PaperlayerView exact revision on arXiv

Source-linked AI summary

Coverage Before Control: Route-Instruction Grounding and Steering for Controllable Retrosynthesis

Xuemin Chen, Xiaozhuang Song, Xinjian Zhao, Yaoyao Xu, Tianshu Yu

arXiv:2609.39955v1cs.AIcs.LG

TL;DR

Single-step retrosynthesis evaluations often recover recorded reactions without establishing whether models can choose among plausible alternatives to meet a chemist’s preference. RIGS learns those preferences in a projector and transfers them through residual adapters to steer a frozen generator. Experiments find instruction-sensitive generation, while coverage and control vary consistently but non-monotonically across model scales.

  • Problem

    Recovering a recorded precursor set does not establish whether a model can generate relevant alternatives and select among them according to a user’s request.

    Source: §1 · pp. 1–2

  • Method

    RIGS trains a language projector on product-specific candidate preferences, then transfers it through lightweight residual adapters to steer a frozen generator.

    Source: §1 · p. 2

  • Results

    Across nested supports and two model capacities, synthetic expansion improves recovery of known alternatives but further expansion does not monotonically improve coverage or instruction control; matched instructions promote preferred candidates and suppress conflicting ones.

    Source: §5 · p. 10

  • Takeaways & Limitations

    Coverage and instruction-dependent control are distinct: broader support supplies alternatives, while language guidance can move preferred alternatives higher in the returned list.

    Source: §1 · p. 2

    The evaluation focuses primarily on single-step retrosynthesis; the multistep case study is illustrative and its benefits for multistep planning have not been systematically evaluated.

    Source: J LIMITATIONS AND FUTURE WORK · p. 42

Abstract

from arXiv · show

Single-step retrosynthesis models are commonly evaluated by their ability to recover recorded reactions. In practice, chemists may need to choose among several precursor sets for the same product, for example to preserve a particular motif. Recovering a recorded answer alone does not establish this ability to follow a preference. Satisfying such requests requires both coverage of relevant alternatives and control over which alternatives are favored. We introduce Route-Instruction Grounding and Steering (RIGS), a two-stage framework for instruction-conditioned retrosynthesis. Stage A trains a language projector, teaching it which alternatives an instruction favors or discourages. Stage B uses the projector learned in Stage A to steer a frozen generative model through lightweight residual adapters. We construct nested one-to-many training supports by pairing each product with increasing numbers of candidate precursor sets. Extensive experiments demonstrate that broader support helps the model generate a wider range of alternatives, and RIGS can learn to guide generation according to instructions. The relationship between coverage and control is consistent across model scales but non-monotone.

1 INTRODUCTION

Single-step retrosynthesis must support both generating plausible precursor alternatives and selecting among them according to a chemist’s request. RIGS addresses this by grounding instructions in product-specific preferences and steering generation, with experiments showing instruction-sensitive ranking and a non-monotone relationship between coverage and control.

1 INTRODUCTION

  • Top-k exact-match accuracy recovers recorded precursor sets but does not directly test whether a model can choose among alternatives to satisfy a user’s request.The paper distinguishes coverage—alternatives recovered under finite sampling—from control, the instruction-dependent choice among them near the top of the returned list.Source: §1 · pp. 1–2
  • A motivating USPTO multi-route subset averages only 2.37 observed precursor sets per product, motivating synthetic expansion while distinguishing broader coverage from language responsiveness.Broader support alone does not establish instruction responsiveness.Source: §1 · p. 2
  • RIGS trains a projector on offline comparisons among candidates for the same product, then transfers it through residual adapters to a frozen GraphDiT generator for direct instruction-conditioned sampling.The framework does not depend on a fixed candidate list at inference.Source: §1 · p. 2
  • 18.52 percentage points: correct instructions improve preferred-set Top-1 recovery over shuffled instructions for the 65M Top30 model.Stage A grounding further helps suppress alternatives that conflict with the instruction.Source: §1 · p. 2

2 RELATED WORK

Prior work broadens retrosynthesis candidate generation and applies language or guidance to route selection. RIGS focuses on grounding instructions in preferences among precursor alternatives for one product and transferring those preferences into generation.

