PaperlayerView exact revision on arXiv

Source-linked AI summary

Solvent: Long-Horizon Reasoning for Multi-Step Retrosynthesis

Yifei Wang, Botian Wang, YiFei Wang, Jiangtao Feng, Xinbo Zhang, Yawen Ouyang, Yixin Wu, Lihao Wang, Xiaoqing Zheng, Xuanjing Huang, Hao Zhou

arXiv:2610.32308v1cs.AI

TL;DR

Multi-step retrosynthesis requires route planning across interdependent decisions, but existing search-based approaches face coverage and cost constraints, and long-horizon LLM reasoning remains underexplored for the task. Solvent uses an 8B model trained with a two-round curriculum to plan, verify precursor availability, and revise routes over multiple turns. Across four benchmarks, it rivals frontier LLMs and outperforms prior search-based systems, while in-silico verification remains short of experimental validation.

  • Problem

    Search-based retrosynthesis depends on single-step model coverage, becomes costly for deeper routes, and leaves global chemical analysis and disconnection rationale implicit; long-horizon LLM reasoning remains underexplored for this task.

    Source: §1 · p. 2

  • Method

    Solvent is an 8B LLM trained with a two-round post-training curriculum to plan routes iteratively, verify precursor availability against inventory, and revise routes from feedback.

    Source: §1 · p. 2, §4.1 · p. 4

  • Results

    Across four benchmarks, Solvent rivals frontier LLMs and outperforms prior search-based retrosynthesis systems; its average success rate is 61.86% on full test sets.

    Source: §5.2 · p. 7, §8 · p. 10

  • Takeaways & Limitations

    Solvent offers a long-horizon paradigm for multi-step retrosynthesis, with case studies showing protecting-group schemes and convergent routes where appropriate.

    Source: §8 · p. 10

    In-silico checks can proxy reaction feasibility but cannot establish experimental success, yield, selectivity, reproducibility, or operational safety.

    Source: §7 · p. 9

Abstract

from arXiv · show

Multi-step retrosynthesis requires recursively transforming a target molecule through successive reaction steps into purchasable building blocks, and recent work has begun to draw on large language model (LLM) reasoning to tackle the task. In its latest evolution, LLM reasoning has entered the long-horizon regime: models pursue objectives across interdependent steps, interact with environments, and adapt to feedback. This shift has markedly advanced LLM capabilities, as illustrated by breakthroughs in AI-driven mathematical research. At this frontier, we introduce Solvent, an 8B LLM that successfully extends long-horizon reasoning to multi-step retrosynthesis. Through a two-round post-training curriculum, Solvent learns to interleave iterative synthesis planning, external availability verification, and adaptive route revision. Experiments show that, following a $2.1\times$ performance boost from RL, Solvent scores 61.86% across four benchmarks, surpassing mid-scale language models and prior search-based retrosynthesis systems, and rivaling frontier trillion-parameter LLMs.

1 INTRODUCTION

Multi-step retrosynthesis requires recursively decomposing a target into accessible precursors, but established search-based methods face coverage and cost constraints. Solvent applies long-horizon reasoning to the task and reports strong results across four benchmarks.

Motivation

  • Traditional search-based planning depends on single-step model coverage, incurs substantial costs for deeper routes, and leaves global analysis and disconnection rationale implicit.Prior work improved local proposals and learned value functions to prioritize search expansions.Source: §1 · p. 2
  • Long-horizon reasoning remains underexplored in multi-step retrosynthesis, where routes require recursive subproblems, precursor checks, and feedback-guided revision.Existing LLM approaches use external optimization or search, plan over externally supplied reaction candidates, or generate a complete route in one turn.Source: §1 · p. 2

Contribution

  • Solvent is an 8B LLM that plans iteratively, verifies precursor availability, revises routes, and uses a two-round post-training curriculum before final RL.The curriculum first develops a seed policy with retrosynthetic knowledge and stable reasoning, then bootstraps a stronger starting policy.Source: §1 · p. 2

Results

  • 61.86% average success across four benchmarks lets Solvent rival frontier LLMs, surpass similarly sized retrosynthesis-specialized models, and outperform prior search-based systems.The reported comparison with R3 is a 42.83-percentage-point margin.Source: §1 · p. 2

2 RELATED WORK

Long-horizon agents work across interdependent steps through sustained interaction and feedback. Prior LLM retrosynthesis approaches use external search or generate routes in one turn, whereas Solvent performs planning and revision over multiple turns.

