PaperlayerView exact revision on arXiv

Source-linked AI summary

Reaction precedents extend retrosynthetic prediction beyond training coverage

Jiaxing Sun, Xuli Shen, Jiang Wu, Ruiyi Ding, Ruikang Zhang, Xiao An, Xibao Li, Yubin Wang, Xingjian Wei, Yinfan Wang, Yu Liu, Lingli Ge, Xi Xue, Xiaojian Wang, Lijun Wu, Chang-Yu Hsieh, Yuqiang Li, Bowen Zhou, Dahua Lin, Conghui He

arXiv:2610.40249v1cs.AI

TL;DR

Retrosynthetic prediction depends on training chemistry, but how far inference-time reaction precedents can compensate for coverage gaps remains unclear. The study restricts training and retrieval coverage, then expands retrieval with model parameters fixed. This improved prediction for missing transformations, and the strongest general-purpose model with precedents surpassed the evaluated specialized models.

  • Problem

    How far reaction precedents can mitigate retrosynthetic prediction limits associated with training-coverage gaps remains unclear, despite precedents already improving predictions in prior studies.

    Source: §1 · p. 3

  • Method

    The study restricted temporal or transformation coverage in training and retrieval, expanded only retrieval after training, and evaluated 14 general-purpose models with precedents without task-specific fine-tuning.

    Source: §1 · p. 3

  • Results

    With parameters fixed, retrieval expansion improved prediction for transformations absent from fine-tuning data, and the strongest general-purpose model with DianShi precedents outperformed all eight evaluated specialized models.

    Source: §1 · p. 3, §3 · p. 9

  • Takeaways & Limitations

    Inference-time precedents extended coverage in these experiments and can complement parameter adaptation, whose combination with inference-time evidence achieved the strongest evaluated prediction.

    Source: §3 · p. 10

    Route solve rates measured computational route finding under the stated protocol and did not establish reaction conditions, selectivity, yield, or experimental executability.

    Source: §7.5 · p. 42

Abstract

from arXiv · show

Reaction precedents guide retrosynthesis, but how far they can compensate for gaps in task-specific training coverage remains unclear. Using 2.71 million reactions, we restricted training and retrieval coverage by transformation or time, then expanded only the retrieval corpus with model parameters fixed. For transformations excluded from training, mean Top-1 accuracy increased from 13.0% to 58.0% on a historical test set and from 4.9% to 23.1% on a temporally held-out test set. For temporal coverage gaps, retrieval expansion enabled a model trained on records before 2017 to match or exceed a non-retrieval model trained on records before 2025 on both test sets. Without task-specific fine-tuning, precedents improved all 14 representative general-purpose models on both test sets and USPTO-50K; the strongest with precedents surpassed all eight representative specialized models. These findings show that inference-time reaction precedents can extend retrosynthetic prediction beyond task-specific training coverage without parameter updates.

1 Introduction

This study asks whether reaction precedents supplied at inference can extend retrosynthetic prediction beyond task-specific training coverage. With model parameters fixed, retrieval expansion improved prediction for transformations missing from training, while precedents also benefited general-purpose models without task-specific fine-tuning.

Motivation

  • Training coverage limits retrosynthetic prediction: accuracy is substantially lower for reactions with rare or absent templates, and forward prediction degrades on later-period reactions excluded from training.New reactions continue to be reported, while substantial proprietary reaction records are generally unavailable for public training corpora.Source: Figure 1 · p. 2, §1 · p. 3

Approach

  • The study combines controlled retrieval-corpus expansion experiments with evaluations of 14 general-purpose models without task-specific fine-tuning and comparisons with eight specialized models.Training and retrieval coverage were restricted by transformation or time; after training, only the retrieval corpus was expanded while parameters and targets remained fixed.Source: Figure 1 · p. 2, §1 · p. 3

Study contribution

  • With parameters fixed, expanding retrieval improved prediction for transformations absent from fine-tuning data, including transformations represented by few records in the expanded corpus.A model trained on temporally restricted data also matched or exceeded a non-retrieval model trained on the full corpus, without further parameter updates.Source: §1 · p. 3

2 Results

The study uses a common reaction dataset and controlled retrieval-expansion interventions, then evaluates inference-time precedents in general-purpose models and multistep planning. Retrieval expansion substantially improved prediction for withheld transformations, while benefits in other settings depended on the test coverage and precedent selection.

