feasible scoring branch. Because TRACE's claim concerns the integrated state–residual controller, we treat this comparison as a local diagnostic rather than the primary attribution experiment. To make the ablation progression explicit, we use the History-guided variant as the historical mean-utility control: it ranks edits by accumulated historical utility while selecting editable parents without state-aware prioritization. We then add a History + residual variant that is identical except for incorporating the Eq. (11) predicted residual-progress signal; neither variant uses state-aware parent scheduling or the Eq. (14) near-feasible branch. TRACE further adds parent-state assessment, state-aware parent prioritization, and state-specific reranking, including the Eq. (14) branch for near-feasible parents. Accordingly, History-guided versus History + residual isolates the value of residual guidance, whereas History + residual versus TRACE evaluates the integrated state–residual module.

All entries below report mean ± sample standard deviation over five runs with Qwen2.5-14B-Instruct under the same fixed scheduled budget and evaluation protocol.

Solid-State Electrolytes

VariantH.R. (%)Stab. (%)G.R. (%)
Candidate-level37.29 ± 15.1024.90 ± 8.5585.13 ± 4.87
Random edits32.94 ± 13.2824.35 ± 9.1095.75 ± 1.62
History-guided40.15 ± 10.3028.28 ± 9.5497.88 ± 1.69
History + residual38.79 ± 7.4430.57 ± 7.1896.50 ± 1.57
TRACE50.19 ± 14.4538.51 ± 8.6996.63 ± 1.44

Wide-Bandgap Semiconductors

VariantH.R. (%)Stab. (%)G.R. (%)
Candidate-level31.44 ± 4.2923.74 ± 2.7386.38 ± 2.04
Random edits25.16 ± 5.4716.28 ± 4.3896.75 ± 2.48
History-guided31.55 ± 10.1220.64 ± 4.3797.88 ± 1.57
History + residual31.71 ± 2.9922.07 ± 3.8996.88 ± 1.71
TRACE44.78 ± 17.6030.94 ± 13.9896.25 ± 1.88

Transparent Conductors

VariantH.R. (%)Stab. (%)G.R. (%)
Candidate-level19.59 ± 3.933.56 ± 1.3191.38 ± 1.12
Random edits37.59 ± 4.769.78 ± 3.7897.00 ± 1.20
History-guided38.46 ± 7.718.67 ± 2.9398.12 ± 1.17
History + residual37.07 ± 5.677.98 ± 2.1697.12 ± 1.14
TRACE39.62 ± 8.248.90 ± 2.1998.38 ± 0.56

Adding residual guidance to the History-guided variant does not produce a uniform gain: H.R. changes by −1.36-1.36, +0.16+0.16, and −1.39-1.39 percentage points on Solid-State Electrolytes, Wide-Bandgap Semiconductors, and Transparent Conductors, respectively. This mixed result is consistent with Eq. (11) serving as an intermediate signal rather than a sufficient standalone controller. TRACE, by contrast, attains the highest H.R. on all three tasks. Relative to History + residual, it improves H.R. by 11.40, 13.07, and 2.55 points and Stability by 7.94, 8.87, and 0.92 points, respectively, while G.R. remains within 1.3 points. On Transparent Conductors, Random edits has the highest mean Stability, but the difference from TRACE is small relative to the run-to-run variation. These results support the integrated state–residual controller as the appropriate contribution unit rather than an isolated effect of Eq. (14). The Wide-Bandgap Semiconductors and Transparent Conductors ablations are reported in Appendix D.1, while the representative Solid-State Electrolytes result remains in the main component-ablation table (Table 2).

Reviewer_Pd97OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses the evaluation on multi-objective materials discovery with LLM agents. This paper proposes TRACE, a transition-aware residual control framework that treats evaluated edits as the basic unit of feedback. Experimental results show that TRACE improves over LLEMA, the state-of-the-art LLM-agent baseline.