2.1 LONG-HORIZON TASKS AND AGENTS

  • Long-horizon tasks involve dependent decisions whose earlier outcomes shape later choices, and agents use sustained environmental interaction to guide subsequent steps.Prior examples span software engineering, web navigation, workplace assistance, code generation, and tool-assisted reasoning.Source: §2.1 · pp. 2–3

2.2 LARGE LANGUAGE MODELS FOR MULTI-STEP RETROSYNTHESIS

  • Prior LLM retrosynthesis methods use external search or single-step models for route exploration and precursor generation, or generate a complete route in one open-loop turn.Examples include evolutionary route search, AND–OR tree search, and LLM selection among externally proposed precursors.Source: §2.2 · p. 3
  • Solvent instead performs global planning, stepwise decomposition, inventory checking, and feedback-guided revision across multiple turns.Source: §2.2 · p. 3

3 PRELIMINARIES

The preliminaries represent a multi-step retrosynthesis solution as a reaction tree and distinguish stock completion from reaction plausibility.

Multi-step retrosynthesis

  • A target is represented by a rooted, acyclic reaction tree whose internal molecules are decomposed into precursors; stock-complete solutions are further assessed for reaction plausibility.Source: §3 · p. 3

4 METHODOLOGY

Solvent treats multi-step retrosynthesis as iterative route planning with inventory feedback. Its training pipeline builds a seed policy, bootstraps a stronger initialization, and applies further RL to improve long-horizon reasoning.

4 METHODOLOGY

  • Solvent plans routes over multiple turns, checks precursor availability against a building-block inventory, and revises routes in response to inventory feedback.The two-round pipeline develops a seed policy, bootstraps a stronger starting policy, and then applies RL.Source: §4 · p. 3

4.1 DEVELOPING A SEED POLICY FOR LONG-HORIZON REASONING

  • A teacher distills target-level strategies and reaction-step analyses from reference routes, then teacher-generated multi-turn trajectories are filtered for valid solutions, consistent inventory interactions, and chemical validity.The curated trajectories fine-tune the warmed-up policy to establish iterative route construction and revision.Source: §4.1 · pp. 3–4
  • GSPO consolidates successful student behaviors into stable reasoning patterns after teacher-guided fine-tuning, which can otherwise produce truncated responses or malformed inventory queries.The seed-stage targets are structurally challenging, diverse, and screened for overlap with evaluation benchmarks.Source: §4.1 · p. 4

4.2 BOOTSTRAPPING AND BOOSTING LONG-HORIZON REASONING

  • Rejection-sampling fine-tuning bootstraps a stronger starting policy from trajectories sampled from the seed policy.The process selects trajectories that are stock-complete, inventory-consistent, chemically valid, and diverse in reasoning quality and style.Source: §4.2 · p. 5, §4.2.1 · p. 5

4.2.1 REASONING BOOTSTRAPPING

  • The first-epoch bootstrap checkpoint improves retrosynthesis performance over the seed policy while avoiding the sharp entropy decline seen with further training.It is selected as an exploration–exploitation trade-off and initializes final-stage RL.Source: §4.2.1 · p. 5

4.2.2 BOOSTING LONG-HORIZON REASONING WITH RL

  • Final-stage RL scores complete trajectories so route construction, inventory interaction, and feedback-guided revision receive credit when they jointly produce a successful route.The final-stage target set is drawn from structurally diverse USPTO-full targets with benchmark overlaps removed.Source: §4.2.2 · p. 5

5 EXPERIMENTS

Solvent is evaluated on four multi-step retrosynthesis benchmarks and compared with LLMs and search-based systems. Its results are competitive with frontier LLMs, lead prior search-based systems across budgets, and improve substantially through successive post-training stages.