Single-step retrosynthesis and candidate coverage

  • Candidate-coverage methods include latent variables, reaction-class prompts, data augmentation, synthetic reactions, and template-based generation with predictive filtering.Tran et al. (2026) distinguish failures to propose candidates from failures to rank them highly.Source: §2 · p. 2

Grounding language in precursor preferences

  • Language-related retrosynthesis work uses generated descriptions to rank reaction centers and precursors or combines descriptions with molecular and route context.RIGS instead grounds instructions in preferences among precursor sets for the same product.Source: §2 · p. 2

Transferring preferences into generation

  • Route planning and retrosynthesis guidance include Pareto objectives, depth and starting-material conditioning, classifier guidance, and inference-time reward steering.RIGS transfers learned precursor preferences to a frozen GraphDiT through residual adapters for direct instruction-conditioned generation.Source: §2 · pp. 2–3

3 RIGS: ROUTE-INSTRUCTION GROUNDING AND STEERING

RIGS first constructs nested candidate supports and learns instruction preferences among alternatives for each product. It then transfers the learned grounding into residual adapters on a frozen generator to steer instruction-conditioned generation.

3 RIGS: ROUTE-INSTRUCTION GROUNDING AND STEERING

  • For each candidate budget and model capacity, RIGS uses a separate no-instruction Base model trained on the corresponding candidate support, with its backbone frozen during both guidance stages.Stage A learns a language projector from offline preferences, and Stage B connects it to generation through residual adapters.Source: §3 · p. 3

3.1 CANDIDATE SUPPORT AND PREFERENCE SUPERVISION

  • Template rules applied to a product yield precursor candidates, which are greedily ranked by heuristic quality and diversity before graph checks and deduplication form the Base support.Figure 1 depicts template-derived candidates broadening training support before instruction grounding.Source: §3.1 · p. 3, Figure 1 · p. 3
  • For budget K, the support contains filtered candidates among the first K ranks, producing nested supports as K increases.Preferred and avoided candidates within each support provide supervision for grounding and steering.Source: §3.1 · p. 3

3.2 STAGE A: GENERATOR-FREE GROUNDING

  • Stage A maps instruction-token features into the reaction-representation space using a trained projector, while the language and reaction-pair encoders remain frozen.Text features are computed offline and cached for reuse in both stages.Source: §3.2 · p. 3
  • Stage A trains the projector to match a predicted candidate-preference distribution to offline target weights that favor candidates better satisfying the instruction.The objective combines distribution matching with a pairwise preference term and updates only the projector.Source: §3.2 · pp. 3–4

3.3 STAGE B: RESIDUAL INSTRUCTION STEERING

  • Stage B initializes projectors from Stage A and uses node-to-text cross-attention adapters to add instruction-conditioned residual updates to graph-node states while keeping GraphDiT frozen.Setting steering strength α to zero recovers the Base model; the backbone uses node updates to adjust edge representations.Source: §3.3 · p. 5
  • Stage B trains projector and adapter parameters to denoise preferred candidates while regularizing departure from the frozen Base reference.Its objective combines positive flow matching, ranking, rollout-drift, and guarded-KL terms; the ranking term favors greater guidance gains for preferred than avoided candidates, while positive training directly reconstructs preferred candidates.Source: §3.3 · p. 5
  • Inference reuses the Base continuous-time Markov chain sampler, then canonicalizes valid draws, merges duplicates, and ranks distinct precursor candidates by sampling frequency.Checkpoint and guidance strength are selected on validation data and fixed during test evaluation.Source: §3.3 · p. 5

4 EXPERIMENTS

The experiments test candidate coverage and instruction control across nested training supports, model sizes, datasets, and guidance suites. Broader support expands recovered alternatives, while instruction following is effective but non-monotone; grounding and generation-time guidance add distinct benefits.