Critical strengths
  1. The paper asks a useful question: remember what each edit did, not just which candidates scored well?
  2. The residual-aware scoring is well motivated.
  3. The experiments is well constructed.
Critical weaknesses
  1. The edit space is limited to structure-preserving elemental substitutions. The current local operator mainly performs same-group elemental substitutions while keeping the lattice, fractional coordinates, and number of sites fixed. This is reasonable for clean credit assignment, but it limits what the method can discover.
  2. Edit effects depend on the parent material, but TRACE averages transitions across parents.
Questions for authors

Please see the weaknesses.

3 comments

Reviewer comment:

The edit space is limited to structure-preserving elemental substitutions. The current local operator mainly performs same-group elemental substitutions while keeping the lattice, fractional coordinates, and number of sites fixed. This is reasonable for clean credit assignment, but it limits what the method can discover.

Response:

We thank the reviewer for highlighting this scope boundary. The current local operator deliberately uses structure-preserving, same-group elemental substitutions, which provide directly executable and consistently identifiable edits for learning transition effects. However, this restriction applies only to the local refinement route, not to TRACE's overall search space. TRACE is explicitly a hybrid search framework: global LLM proposals explore new compositions, material families, and structural candidates beyond the substitution neighborhood, while transition-guided local edits exploit accumulated evaluation evidence around promising parents. Evaluated global proposals can subsequently become local-edit parents, and transition summaries are returned to the global route, allowing the two routes to exchange information rather than operate independently.

The experiments support this complementary design. In the global–local allocation study on Solid-State Electrolytes, LLM-only and local-edit-only search achieve H.R. values of 32.4% and 27.7%, respectively, whereas their balanced combination reaches 50.19% under the same scheduled budget (Figure 4). Moreover, within the same fixed substitution space, H.R. increases from 32.94% with random edits to 40.15% with history-guided selection and 50.19% with full TRACE (Table 2). Thus, the improvement comes from combining broad global exploration with more effective transition-aware local control, rather than from expanding the local operator set.

Reviewer comment:

Edit effects depend on the parent material, but TRACE averages transitions across parents.

Response:

We clarify an important distinction: although the realized effect of an edit necessarily depends on the parent material, TRACE does not infer edit behavior from a single execution on one parent structure. It instead aggregates repeated observations of the same edit across different parents to estimate a reusable tendency in property changes. This pooled estimate is not treated as parent-invariant; it is applied to the current parent's observed properties and evaluated against its specific satisfied and violated constraints. Thus, transition evidence is aggregated across parents, while the resulting edit decision remains conditioned on the current parent state (Section 3, Eqs. (9)–(11)).

To characterize the scope of this reuse, we added a chronological parent-context analysis in Appendix D.5 of the revised manuscript.

Prior exact-edit supportCross-context (<0.75) NDir. acc.Near-context (≥0.75) NDir. acc.
All (≥2)61/11956.30%472/93678.10%
2–342/8159.26%183/36173.68%
4–514/2846.43%109/21773.27%
≥65/1060.00%180/35885.47%

Here, NN denotes decisions/property comparisons. The analysis covers 533 decisions and 1,055 property comparisons, with no current parent formula exactly repeated in its historical evidence set. Direction agreement increases from 56.30% in cross-context transfers to 78.10% in near contexts, and the separation remains within the comparable 2–3 and 4–5 support strata. Together, these results indicate that TRACE captures reusable cross-parent edit regularities rather than memorizing identical parent formulas, while the context-dependent agreement favors conditional rather than parent-invariant reuse.

Response:

Thank you for your thoughtful engagement with our work. We particularly appreciate your confirmation that our rebuttal addressed your concerns and your continued positive assessment of the manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Contributions and Rebuttal

We thank the Area Chair and all reviewers for their constructive feedback. TRACE addresses multi-objective materials discovery by turning costly evaluations into reusable guidance for subsequent search. It learns how executable edits change material properties and combines these estimates with each candidate’s remaining constraint violations to select promising refinements. Our analyses show that this guidance becomes more reliable as relevant experience accumulates and is more consistent across compositionally related materials, helping explain the larger search gains observed over longer trajectories.

During the discussion, we clarified TRACE’s methodological contribution, strengthened the evidence for its integrated controller, and expanded the analyses of evaluation fairness and the transferability of edit effects. We have addressed the reviewers’ concerns through these revisions and additional analyses, and the work currently has positive assessments from all three reviewers.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.