2.1 A common data foundation for controlled retrieval corpus expansion

  • DianShi-PatentRxnStep-Dataset contains 2,712,259 pre-2025 training reactions and separate historical and 2025 evaluation sets, disjoint by standardized target product.The 2001–2024 validation and test sets each contain 3,798 reactions; the 2025 sets each contain 1,886.Source: §2.1 · pp. 3–4
  • RetroStepLM-RAG receives five retrieved reaction precedents with the target product during retrieval-conditioned supervised fine-tuning and inference.The paired RetroStepLM-Direct model predicts precursor sets from the target product alone.Source: §2.1 · p. 4

2.2 Expanding the retrieval corpus extends predictive coverage without further training

  • Mean Top-1 accuracy on withheld-template reactions rose from 13.0% to 58.0% in Test-2001-2024 and from 4.9% to 23.1% in Test-2025 after retrieval expansion.Retained-template means decreased modestly, from 69.9% to 68.5% and from 38.2% to 34.6%, respectively.Source: §2.2 · p. 4, Figure 2 · p. 5
  • With expanded retrieval, a Pre2017-trained model reached 70.7% versus 62.0% on Test-2001-2024 and 36.5% versus 36.4% on Test-2025, matching or exceeding full-corpus RetroStepLM-Direct.The Pre2017 model used 1.72 million reactions; full-corpus RetroStepLM-RAG scored 72.4% and 39.4% on those test sets.Source: §2.2 · pp. 5–6
  • On USPTO-50K, retrieval expansion left accuracy essentially unchanged at 73.5% versus 73.3%; the benchmark covers 1976–2016, within the Pre2017 training period.The expanded corpus added records from 2017–2024.Source: §2.2 · p. 6

2.3 Reaction precedents improve retrosynthesis without task-specific training

  • Across 42 model–test-set combinations, DianShi was the highest-performing retrieval corpus, and the strongest general-purpose model with DianShi precedents exceeded all eight evaluated specialized models on each test set.The strongest model was gemini-3.8-flash with DianShi precedents.Source: §2.3 · p. 6
  • Precedent selection affected prediction: reference-template restriction improved accuracy for all five diagnostic models, while matching-template exclusion reduced accuracy on all three test sets.On Test-2025, exclusion fell below the no-precedent baseline for gemini-3.8-flash and deepseek-v4.1-flash.Source: §2.3 · p. 6

2.4 DianShi single-step models improve solve rates across planning frameworks

  • Both DianShi single-step models had higher solve rates than every evaluated non-DianShi model in all nine planner–benchmark settings at N=50 and N=500.At N=500, their solve rates ranged from 84.00% to 97.37%, exceeding each setting’s strongest non-DianShi model by 5.00–31.06 percentage points; RetroStepLM-RAG did not lead RetroStepLM-Direct in every setting.Source: §2.4 · pp. 8–9, Figure 4 · p. 8

3 Discussion

The discussion argues that inference-time reaction precedents can extend prediction beyond task-specific training coverage, while emphasizing that benefits depend on precedent selection. It connects these gains to general-purpose model performance and multistep planning, while noting limits of exact-match evaluation.

Coverage beyond training data

  • With parameters fixed, expanded retrieval can improve predictions for transformations absent from training, including transformations with few precedents and records beyond the training period.The authors interpret this as a way to incorporate newly reported or proprietary reactions without updating model parameters.Source: §3 · p. 9

Precedent selection

  • Expanded retrieval does not improve accuracy uniformly: gains depend on whether the corpus contains the needed transformation and whether relevant records enter the precedent set.Template-informed diagnostics increased accuracy when selecting matching-template precedents and reduced it when excluding them.Source: §3 · p. 9

General-purpose models

  • The strongest general-purpose model with DianShi precedents outperformed all eight representative released specialized models evaluated across the three test sets.The discussion also notes that performance depended on the retrieval corpus.Source: §3 · p. 9