4.1 EXPERIMENTAL SETUP

  • The evaluation uses USPTO-MR and synthetic Top15–Top60 supports, two backbone sizes, a shared Top100 test inventory, and validation-selected checkpoints and guidance scales.Matched uses each case’s intended instruction, Shuffled uses a different instruction, and each arm receives 100 raw draws per group.Source: §4.1 · pp. 6–7

4.2 BROADER TRAINING SUPPORT IMPROVES CANDIDATE COVERAGE

  • From Top15 to Top45, pooled recovery rises from 2,348 to 12,880 for 8M and from 1,723 to 13,661 for 65M, then is nearly unchanged or falls at Top60.More distinct valid outputs at Top60 do not necessarily mean more recovered candidates in the test inventory.Source: §4.2 · p. 7, Figure 3 · p. 7

4.3 INSTRUCTIONS HELP THE MODEL RETURN CANDIDATES THAT MATCH THE REQUEST

  • At 65M Top30, Matched improves P@1 by 18.52 percentage points over Shuffled; instruction-following gains peak at Top30/Top45 and decline at Top60 across both capacities.At Top30, Matched also improves 65M P@1 by 15.90 percentage points over unguided generation, while its P@10 gain is 4.57 points.Source: §4.3 · pp. 7–8, Table 1 · p. 8

4.4 ROLES OF GROUNDING AND GENERATION-TIME GUIDANCE

  • Stage-A grounding increases the Matched–Shuffled reduction in N@10 by 6.82 percentage points for 8M Top30, chiefly suppressing candidates that conflict with the request.The comparison uses the same Stage-B training and inference settings on a separate test cohort.Source: §4.4 · p. 8, Table 2 · p. 9
  • With 100 samples per group, RIGS achieves higher P@1 than reranking, 32.73% versus 22.23%, while both methods include a preferred candidate in about 64% of groups.The reranker lowers N@1 but its P@1 falls below the unguided baseline’s 24.57%; RIGS improves both outcomes over unguided generation.Source: §4.4 · p. 8, Table 3 · p. 9

4.5 APPLICABILITY ACROSS DATASETS AND GUIDANCE SUITES

  • On Pistachio, Matched improves P@1 by 1.52 points and reduces N@10 by 7.19 points versus Shuffled; structural instructions yield 65M gains of 8.09 P@1 points and an 18.44-point N@10 decrease.Metric4 shows preferred-candidate promotion and avoided-candidate suppression at both capacities, with improved instruction following in seven of eight synthetic-support settings.Source: §4.5 · p. 9, Table 4 · p. 9
  • In a case study, conditioning each expansion on using fewer halogen atoms yields a complete multistep route with halogen-free starting materials.RIGS serves as the single-step precursor generator in a custom value-guided AND–OR tree-search planner.Source: §4.5 · pp. 9–10

5 CONCLUSION

The conclusion presents RIGS as a way to ground chemists’ preferences in precursor alternatives and steer a frozen graph generator. Its results show that coverage and instruction control interact non-monotonically, with Stage-A grounding especially benefiting conflict suppression.

5 CONCLUSION

  • RIGS grounds instructions in preferences among alternatives for the same product, then transfers the learned representation to a frozen graph generator through lightweight residual adapters.The evaluation concerns single-step retrosynthesis and instruction-responsive precursor alternatives.Source: §5 · p. 10
  • Synthetic expansion improves recovery over observed-reaction training, but further expansion does not monotonically improve coverage or instruction control across two model capacities.The conclusion describes this coverage–control relationship as consistent but non-monotone.Source: §5 · p. 10
  • Stage-A grounding provides its clearest additional benefit in suppressing conflicting alternatives, while case-specific structural instructions support control beyond predefined molecular descriptors.Source: §5 · p. 10

A COMMON-INVENTORY PROTOCOL AND FROZEN COORDINATES

The common-inventory study compares frozen model coordinates selected on validation across nested supports and capacities. It evaluates instruction effects, candidate recovery, and validity on a shared test while documenting product overlap with an earlier exploratory cohort.