5.2 MAIN RESULTS

  • 61.86% average success across the full test sets makes 8B Solvent competitive with frontier LLMs and substantially stronger than similarly sized open-weight models.The comparison covers four multi-step retrosynthesis benchmarks.Source: Table 1 · p. 6, §5.2 · p. 7

5.1 EXPERIMENTAL SETUP

  • Evaluation uses four widely adopted benchmarks and eMolecules as the inventory for commercial availability, with success requiring a valid route tree whose leaves are in stock.The benchmarks are USPTO-190, ChEMBL-1000, Pistachio Reachable, and Pistachio Hard.Source: §5.1 · p. 6
  • The LLM comparisons include frontier, mid-scale, and 8B retrosynthesis-specialized models, while search-system comparisons cover all four benchmarks.LLM comparisons use a common 32k-token output limit and up to five interaction turns; search comparisons use aligned budgets.Source: §5.1 · p. 6

5.2 MAIN RESULTS

  • Solvent achieves the highest success rates across all four benchmarks throughout the search-budget range against prior search-based systems.The comparison uses aligned search budgets.Source: §5.1 · p. 6, Table 2 · p. 7, §5.2 · p. 8

5.3 ABLATION STUDY

  • Average success rises from 0.18% to 61.86% across post-training stages, with final-stage RL producing the largest gain.After Round 1 RL, success reaches 22.82%; bootstrapping raises it to 29.19%, and final-stage RL raises it to 61.86%.Source: Table 3 · p. 8, §5.3 · p. 8
  • Teacher-guided cold start raises average success from 0.18% to 6.26%, and Round 1 RL raises it to 22.82% while stabilizing response patterns.Truncated responses and malformed inventory queries remain frequent after cold start.Source: §5.3 · p. 8

6 ANALYSIS

Across post-training, Solvent increasingly uses environmental feedback to revise route proposals. Final-stage RL improves chemical validity and stock completion, with validity gains preceding stock-completion gains.

6.1 THE CAPACITY FOR ADAPTIVE REVISION EMERGES ACROSS POST-TRAINING

  • As post-training progresses, Solvent learns to improve proposals through environmental feedback, with success gains from additional turns reported across stages.Figure 3 evaluates turn limits of two to five and reports the ordinary least-squares success-gain slope in percentage points per turn.Source: §6.1 · p. 8, Figure 3 · p. 9

6.2 HOW FINAL-STAGE RL IMPROVES LONG-HORIZON ROUTE CONSTRUCTION

  • Final-stage RL progressively improves chemical validity and stock completion, and validation success rises throughout training.Chemical validity improves before stock completion, while mean response length remains stable; after RL, Solvent exhibits strategic planning such as protecting-group use and convergent route design.Source: §6.2 · pp. 8–9, Figure 4 · p. 9

7 LIMITATIONS AND FUTURE DIRECTIONS

In silico route verification cannot establish experimental success or practical laboratory outcomes. The paper identifies experimental evaluation as a future direction for grounding planning and revision in measured results.

7 LIMITATIONS AND FUTURE DIRECTIONS

  • In silico checks proxy reaction feasibility but cannot establish experimental success, yield, selectivity, reproducibility, or operational safety.The authors suggest evaluating routes and conditions on automated wet-lab platforms against measured outcomes and practical laboratory constraints.Source: §7 · p. 9

8 CONCLUSION

Solvent frames multi-step retrosynthesis as long-horizon reasoning, iteratively planning routes, checking precursor availability, and revising decisions from inventory feedback. Across four benchmarks it rivals frontier LLMs and consistently outperforms prior search-based systems.

8 CONCLUSION

  • Solvent is an 8B LLM that iteratively plans routes, checks precursor availability, and revises decisions using inventory feedback.A two-round post-training curriculum develops this capability.Source: §8 · p. 10
  • Across four benchmarks, Solvent rivals frontier LLMs and consistently outperforms prior search-based retrosynthesis systems.Source: §8 · p. 10
  • Case studies show strategic chemical reasoning, including protecting-group schemes and convergent routes, and the authors describe the approach as a path to further scaling.Source: §8 · p. 10

AI USE STATEMENT

The authors report using generative AI for synthetic training-data generation, coding assistance, and manuscript polishing, while excluding it from specified research and analysis tasks.