Complementary sources of knowledge

  • Inference-time reaction evidence can complement parameter-based learning by extending coverage without encoding every additional reaction in model parameters.The strongest evaluated predictor combined retrieval-conditioned fine-tuning with inference-time precedents.Source: §3 · p. 10

Multistep planning

  • Both DianShi models increased solve rates across all nine planner–benchmark settings, but higher single-step accuracy did not ensure an advantage in every planning setting.Exact matching against one recorded precursor set omits chemically plausible alternatives, and proposed routes still require experimental validation.Source: §3 · p. 10

4 Methods

The methods describe construction of a curated reaction dataset, controlled training-coverage interventions, and evaluations of retrieval-based, general-purpose, and specialized models. They also specify single-step scoring and multistep planning procedures.

4.1 Reaction dataset construction and partitioning

  • The dataset contains 2,712,259 pre-2025 training reactions, with historical and 2025 evaluation partitions constructed to be disjoint by standardized target product.Curation included standardization, deduplication, quality filtering, and temporal partitioning.Source: §4.1 · pp. 10–11

4.2 Fixed-parameter retrieval corpus expansion experiments

  • The coverage experiments withheld reaction templates or restricted training records to before 2017, then compared restricted and full retrieval corpora with the trained checkpoint fixed.Template withholding targeted approximately half of distinct test-set templates; the three partitions were alternative assignments, not independent training replicates.Source: §4.2 · p. 11

4.3 Reaction corpora and precedent retrieval

  • Retrieval used product fingerprints to rank candidate reactions by structural similarity, with five eligible precedents supplied at standard inference.External retrieval corpora included USPTO-derived and RXNGraphormer training reactions.Source: §4.3 · p. 12

4.4 DianShi single-step models and training

  • DianShi-RetroStepLM-Direct received the target product alone, whereas RetroStepLM-RAG was fine-tuned with five precedents to generate the same structured target sequence.Training minimized autoregressive negative log-likelihood over the complete target sequence.Source: §4.4 · pp. 12–13

4.5 Precedent-relevance diagnostics

  • Precedent-selection diagnostics compared standard structural-similarity retrieval with reference-informed restrictions and exclusions while holding the model and evaluation settings fixed.Reference classes and templates came from recorded answers, so these diagnostics were privileged rather than deployable retrieval settings.Source: §4.5 · p. 13

4.6 General-purpose model evaluation and specialized-model comparisons

  • Fourteen general-purpose models were evaluated with and without retrieved precedents, and eight released specialized retrosynthesis models were evaluated without retraining on DianShi data.General-purpose models were tested with precedents from three retrieval corpora across USPTO-50K and two DianShi test sets.Source: §4.6 · pp. 13–14

4.7 Single-step evaluation

  • Single-step scoring required an exact match to the complete standardized precursor set, with component order ignored and partial matches counted incorrect.Specialized models used Top-k accuracy across ranked candidates, while general-purpose models were assessed at Top-1.Source: §4.7 · p. 14

4.8 Multistep planning and evaluation

  • Multistep solve rate measured the fraction of targets with a complete route to inventory materials within a search budget across three planning frameworks and three benchmarks.Within each framework, planner, inventory, and search protocol were held fixed across single-step model comparisons.Source: §4.8 · pp. 14–15

5 Data availability

The paper states that its curated datasets and derived resources are available under a Creative Commons Attribution 4.0 license.

Data availability

  • The curated datasets and derived resources are available under a Creative Commons Attribution 4.0 International License.The listed resources include evaluation partitions, the PatMol-150 benchmark, the Train-Full template library, and the template-coverage cutoff scheme.Source: §5 · p. 15

Evaluation partitions

  • The available evaluation partitions include DianShi-PatentRxnStep-Test-2001-2024 and DianShi-PatentRxnStep-Test-2025.The passage lists 3,798 reactions for the former and 1,886 for the latter.Source: §5 · p. 15

Other released resources

  • The listed derived resources also include the PatMol-150 multistep benchmark and the Train-Full template library used for reaction validation.The template-coverage cutoff scheme includes retained and withheld assignments for seeds 42, 43, and 44.Source: §5 · p. 15

8 Extended Data

The Extended Data figures cover template-restricted accuracy, precedent-frequency ranges, temporal restriction, precedent-selection conditions, and multistep solve rates.