A COMMON-INVENTORY PROTOCOL AND FROZEN COORDINATES

  • Top15/30/45/60 are nested synthetic-only prefixes of one product-local ordering; USPTO-MR is a separate natural-data reference.The primary recipe uses within-family data passes, so realized update counts vary with support size.Source: A COMMON-INVENTORY PROTOCOL AND FROZEN… · p. 15
  • The ten model coordinates were selected on validation before opening the Common Top100 test, with Reference, Matched, and Shuffled arms for every model.The models span 8M and 65M capacities and the listed supports.Source: Table 5 · p. 15
  • The test contains 1,201 cases, 1,280 paired semantic families, and 2,560 signed groups balanced across five RDKit axes and both polarities.The ten-model, three-arm study contains 7,680,000 trajectories.Source: A COMMON-INVENTORY PROTOCOL AND FROZEN… · p. 15
  • P@k and N@k measure whether ranked candidates include a preferred or avoided set, while Hitall@B and Hit+@B measure inventory and preferred-set recovery among raw draws.Raw validity concerns molecular representation, not experimental feasibility.Source: A COMMON-INVENTORY PROTOCOL AND FROZEN… · p. 15
  • The earlier exploratory cohort shares 154 of its 988 products with Common Top100, so the test products are not wholly independent of that evaluation.The overlap is 12.82% of Common Top100; reported model coordinates were fixed before its outputs were inspected.Source: A COMMON-INVENTORY PROTOCOL AND FROZEN… · p. 16

B COMMON TOP100 RESULTS AND DIAGNOSTICS

The Common Top100 results show that broader support can expand generated alternatives, while instruction guidance primarily changes which sampled candidates are promoted or suppressed. Results vary with support, capacity, and inference calibration, and relaxed round-trip findings carry a reference-precursor shortcut caveat.

B COMMON TOP100 RESULTS AND DIAGNOSTICS

  • Top60 generates the most distinct valid outputs at B = 100 but has the lowest Hit@1 among synthetic-support models at each capacity.Table 6 reports reference coverage endpoints; its columns other than Hit@1 use B = 100 without backfilling.Source: B COMMON TOP100 RESULTS AND DIAGNOSTICS · p. 16, Table 6 · p. 16
  • Top30 and Top45 show preferred-route promotion and avoided-route suppression at both capacities, while 65M Top60 remains unresolved at these operating points.At 8M, Top30–Top45 intervals include zero; Top30 improves over Top15, while Top60 weakens relative to Top45. Effects are not uniform across axes.Source: B COMMON TOP100 RESULTS AND DIAGNOSTICS · p. 16, Table 7 · p. 17

B.1 ABSOLUTE ARM RESULTS

  • Matched and Shuffled can have similar validity while both lose validity relative to Reference.The absolute-arm results use the primary equal-axis, case-balanced aggregation at B = 100.Source: §B.1 · p. 16, Table 8 · p. 18

B.2 RETURNED-LIST CONTRASTS WITH THE UNGUIDED MODEL

  • Early-rank preferred-route gains relative to unguided generation do not necessarily extend to Top-10, while avoided-route suppression can persist across the list.Guidance can move an already available preferred route to the front even when later-rank promotion is small.Source: §B.2 · p. 16, Table 9 · p. 19

B.3 COMMON-UPDATE-CEILING SUPPORT SENSITIVITY

  • Positive control persists under common training ceilings, but the 8M peak shifts to Top30, so support choice depends on training and selection rather than candidate count alone.Top60 reuses its primary coordinates; the table reports common-ceiling deployment coordinates and Matched–Shuffled effects.Source: §B.3 · p. 18, Table 10 · p. 20