AI USE STATEMENT

  • Generative AI was used to generate synthetic training data, assist with coding, and polish the manuscript.The authors state they reviewed all AI-assisted work.Source: AI USE STATEMENT · p. 11
  • The authors report no AI use in developing theoretical models, formulating mathematical claims or hypotheses, designing or evaluating methodology, or interpreting results.They also exclude dataset cleaning or reformatting and qualitative or thematic analysis; mathematical proofs were not applicable.Source: AI USE STATEMENT · p. 11
  • The authors accept responsibility for the final work, including AI-assisted text, claims, and artifacts.Source: AI USE STATEMENT · p. 11

ETHICS STATEMENT

The passages describe the study’s ethics declarations, prior work on retrosynthesis and reasoning, and methods for agentic learning and environment interaction.

A ADDITIONAL RELATED WORK

B AGENTIC REINFORCEMENT LEARNING

C.1 BENCHMARKS AND EVALUATION

The evaluation uses four retrosynthesis benchmarks, checks structural separation from training data, and applies fixed inventory and chemical-validity criteria. Solvent’s target-matched comparison remains competitive with frontier LLMs and substantially exceeds the same-sized R3.

C.1 BENCHMARKS AND EVALUATION

  • The four benchmarks cover patent-derived routes and bioactive targets; Pistachio Reachable emphasizes shorter tractable routes, while Pistachio Hard emphasizes longer, deeper plans.ChEMBL-1000 has no reference routes, so reaction-count and depth statistics are unavailable for it.Source: §C.1 · p. 17, Table 4 · p. 18
  • Training targets are decontaminated against all four evaluation benchmarks, and structural separation is assessed with Morgan-fingerprint similarity and Murcko-scaffold matches.Targets with empty acyclic scaffolds are excluded from the scaffold-novel calculation.Source: §C.1 · p. 18
  • 5.28% of targets have a training neighbor at similarity ≥0.8; on the novel-scaffold subset, Solvent’s benchmark-macro success rate is 60.26%.81.49% of scaffold-bearing targets lack a Murcko-scaffold match, and success is 55.37% among 410 targets below 0.4 maximum similarity.Source: §C.1 · p. 18
  • Route success requires a nontrivial target-rooted decomposition, unique parent for each non-root molecule, inventory-complete leaves, and both reaction-support and atom-source checks at every step.The reaction-support check uses ASKCOS-template substructures while permitting new substrate combinations.Source: §C.1 · p. 19
  • Experimental feasibility is not established by the computational plausibility checks; yield, selectivity, and competing reactions remain open proxy challenges.Source: §C.1 · p. 19

C.2 EVALUATION CONFIGURATIONS AND SEARCH BUDGETS

  • The LLM comparison uses target-only prompts, a shared inventory interface and scoring criteria, one generation per target, up to five turns, and a 32,000-token response limit.The cost-controlled comparisons use a shared, frozen subset stratified by structure and complexity.Source: §C.2 · p. 19
  • 63.26% macro average on the shared 400 targets places Solvent within 0.31 percentage points of GPT-5.5, 0.23 above Kimi K3, and 43.08 above R3.The target-matched results preserve the reported conclusion that Solvent is competitive with frontier LLMs and stronger than mid-scale alternatives.Source: Table 6 · p. 20, §C.2 · p. 20
  • Solvent’s search budget counts reaction steps across submitted routes, requiring a successful route by cumulative cost N for a target to count as solved.This provides a common search-effort measure across methods with different operation-counting schemes.Source: §C.2 · p. 21

D POST-TRAINING DATA CONSTRUCTION

This section describes how Solvent’s post-training data and target curricula were curated for long-horizon retrosynthesis and used in staged SFT and RL. It covers route-grounded retrosynthetic-intuition warm-up and teacher-generated tool-interaction trajectories, using target-only prompts and excluding evaluation-benchmark overlaps; the warm-up improved subsequent reasoning-bootstrapping results.