Extended Data Fig. 1

  • Extended Data Fig. 1 concerns Top-k accuracy under template restriction and retrieval-corpus expansion.The associated description identifies retained- and withheld-template test subsets and averages across three template partitions.Source: Figure 1 · p. 19, §8 · p. 19

Extended Data Fig. 2

  • Extended Data Fig. 2 concerns prediction accuracy across precedent-frequency ranges for withheld-template reactions.The ranges are based on the frequency of the local reaction template in Train-Full.Source: Figure 2 · p. 19, §8 · p. 20

Extended Data Fig. 3

  • Extended Data Fig. 3 concerns Top-k accuracy under temporal restriction and retrieval-corpus expansion.The temporal intervention compares a Pre2017 corpus with the full corpus while keeping the checkpoint fixed.Source: Figure 3 · p. 20

Extended Data Fig. 4

  • Extended Data Fig. 4 concerns prediction accuracy under different precedent-selection conditions.The described conditions include standard retrieval, reference-class or reference-template restrictions, and matching-template exclusion.Source: Figure 4 · p. 20, §8 · p. 20

Extended Data Fig. 5

  • Extended Data Fig. 5 presents multistep solve rates across search budgets for three planning frameworks and three benchmarks.Within each framework, the planner, starting-material inventory, and search protocol are fixed.Source: Figure 5 · p. 21, §8 · p. 21

9 Supplementary Information

The dataset uses five mutually disjoint base partitions, and coverage experiments compare retrieval conditions using the same trained checkpoint. The supplementary configuration labels distinguish full, time-restricted and template-restricted resources.

1.1 Dataset partitions

  • The five base partitions are mutually disjoint by standardized target product, and their membership stays fixed across training and retrieval configurations.Years refer to the source-patent year recorded in DianShi-RxnDB.Source: §1.1 · p. 22

1.2 Training and retrieval configurations

  • Full denotes Train-Full, Pre2017 denotes its pre-2017 subset, and TemplateRestricted labels denote subsets with the corresponding withheld templates removed.The template-restricted variants are S42, S43 and S44.Source: §1.2 · p. 22
  • For each restricted RetroStepLM-RAG configuration, the restricted record set supplies both fine-tuning reactions and the matching restricted retrieval corpus.The same checkpoint is then evaluated with either that restricted corpus or the full retrieval corpus.Source: §1.2 · p. 22

1.3 Template-defined evaluation subsets and reporting

Template-based evaluation separates reactions whose templates remain represented in restricted training and retrieval corpora from those whose templates are withheld. The supplement also specifies configuration comparisons, scoring summaries and reaction-record preprocessing.

1.3 Template-defined evaluation subsets and reporting

  • Within each template partition, withheld templates have no reactions in the restricted training or matching retrieval corpus, while retained templates have at least one.The three alternative partitions withhold approximately half of the distinct templates represented in the two base test sets.Source: §1.3 · p. 23
  • Retained- and withheld-template subsets are mutually exclusive and together comprise each corresponding base test set.Subset names combine the evaluation window, template status and partition seed.Source: §1.3 · p. 23
  • Single-step accuracy measures exact agreement with the complete standardized reference precursor set, and cross-partition summaries average the three partition-specific accuracies.Mean gains average the three within-partition accuracy differences; evaluation reactions are not pooled across partitions as independent samples.Source: §1.3 · p. 23

1.4 Configuration correspondence

  • RetroStepLM-Direct receives no retrieved precedents, while RetroStepLM-RAG’s training retrieval corpus matches its training corpus and its inference corpus is specified separately.The full-corpus RetroStepLM-RAG reference has separately trained parameters, unlike a restricted-training checkpoint evaluated with full retrieval.Source: §1.4 · pp. 23–24
  • In paired RetroStepLM-RAG comparisons, the checkpoint, targets, retrieval procedure, precedent serialization, prompting, decoding and scoring stay fixed; only the retrieval corpus and index change.Standard inference supplies five retrieved precedents per query.Source: §1.4 · pp. 23–24