B.4 RELAXED ROUND-TRIP COVERAGE AND ACCURACY ACROSS CANDIDATE RANKS

  • Guidance raises relaxed round-trip hits at K = 1 by 3.00–13.50 points (paired stratified bootstrap 95% interval), but the advantage fades deeper in the list.At K = 10, equal group coverage with lower fixed-slot accuracy means guidance reaches as many groups with a hit but returns fewer successful candidates per group; the reference-precursor shortcut can accept candidates without forward-model closure.Source: §B.4 · p. 19, Table 11 · p. 20

C.1 VALIDATION GUIDANCE-STRENGTH TRADE-OFF

  • Stronger inference intervention can favor preferred routes while disrupting valid generation, motivating calibration against both preference recovery and output validity.The 65M scale sweeps use each support’s original epoch-14 validation cohort and do not isolate a causal effect of training support.Source: §C.1 · p. 20, Figure 5 · p. 21

C.2 MATCHED RESIDUAL-SCALE SENSITIVITY

  • At matched α, the Top45–Top60 paired 95% intervals include zero at all four 8M recovery endpoints, so the support ranking partly depends on deployment strength.The sensitivity uses αeval = .25.Source: §C.2 · p. 21, Table 12 · p. 21

C.4 COMPLETE THREE-SEED ANALYSIS ACROSS SUPPORTS AND CAPACITIES

Three-seed results support instruction control at intermediate synthetic supports, but do not identify a universal best support because control and validity costs vary. The Stage-A initialization test most consistently supports avoided-route suppression; promotion and validity evidence remain uncertain.

C.4 COMPLETE THREE-SEED ANALYSIS ACROSS SUPPORTS AND CAPACITIES

  • Across three seeds, all 8M synthetic supports improve P@1 against Shuffled and Reference, while 65M Top30/Top45 retain strong gains and Top60 has negative net P@1 effects in every seed.USPTO-MR has near-zero P@1 instruction-content effects at both capacities; repeats do not establish a universal best support.Source: §C.4 · p. 22, Table 13 · p. 22
  • Validity costs differ across supports: 8M Top30 loses more validity than Top45, and all three 65M Top15 runs use fallback selections that failed the original validation hard gate.The comparisons reflect different validity costs and selection rules, not a shared validity budget.Source: §C.4 · p. 22, Table 13 · p. 22

D.1 DIRECT STAGE-B: EFFECT OF STAGE-A-ALIGNED INITIALIZATION

  • In single-seed validation sweeps, Stage-A-aligned RIGS improves N@10 at every 8M checkpoint, while preferred-route recovery does not improve uniformly across ranks.The 65M matched-scale sweep also strengthens control with a modest validity cost.Source: §D.1 · p. 23, Table 14 · p. 23
  • On the frozen 8M test, all four N@k difference-in-differences intervals favor Stage A, but preferred-route and validity intervals cross zero.The evidence does not establish a promotion gain; intervals lack multiplicity correction and training-seed uncertainty, and 65M evidence is validation-only.Source: §D.1 · pp. 23–24, Table 15 · p. 23

D.2 STAGE-B LOSS ABLATION

The ablation identifies positive DFM and route ranking as the clearest contributors to instruction-specific control, while other objective terms show validity effects or unresolved control gains.

D.2 STAGE-B LOSS ABLATION

  • Each leave-one-out arm zeroes one coefficient while sharing the checkpoint, selected Stage-A projector, initialization, data order, optimizer, and 14-epoch, 57,260-update budget.Evaluation uses one frozen validation cohort at epoch 14 and α_eval = .50, so the comparison is single-seed and validation-only.Source: §D.2 · p. 24, Figure 7 · p. 24
  • The full objective combines positive DFM, route ranking, metric-teacher KL, full-reference KL guard, and rollout-reference KL at weights 0.10, 0.05, 0.05, 0.05, and 1.00.The full arm has E_P = 14.89, E_N = 14.40, and 68.97% matched validity versus 70.22% for Reference.Source: §D.2 · p. 24
  • Removing positive DFM or route ranking weakens promotion and suppression, with paired intervals excluding zero.Metric-teacher and rollout-reference KL support validity without a resolved control gain; deleting the guard improves validity, but its protective role is not established.Source: §D.2 · p. 24

E.1 BUDGETED PROPOSAL–SELECTION COUNTERFACTUAL

The proposal–selection analysis separates candidate availability from ranking, finding that guidance can prioritize preferred routes within sampled pools. Free-form structural instructions also promote and suppress case-specific routes at both model capacities.