D.1 BUILDING RETROSYNTHETIC INTUITION

  • Warm-up annotations pair target-only structural analysis with chemically grounded explanations for each route transformation, preserving the source route while excluding disconnection information from target analysis.Annotations were retained only when they contained one valid JSON answer matching the source route and an ordered explanation for every transformation.Source: §D.1 · pp. 22, 25
  • 102,212 validated annotations from 104,867 deduplicated target–route pairs were retained for warm-up fine-tuning, with route-guided annotations generated by Intern-S1 from R3 data.The route informed annotation generation but was not included in the fine-tuning input; validation required a matching JSON route and complete structured explanations, and removed hidden reasoning.Source: §D.1 · pp. 22, 25

D.2 TEACHER-GUIDED COLD START FOR LONG-HORIZON REASONING

  • 2,200 structurally challenging USPTO-full targets (Dai et al., 2019), selected by molecular features and scaffold diversity without evaluation overlap, seeded the teacher-guided cold start.DeepSeek-V4-Pro (DeepSeek-AI, 2026) generated eight independent tool-interaction trajectories per target.Source: §D.2 · p. 25
  • Each cold-start trajectory allowed at most three inventory-query turns plus one finalization turn, retaining opportunities for route revision within a 32k-token generation budget.The same target-only prompt was used for rollout generation and fine-tuning, with the full teacher reasoning, queries, tool observations, and final answer as the autoregressive target.Source: §D.2 · p. 26
  • Teacher-guided cold-start curation required chemically plausible routes, full-route stock verification, and revision after failed checks, yielding 700 trajectories for fine-tuning πwarmed.Surviving candidates were ranked by chemistry tier, plausibility, complexity, trace length, route depth, and diversity, with at most six trajectories per Murcko scaffold.Source: §D.2 · pp. 26–27

D.3 TARGET SELECTION FOR FIRST-ROUND RL

  • The first-round RL curriculum contains 10,000 evaluation-deduplicated USPTO-full targets, stratified by structural complexity to cover increasingly demanding route-construction problems.Selection increases representation of complex ring systems, heteroaromatics, stereocenters, flexible linkers, and densely functionalized molecules while controlling scaffold concentration.Source: §D.3 · p. 27

D.4 CURATING TRAJECTORIES FOR REASONING BOOTSTRAPPING

  • From 400,000 bootstrapped candidates, route, interaction, and reaction filters plus reasoning-quality ranking retained one trajectory per target, yielding 9,550 trajectories after decontamination.Reasoning and style scores assess elaboration, self-verification, revision, granularity, balanced coverage, uncertainty, and tool use.Source: §D.4 · pp. 27–28

D.5 TARGET CURRICULUM FOR FINAL-STAGE RL

  • The final-stage RL curriculum replaces 2,909 lower-priority examples with persistent hard cases whose bootstrapping trajectories had 0% success, preserving a 10,000-target curriculum.Targets are chosen from inventory-absent USPTO-full products using failure profile, molecular complexity, structural attributes, and scaffold diversity, then decontaminated against four benchmarks.Source: §D.5 · p. 28

E.1 WARM-UP FOR RETROSYNTHETIC INTUITION; E.2 TEACHER-GUIDED COLD START

  • The retrosynthesis-pretrained Qwen3-8B backbone was warmed with 102,212 curated annotations, then πwarmed was fine-tuned on 700 teacher trajectories for long-horizon reasoning.The warm-up reserved 1,000 examples for validation; the cold-start checkpoint with the lowest validation loss initialized first-round RL.Source: §E.1 · p. 28, E TRAINING RECIPE · p. 28, §E.2 · p. 29

E.3 FIRST-ROUND RL FOR STABLE LONG-HORIZON INTERACTION

  • First-round GSPO trained on the 10,000-target curriculum with rewards weighted 0.6 for route success, 0.2 for inventory verification, and 0.2 for interaction validity, producing πseed.Dynamic group filtering, excluding truncated trajectories, and overlong-response penalties were used to improve training stability.Source: §E.3 · p. 29

E.4 REASONING BOOTSTRAPPING; E.5 FINAL-STAGE RL

  • Reasoning bootstrapping SFT on 9,550 trajectories selected the first-epoch πboot, which final-stage GSPO optimized on the 10,000-target curriculum using gated rewards and dynamic sampling.Final-stage RL excluded truncated trajectories, penalized overlong responses, and applied a KL penalty with coefficient 0.01 to limit policy drift.Source: §E.4 · p. 30, §E.5 · p. 30

F.1.1 RETROSYNTHETIC-INTUITION WARM-UP

  • Warm-up improved macro Pass@1 and Pass@8 at all three training depths and lowered early losses, though the runs converged after two epochs.The optimization advantage translated into stronger route-solving performance across the four benchmarks.Source: §F.1.1 · p. 31, Figure 5 · p. 31, Table 15 · p. 32

F.1.2 TRAINING DEPTH FOR REASONING BOOTSTRAPPING

This section examines how bootstrapping and related training and inference choices affect route success, quality, and cost, alongside case studies of inventory-guided route revision and long-horizon behavior before and after RL.