2.1 Reaction standardization and deduplication

  • DianShi-RxnDB is a provenance-linked reaction-data platform that extracts structured records from patent text, images and reaction schemes, primarily from USPTO and EPO organic-synthesis patents.Records represent reported single-step experiments and link to their source patents.Source: §2.1 · p. 24
  • Reaction records were standardized with RDKit, and duplicates were identified by standardized reactant-set–product-set signatures after removing mapping numbers and sorting components.For duplicates spanning multiple years, the earliest recorded year was retained.Source: §2.1 · p. 24

2.2 Reaction-quality filtering and template processing

Reaction-quality filtering excludes records that fail structural or mapping checks, while template-frequency filtering removes rare local templates before temporal partitioning. Standardization and deduplication help define the records entering the resulting partitions.

2.2 Reaction-quality filtering and template processing

  • Quality filtering excludes records with multiple products, repeated molecules, reactant–product structural overlap, more than four reactant components, or inorganic products.Records with inconsistent atom mappings or unsuccessful reaction-centre or template extraction are also excluded.Source: §2.2 · p. 24
  • Local reaction templates represented by fewer than six reactions are removed before temporal partitioning.The threshold applies to template frequency in the corpus at that processing stage.Source: §2.2 · p. 24

2.3 Temporal partitioning and product-level exclusions

Temporal evaluation separates pre-2025 records from a 2025 held-out pool, with the base partitions jointly constructed to avoid shared standardized target products. Product-level screening also controls overlap with external comparison datasets.

2.3 Temporal partitioning and product-level exclusions

  • Records before 2025 supply training and historical evaluation pools, while records from 2025 supply the temporally held-out evaluation pool.The five base partitions are mutually disjoint by standardized target product, so overlapping year ranges do not imply shared target products.Source: §2.3 · p. 25
  • Product-level exclusions screen training candidates and DianShi evaluation candidates against specified USPTO, USPTO-MIT and RXNGraphormer-pretraining dataset partitions.These exclusions compare standardized target products irrespective of precursor set.Source: §2.3 · p. 25

2.4 Final dataset composition

  • Train-Full contains 2,712,259 reactions, while each historical validation and test partition contains 3,798 reactions and each 2025 validation and test partition contains 1,886.Train-Full spans 1976–2024 and supplies all records in the full DianShi retrieval corpus.Source: §2.4 · p. 25

9.3 Supplementary Note 3. Precedent retrieval, intermediate-label supervision and model training

The retrieval procedure ranks product-similar reactions, selects precedents for training or inference, and supplies precedent fields alongside the query product. Corpus-expansion comparisons keep the model checkpoint and retrieval procedure fixed.

3.1 Retrieval implementation and precedent selection

  • A product fingerprint retrieves 200 candidates by Hamming distance, which are reranked by Tanimoto similarity.The query uses no reaction-class or template information.Source: §3.1 · p. 25
  • Training randomly selects five eligible precedents from the ten highest-similarity candidates; standard inference uses the five highest-ranked eligible precedents deterministically.Training excludes the query reaction, duplicates, and potentially answer-disclosing records from the same patent.Source: §3.1 · pp. 25–26
  • Corpus-expansion comparisons keep the checkpoint and retrieval procedure unchanged, switching only between indices built from restricted and full record sets.Indices use the full training partition or the corresponding restricted training corpus, without corpus-level subsampling.Source: §3.1 · p. 26

3.2 Precedent representation and model inputs

  • Each precedent supplies product and reactant SMILES plus synthons and leaving groups; RetroStepLM-RAG receives these with the query product, unlike RetroStepLM-Direct.The query reaction's reference reactants, synthons, and leaving groups construct training targets, not query inputs.Source: §3.2 · p. 26

3.3 Intermediate-label supervision and prompt templates

Both DianShi models use intermediate-label supervision, with prompts that guide staged reconstruction of reactants. The RAG prompt adds retrieved reactions as context, while the direct prompt receives only the product.