E.1 BUDGETED PROPOSAL–SELECTION COUNTERFACTUAL

  • Selectors reorder valid distinct candidates within realized pools, while Hit@B and unique-valid counts remain unchanged by reordering.Frequency, learned scalar-head, Stage-A cosine, and RDKit-oracle readouts provide distinct selection counterfactuals.Source: §E.1 · p. 25
  • At B = 100, guidance resolves higher P@1 and lower N@1 without a resolved increase in preferred-set Hit@100, supporting prioritization within sampled pools.Frequency ordering also yields fewer unique-valid candidates; the table reports the primary Reference and Matched pools.Source: §E.1 · p. 25, Table 16 · p. 26
  • Scalar heads reduce N@1 and metric regret on both pools but lower high-budget P@1; descriptor-optimal oracle ranking can still miss frozen-set positives.The trade-off varies across head seeds, and Table 17 reports seed sensitivity.Source: §E.1 · p. 25, Table 17 · p. 26

F.1 FREE-FORM STRUCTURAL INSTRUCTIONS AND THE FROZEN LANGUAGE REPRESENTATION

  • Structure guidance uses case-specific retain-or-avoid instructions and frozen language features, which the projector transfers to route scoring and node-to-text memory.The frozen test split contains 2,229 cases, 6,752 semantic groups, and 4,606 instruction strings.Source: §F.1 · pp. 25, 27
  • At 8M and 65M, Matched–Reference P@1 gains are 5.54 and 14.74 percentage points, while N@10 reductions are 10.57 and 35.79 points, respectively.Both capacities promote preferred routes and suppress avoided routes at Top-1 and Top-10; Table 19 reports case-balanced Matched–Shuffled effects.Source: §F.1 · p. 27, Table 19 · p. 27

F.2 A CHEMISTRY-AUDITED PAIRED CASE: FEWER PRECURSOR HALOGENS

Instruction guidance changes which recorded routes are returned first and reallocates repeated draws toward preferred routes. Paired cases also illustrate compositional shifts and possible effects on downstream plans, without estimating average planning gains.

F.2 A CHEMISTRY-AUDITED PAIRED CASE: FEWER PRECURSOR HALOGENS

  • In the halogen-low case, the paired fixed-seed outputs shift from a one-halogen Ullmann disconnection to a zero-halogen Chan–Lam disconnection.Across 100 draws, Reference has 21 negative and 19 positive hits at 91% validity; Guided has 0 negative and 24 positive hits at 85%. The examples are curated, not unfiltered Top-5 outputs.Source: §F.2 · p. 28, Figure 8 · p. 28

F.3 DOWNSTREAM MULTISTEP PLANNING WITNESSES

  • In the ambiguity-low witness, guidance changes every selected reaction and terminal precursor while keeping proof depth equal; automatic checks pass, but the chemistry was not manually audited.In a separate halogen-low witness, guidance changes the root and later reactions to yield halogen-free terminal precursors. These paired witnesses do not estimate average gains, cross-seed stability, or experimental feasibility.Source: §F.3 · pp. 28–29

F.4 DIRECT EMPIRICAL SAMPLED-MASS REALLOCATION

  • Matched instructions reallocate repeated draws toward preferred routes and away from avoided routes at all ten coordinates, with paired intervals excluding zero in both directions.Validity contrasts remain unresolved, and invalid draws remain in the denominator; Table 20 reports the B = 100 effects.Source: §F.4 · p. 29, Table 20 · p. 31

G.1 PISTACHIO TOP30: CONTROL ON RECORDED REACTION ALTERNATIVES

  • On recorded Pistachio alternatives, Matched improves P@1 and P@10 and reduces N@10 versus Reference, while preferred-set Hit@100 has no resolved increase.Against Shuffled, paired intervals resolve preferred-route promotion and avoided-route suppression at Top-10; a validity cost is visible in Table 21.Source: §G.1 · p. 31, Table 21 · p. 31

G.2 APPLICABILITY TO ANOTHER METRIC SUITE

Metric4 results indicate instruction control beyond RDKit5, with outcomes depending on support and deployment setting. The accompanying appendices specify the evaluation scope, data construction, candidate ranking, and guidance supervision.