F.1.2 TRAINING DEPTH FOR REASONING BOOTSTRAPPING

  • One-epoch bootstrapping retains more exploratory capacity and ultimately outperforms three epochs after RL (53.40% vs. 47.43%), despite lower pre-RL success.Before RL, macro Pass@1 rises from 26.42% to 32.94% and Pass@8 from 69.62% to 75.26% with deeper training, while held-out entropy and rollout NLL decline; the one-epoch policy is selected for final-stage RL.Source: §F.1.2 · pp. 32–33, Table 16 · p. 32

F.1.3 THE LARGE FINAL-STAGE RL GAIN DEPENDS ON REASONING BOOTSTRAPPING

  • 45.69% validation success with reasoning bootstrapping compares with 9.31% when the same RL procedure starts from Round 1 SFT.The comparison holds the target curriculum, reward, optimizer, rollout configuration, and validation protocol fixed.Source: §F.1.3 · p. 33, Figure 7 · p. 33

F.1.4 THE ROLE OF DYNAMIC SAMPLING IN TRAINING SOLVENT

  • 53.40% validation success with dynamic sampling exceeds 48.26% without it, a gain of 5.14 percentage points.Groups with uniformly failed or uniformly successful trajectories are excluded, retaining mixed outcomes for updates.Source: §F.1.4 · pp. 33–34, Figure 8 · p. 34

F.1.5 ADAPTIVE REVISION FROM AVAILABILITY FEEDBACK

  • True precursor-availability feedback improves success by 29.2 points over shuffled feedback and 50.6 points over masked feedback.The paired study evaluates one trajectory for each of 500 held-out targets; the 95% paired bootstrap confidence intervals are 24.4–34.2 and 45.8–55.4 points, respectively.Source: §F.1.5 · p. 34, Table 17 · p. 34

F.2.1 SELECTING THE NUMBER OF INTERACTION TURNS ON HELD-OUT DATA; F.2.2 SELECTING THE SAMPLING TEMPERATURE ON HELD-OUT DATA

  • 68.00% held-out success at five turns supports the prespecified limit, while temperature 0.8 yields the highest success and improves paired success by 5.0 points over temperature 1.0.Six- and seven-turn limits add 1.2 and 2.6 points, respectively, but their paired confidence intervals include zero; temperature 0.8 is fixed with top-p 1.0 for the main results.Source: §F.2.1 · p. 35, Table 18 · p. 35, §F.2.2 · p. 36

F.2.3 INFERENCE COST

  • 61.86% average success for Solvent is comparable to Kimi K3’s 63.03%, with 8B versus 2.8T parameters and 9.20k versus 18.45k output tokens per trajectory.The cost comparison reports macro-averaged success and usage.Source: §F.2.3 · p. 36, Table 20 · p. 37

F.2.4 COMPLEMENTARY IN SILICO ASSESSMENTS OF ROUTE QUALITY

  • Solvent’s routes show a 0.827 target-to-leaf SA-score decrease and 84.69% SA-reducing steps, while all-step forward round-trip consistency is 9.02%.ChemCensor results are comparable to several LLMs, and independent diagnostics provide complementary evidence of route plausibility rather than establishing experimental feasibility.Source: §F.2.4 · p. 37, Table 21 · p. 38

F.2.5 BEHAVIORAL DYNAMICS OF THE TWO RL STAGES; G.1.1 INVENTORY-GUIDED CONVERGENT ROUTE PLANNING

  • Over Round 1 RL’s first 49 updates, five-update averages showed reward rising from −0.255 to 0.569, response length falling from 22.9k to 6.1k tokens, and truncation from 43.2% to 0.12%; final-stage mean rollout reward rose from 0.274 to 0.484 across the first and last 25 updates while policy entropy was maintained.Source: §F.2.5 · p. 38