3.3 Intermediate-label supervision and prompt templates

  • Both DianShi models supervise synthons, leaving groups, and final reactants in that order, with the reactant string enclosed in <answer> tags.The labels are derived from atom-mapped reactions and correspond to the retrosynthetic disconnection and standardized reactants.Source: §3.3 · p. 26
  • The direct prompt gives the product alone and requests the same staged inference of synthons, leaving groups, and reactants.The structured prompt and supervised target are shown separately, with <think> tags delimiting serialized intermediate fields.Source: §3.3 · p. 26
  • The supervised target serializes the query product, then its synthons and leaving groups, before the final inferred reactants.The target text presents intermediate deductions before the final reactant(s).Source: §3.3 · p. 27
  • The RAG prompt asks the model to infer synthons, identify leaving groups, and reconstruct reactants using the product and similar retrieved reactions.The final reactant SMILES must appear in <answer> tags; multiple reactants are separated by periods.Source: §3.3 · p. 27

RetroStepLM-RAG supervised target

This section examines retrieval-conditioned retrosynthetic prediction across task-specific coverage gaps and single-step and multistep evaluations. It considers how precedent selection, training choices, and retrieval-corpus coverage relate to performance, while interpreting multistep solve rates within the stated inventory and search protocol rather than as experimental executability.

3.4 Model training

  • RetroStepLM-RAG used retrieval-conditioned supervised fine-tuning with the target product and five precedents, and evaluated ranked beam-search candidates by complete precursor-set matching.Training minimized autoregressive negative log-likelihood over the supervised target, but only the final reactant set contributed to precursor-set accuracy; standardized component order did not affect correctness.Source: §3.4 · p. 27, RetroStepLM-RAG supervised target · p. 27, §3.5 · p. 28

4.1 Template-coverage intervention

  • Withheld-template Top-1 accuracy rose from 13.0% to 58.0% on Test-2001-2024 and from 4.9% to 23.1% on Test-2025 after retrieval expansion.Retained-template accuracy decreased from 69.9% to 68.5% and from 38.2% to 34.6%, respectively.Source: Table 7 · p. 29, §4.1 · p. 30

4.2 Precedent-frequency stratification

  • For templates represented by 1–100 reactions in the full corpus, retrieval expansion gained 36.7 percentage points on Test-2001-2024 and 7.6 percentage points on Test-2025.Expansion improved Top-1 accuracy in every partition and frequency range in both test sets.Source: §4.2 · p. 31

4.3 Temporal-coverage intervention

  • Retrieval expansion increased Top-1 accuracy by 12.1 percentage points on Test-2001-2024 and 7.8 percentage points on Test-2025 for the fixed Pre2017-trained checkpoint.USPTO-50K accuracy changed from 73.5% to 73.3%; the comparison does not isolate corpus size, recency, and composition effects.Source: §4.3 · pp. 31–32

5.3 Diagnostic results

  • Reference-template restriction improved accuracy for all five models on all three test sets, whereas matching-template exclusion reduced accuracy in every model–test-set combination.Reference-class restriction helped all four general-purpose models across the three test sets, but its effect on RetroStepLM-RAG varied by test set.Source: Table 11 · p. 33, §5.3 · p. 34

5.4 Interpretation and scope

  • Reference-informed selection uses classes or templates derived from the recorded answer, so its accuracy reflects privileged diagnostic conditions rather than generally available prospective retrieval.The diagnostics indicate that structurally similar products can arise from different transformations, and standard retrieval does not ensure transformation-matched precedents.Source: §5.2 · p. 33, §5.4 · p. 34

6.1 General-purpose models and inference settings

  • The evaluation covered 14 general-purpose models without task-specific updates and eight released specialized models without DianShi retraining.General-purpose models were tested on USPTO-50K, Test-2001-2024, and Test-2025; retrieval corpora differed in source, scale, temporal coverage, and curation.Source: §6.1 · p. 34, §6.2 · p. 35, §6.4 · p. 37