G.2 APPLICABILITY TO ANOTHER METRIC SUITE

  • Seven of eight synthetic-support settings show preferred Top-1 promotion and avoided Top-10 suppression against Shuffled, with preferred Top-1 gains over Reference; USPTO-MR and 65M Top60 do not resolve P@1 effects.Validity costs versus Reference remain visible, while Matched–Shuffled validity intervals include zero at all ten settings; Table 22 covers the ten Metric4 settings.Source: §G.2 · pp. 31–32, Table 22 · p. 32

H.1 IDENTIFIERS, DATA UNITS, AND REPORTING SCOPE

  • Reporting distinguishes model identifiers and evaluation conditions, as well as product/case, route, group, and sampled-rollout counting units.Base identifies a no-instruction generator; Guide names the supervision suite, while scale and support identify the backbone and training candidate view.Source: §H.1 · p. 32, Table 23 · p. 33

H.2 CONSTRUCTION OF USPTO-MR

  • USPTO-MR retains products with at least two distinct recorded precursor combinations and contains 32,584 products and 77,110 product–precursor pairs after preprocessing.All reactions for a canonical product stay in one split to prevent product overlap.Source: §H.2 · p. 33, Table 24 · p. 33

H.3 BASE-GENERATOR DATA MATRIX

  • Synthetic supports share 270,794 templates; USPTO Top15/30/45/60 are nested synthetic-only prefixes, whereas USPTO-MR is a natural-data baseline.The Pistachio Top30 training and validation sets are expanded, but its registered test split contains recorded reactions only.Source: §H.3 · p. 33, Table 24 · p. 33

H.4 SYNTHETIC PRECURSOR CONSTRUCTION AND RANKING

  • The synthetic candidate selector greedily balances heuristic quality and precursor diversity, producing nested prefixes as K changes only the stopping point.Candidates are complete precursor sets; graph checks and deduplication can make usable Base support smaller than K.Source: §H.4 · pp. 34–35

H.5 GUIDANCE SUITES AND CONSTRUCTION

  • Metric4 and RDKit5 use fixed metric registries with signed directions, whereas Structure guidance is case-specific; preference targets use retained upper- and lower-quartile candidates.Middle or missing-value candidates and groups without a valid contrast are excluded.Source: §H.5 · p. 35, §H.6 · p. 35

H.6 GUIDANCE-DATA MATRIX

  • After support restriction and graph filtering, RDKit5 and Metric4 retain only groups containing both preferred and avoided candidates, so group counts need not rise monotonically with TopK.The reported counts are graph-bound training identities, not candidate occurrences multiplied by semantic groups.Source: §H.6 · p. 35

H.7 FROZEN EVALUATION COHORTS

The appendices specify the evaluation cohorts and protocols, then detail how frozen representations, preference labels, a projector, and residual adapters implement instruction-conditioned steering.

H.7 FROZEN EVALUATION COHORTS

  • The RDKit5 scaling cohort fixes a shared strict-Top100 test universe, with 256 paired families and 512 signed groups per axis across five axes.All selected cases have 100 raw candidates, though fewer routes may survive graph binding; Table 28 distinguishes raw inventory from graph-resolved routes.Source: §H.7 · pp. 35, 37, Table 28 · p. 37