G.1.1 INVENTORY-GUIDED CONVERGENT ROUTE PLANNING

  • Inventory feedback expands both aryl branches after rejecting an unavailable isocyanate route, yielding a stock-complete convergent plan with the anilines joined by a final urea formation.Manual review found all four transformations passed ChemCensor’s reaction-center criterion, three passed its functional-group criterion, and Chemformer recovered every product within its top-5 forward predictions.Source: §G.1.1 · pp. 41–42

G.1.2 BOC PROTECTION ENABLES CHEMOSELECTIVE PIPERAZINE FUNCTIONALIZATION

  • An unavailable 1-(4-methoxyphenyl)piperazine prompts Solvent to build it from N-Boc-piperazine and 4-bromoanisole, then remove Boc before coupling the remaining nitrogen to activated maleimide.The revised route’s three terminal precursors are available, and the second inventory query verifies the complete route; the primary chemical-validity assessment accepts all three transformations.Source: §G.1.2 · pp. 42–43

G.2 SOLVENT GENERATES DIVERSE PLAUSIBLE ROUTES FOR THE SAME TARGET

  • Solvent generates distinct stock-complete synthesis plans for the same target, using different chemistry to form the benzylic C–N bond.One route uses benzyl-bromide substitution; the other uses aldehyde reductive amination, with different aromatic precursors.Source: §G.2 · p. 43, §G.2.2 · p. 45

G.2.1 ROUTE A: BENZYL SUBSTITUTION

  • After an aldehyde-based proposal fails inventory checking, Route A switches from reductive amination to benzyl-bromide substitution, and inventory confirms all three terminal precursors are available.The revised route first forms the aryl amide, then couples the resulting benzyl bromide with 2-(pyridin-2-yl)ethan-1-amine.Source: §G.2.1 · p. 44

G.2.2 ROUTE B: REDUCTIVE AMINATION

  • Route B retains the aldehyde disconnection but replaces an unavailable acid with an available methyl ester, using aminolysis followed by reductive amination; inventory confirms all three terminal precursors are available.This provides a route to the same target through different chemistry from Route A.Source: §G.2.2 · pp. 44–45

G.3 REPRESENTATIVE CASES OF LONG-HORIZON REASONING ACROSS RL STAGES

  • The representative cases compare policies before and after Round 1 RL and final-stage RL on the same targets to illustrate qualitative changes in Solvent’s long-horizon behavior.The examples describe a shift from unstable teacher-guided traces to concise, reliable interaction and a post-RL route incorporating a protection-group strategy.Source: §G.3 · p. 45

G.3.1 FROM UNSTABLE IMITATION TO CONCISE, RELIABLE REASONING

Round 1 RL improves rollout reliability and produces shorter, tool-valid route planning. Later examples show a protection-group strategy incorporated into a chemically valid route, while the discussion distinguishes computational checks from experimental success.

G.3.1 FROM UNSTABLE IMITATION TO CONCISE, RELIABLE REASONING

  • 78.27% to 0.79% truncation, 71.32% to 39.69% malformed or incomplete interactions, and 12.93% to 70.91% final-answer parse rate after Round 1 RL under matched decoding.The comparison covers 12,896 paired trajectories.Source: Table 24 · p. 45, §G.3.1 · p. 45
  • On the same target, the SFT policy uses 31,964 tokens and three unsuccessful inventory queries without a final answer, while πseed returns a stock-complete route after 2,856 tokens and one query.Source: §G.3.1 · p. 45
  • πseed submits two amide disconnections for inventory verification and returns the route after availability is confirmed, without revisiting its plan.Source: §G.3.1 · p. 47

H FURTHER DISCUSSION

  • The evaluator screens evident chemical inconsistencies but cannot establish experimental success; the authors propose integrating experimental observations into later route decisions.Laboratory outcomes and practical route value depend on factors that cannot be fully assessed in silico and require experimental validation.Source: H FURTHER DISCUSSION · p. 48

G.3.2 A PROTECTION-GROUP STRATEGY AFTER FINAL-STAGE RL

  • Before final-stage RL, the route is stock-complete but chemically invalid; afterward, it carries Boc protection through selective bond formation and removes it in the final step.The earlier route’s initial methylation fails to account for the exocyclic nitrogen.Source: §G.3.2 · pp. 47–48, Figure 18 · p. 49

Related papers

Loading 2610.32308v1…