6.3 General-purpose model results

  • DianShi retrieval produced the highest accuracy in all 42 model–test-set combinations and improved every general-purpose model over its no-precedent condition on each test set.On Test-2025, gemini-3.8-flash scored 18.12% with RXNGraphormer retrieval, 22.85% without precedents, and 35.74% with DianShi retrieval.Source: Table 15 · p. 36, §6.3 · p. 37

6.5 Complete single-step comparison

  • Gemini-3.8-flash with DianShi retrieval exceeded the Top-1 accuracy of all eight evaluated released specialized models on each test set, and full-corpus RetroStepLM-RAG had the highest Top-1 accuracy among predictors reported in the note.Full-corpus RetroStepLM-RAG combines retrieval-conditioned fine-tuning with inference-time reaction precedents.Source: Table 17 · p. 38, §6.5 · p. 39

7.1 Multistep benchmarks; 7.3 Starting-material inventory and planning frameworks

  • Multistep evaluation used Pistachio Hard, USPTO-190 and PatMol-150, with a common inventory of 23,081,629 eMolecules building blocks and Qwen3.5-397B as planner across the three frameworks.Pistachio Hard was evaluated without its specified-starting-material constraint, and all benchmarks used the common inventory to define terminal precursor availability.Source: §7.1 · p. 39, §7.3 · p. 40

7.2 PatMol-150 construction

  • PatMol-150 contained 150 target molecules and 150 patent reference routes; route depths ranged from 7 to 13 steps (mean 8.17; median 8), and 41 routes (27.3%) were at least nine steps deep.Source: §7.2 · p. 40

7.4 Reaction validation and fallback; 7.5 Search budgets and solve-rate calculation

  • Solve rate measured complete-route finding under the specified inventory, validation and search protocol, not reaction conditions, selectivity, yield or experimental executability.Although the Retro-R1 default and other model’s template libraries were similar in size (381,302 versus 412,027 entries), Retro-R1 solved 49 percentage points fewer USPTO-190 targets within the same framework, so library size alone did not account for the difference.Source: §7.5 · pp. 41–42, §7.4 · p. 41

7.6 Complete planning results

  • At N = 50 and N = 500, both DianShi models exceeded every evaluated non-DianShi model in all nine framework–benchmark settings; at N = 500, DianShi solve rates ranged from 84.00% to 97.37%.RetroStepLM-Direct solved more targets than RetroStepLM-RAG in four of nine settings at N = 500 and two of nine at N = 50, despite RetroStepLM-RAG’s higher single-step Top-1 accuracy.Source: Table 23 · p. 42, §7.6 · p. 43

8.1 Intermediate-label and retrieval ablations

  • Intermediate labels increased accuracy at every reported rank in retrieved and nonretrieved trained models, and combining labels with retrieval achieved the highest accuracy among these models on all three test sets.Without intermediate labels, retrieval reduced Test-2025 Top-1 accuracy from 35.63% to 34.36%, while increasing Top-3, Top-5 and Top-10 accuracy.Source: Table 24 · p. 44, §8.1 · p. 44

8.2 Training-time precedent sampling; 8.3 Additional USPTO-50K adaptation

  • Random selection from the Top-10 eligible precedent candidates produced higher accuracy at every reported rank on all three test sets than deterministic selection of the five highest-similarity precedents.USPTO-50K adaptation increased accuracy at all reported ranks on USPTO-50K, but on both DianShi test sets Top-1 and Top-3 fell while Top-5 and Top-10 rose.Source: §8.2 · pp. 44–45, Table 25 · p. 44, §8.3 · p. 45

8.4 Retrieval corpus transfer for a USPTO-only RetroStepLM-RAG model

  • Both DianShi retrieval corpora improved every reported rank over USPTO-50K-train retrieval on all three test sets; full retrieval further improved every reported metric over Pre2017 on both DianShi test sets.On USPTO-50K, differences between the two DianShi corpora were small and varied by rank; the checkpoint’s task-specific training history remained fixed across retrieval conditions.Source: Table 27 · p. 46, §8.4 · p. 46

Related papers

Loading 2610.40249v1…