H.8 BASE-GENERATOR CHECKPOINT PROVENANCE; H.9 SAMPLING AND STATISTICAL PROTOCOL

  • Base checkpoints are selected by validation deduplicated Top-3 accuracy, while sampling uses ordered RC Top-5 proposals and 100 raw trajectories per group and arm.Matched and Shuffled share proposals, root choices, and seeds; outputs are canonicalized, deduplicated, and frequency-ranked without backfilling short lists. Cohorts and uncertainty use fixed sampling and case-balanced aggregation with 10,000 shared case-cluster bootstrap resamples.Source: §H.8 · p. 37, §H.9 · p. 37

I ARCHITECTURE, OBJECTIVES, AND IMPLEMENTATION; I.1 FROZEN STAGE-A REPRESENTATIONS

  • Stage A uses frozen instruction and route encoders, while guided scaling varies route representations as well as candidate support.Projector fitting uses cached features, but native-cache construction runs the frozen GraphDiT backbone; the reported Stage-A compute budget excludes cache construction and historical semantic-teacher pretraining.Source: §I.1 · p. 38, I ARCHITECTURE, OBJECTIVES, AND… · p. 38

I.2 STRUCTURE ANNOTATION AND QUALITY CHECKS

  • Structure instructions and candidate labels come from an annotator, with automated checks enforcing executable preference contrasts but not independently adjudicating chemical correctness.Chemical label correctness and distinctness of instruction axes remain LLM judgments without exhaustive expert review.Source: §I.2 · p. 38

I.3 STAGE-A GROUNDING AND EXACT ROUTE SCORING

  • Stage A projects instruction tokens, averages and normalizes them into an instruction representation, then scores candidate routes against it.Candidate weights are instruction-independent, derived from heuristic selection scores or recorded route counts, normalized over retained candidates, and floored at 10^-8; transfer copies the validation-selected projector while encoders remain frozen.Source: §I.3 · p. 39

I.4 STAGE-B RESIDUAL ADAPTER PATH

  • Stage B transfers the learned projector into trainable node-to-text adapters at selected layers, injecting guidance into node states while edge states respond through the frozen backbone.Instruction-token memory is cached for inference; the 8M and 65M models use different projector-sharing groups and each listed layer has its own cross-attention adapter.Source: §I.4 · p. 39, Figure 10 · p. 40

I.5 STAGE-B LOSS COMPONENTS AND METRIC TEACHER

  • Stage B’s positive flow-matching term anchors preferred candidates, while ranking compares preferred–avoided gains and KL terms constrain drift on training and sampler-reached states.The metric teacher is an offline deterministic transformation of route metrics and is absent at inference; the full-reference guard applies a threshold before averaging.Source: §I.5 · pp. 39–41

I.6 NOTATION

  • The notation distinguishes support size K, raw-draw budget B, and returned-list cutoff k, which denote separate quantities.The section also uses validation-selected parameters marked by ⋆ and α as shorthand for αeval in tables and plots.Source: §I.6 · p. 41, Table 30 · p. 41

I.7 IMPLEMENTATION AND REPRODUCIBILITY DETAILS

Implementation settings fix the training objectives and sampling procedure; the paper’s evaluation is primarily single-step, with multistep benefits not systematically tested.

I.7 IMPLEMENTATION AND REPRODUCIBILITY DETAILS

  • Stage A and Stage B use learning rates 10^-4 and 3 × 10^-5, respectively, with RDKit5 Common Top100 training for 30/14 epochs at 8M and 65M.The selected Structure checkpoints are reported in Table 31; both capacities use Matched-P@1-first selection.Source: §I.7 · pp. 41–42, Table 31 · p. 42
  • Inference scales only the cross-attention output residual, and the unchanged CTMC sampler uses 50 steps in the reported evaluations.The terminal graph is marginalized under the model prediction rather than supplied as ground truth.Source: §I.7 · p. 42

J LIMITATIONS AND FUTURE WORK

  • Evaluation focuses primarily on single-step retrosynthesis; the multistep case study does not systematically evaluate benefits for multistep planning.Future work includes planner integration and post-training to improve output validity and instruction following.Source: J LIMITATIONS AND FUTURE WORK · p. 42

Related papers

Loading 2609.39955v1…