RetroPlanner: Multi-Route Expansion and Multi-Objective Reasoning for LLM-Based Retrosynthesis Planning

• Retrosynthesis planning searches for a sequence of reactions connecting a target molecule to starting materials available as purchasable building blocks.Source: §1 · p. 1
• Existing LLM planners expand one frontier molecule per turn, making it difficult to compare competing routes under a limited search budget.The relevant route-quality signals include reaction plausibility, round-trip consistency, and starting-material price.Source: §1 · p. 2
• In a 60-target USPTO-190 subset, 78.6% of the 42 targets solved by RetroAgent had a dominating route in the search space, and 64.3% had one already exposed but unexpanded.The cited routes dominate on reaction plausibility, round-trip consistency, and starting-material price.Source: §1 · p. 2
• RETROPLANNER jointly expands multiple frontier molecules and compares route candidates in one multi-objective decision to recover alternatives across the three synthesizability objectives.The introduction identifies multi-route expansion, evidence-grounded multi-objective reasoning, and improved discovery of alternative synthesizable routes as its contributions.Source: Figure 1 · p. 1, §1 · p. 2

2.1 PROBLEM FORMULATION

• The search state is a directed acyclic AND-OR graph in which molecules are shared across routes, and unsolved frontier molecules remain eligible for later expansion.Opened molecules retain their candidates, allowing the planner to revisit them and expand several at once.Source: §2.1 · pp. 2–3

2.2 OVERVIEW OF RETROPLANNER

• Each turn opens multiple frontier molecules, ranks their reaction candidates across plausibility, round-trip consistency, and starting-material price, then commits routes judged complete.Candidate comparisons also use auxiliary signals such as single-step confidence and molecular structural descriptors; the setup uses K = 10 and L = 3.Source: §2.2 · pp. 3–4

2.3 REASONING SCHEMA

• Reasoning traces describe structural progress, bond changes and reaction classes, then compare candidates with explored alternatives and identify the objective that orders them.Single-step confidence is an auxiliary signal when leading candidates are nearly tied on the primary objectives.Source: §2.3 · pp. 4–5

2.4 TRAINING

• Training traces are generated by replaying Pareto-ranked route graphs and are verified against the search graph before verified reasoning is used for supervised fine-tuning.Verification checks numerical values, the existence and stated differences of comparison routes, and whether the cited objective separates the candidate from competitors.Source: §2.4 · p. 5

3.1 EXPERIMENTAL SETUP

• Baselines include conventional search algorithms and trained LLM planners; RETROPLANNER is initialized from GPT-OSS-20B and supervised fine-tuned to learn multi-route expansion and multi-objective reasoning.Conventional search baselines use an equal-weight scalarization of the three route-quality objectives; MORetro* retains its Bayesian-optimization sampling procedure.Source: §3.1 · pp. 5–6
• Evaluation uses NB = 50 on in-distribution USPTO-190 and out-of-distribution ChEMBL-1000; route quality selects each method’s composite-best route across plausibility, round-trip consistency, and price.Success Rate measures targets reaching purchasable building blocks within budget, while Routes counts routes per solved target.Source: §3.1 · p. 6, Q1. Does the planner efficiently… · p. 6

Q1. Does the planner efficiently recover diverse, high-quality routes under a constrained search budget?

• RETROPLANNER’s success rates are 4.73 and 5.90 percentage points lower than RetroAgent’s on USPTO-190 and ChEMBL-1000, respectively, but it returns substantially more alternatives with favorable objective trade-offs.The comparison uses an expansion budget of NB = 50.Source: Q1. Does the planner efficiently… · p. 6, Table 1 · p. 7
• On USPTO-190, RETROPLANNER achieves the highest Better-Than-Ref and improves all three route-quality metrics over MORetro* on both benchmarks; RetroAgent has substantially lower round-trip consistency.ChEMBL-1000 has no reference routes, so Better-Than-Ref is unavailable there.Source: Q1. Does the planner efficiently… · p. 6, Table 1 · p. 7
• Table 2 isolates expansion strategy with a shared GPT-OSS-20B backbone; its daggered first-cut disagreement is computed only for targets with at least two routes, so 1.00 does not indicate route coverage.The ablation uses an extended budget of NB = 300.Source: Table 2 · p. 7

4 ABLATION AND ANALYSIS

• The ablations vary frontier selection, the number of molecules expanded per turn, parallel versus sequential expansion, reasoning evidence, and single-step or forward-model choice.Expansion-strategy variants use an extended search budget of NB = 300; the reasoning-schema variants share action formats and differ in their reasoning traces.Source: §4 · p. 7, Q2. How does the expansion strategy… · pp. 7–8, Q3. Which components of the reasoning… · p. 8

Q2. How does the expansion strategy shape route-set coverage and search efficiency under extended search?

• 3.88 distinct first cuts for Multi-Route (Sequential) versus 2.23 for Single-Route; Parallel recovers fewer routes but achieves the highest success rate and Better-Than-Ref.Parallel expansion allocates budget toward alternatives that differ from routes already in the graph rather than repeating the same first cut.Source: Q2. How does the expansion strategy… · p. 8

4.2 REASONING SCHEMA

• 13.0 s per molecule for Sequential versus 10.6 s for Parallel, despite opening 30.4 versus 29.6 molecules per rollout; parallel expansion also reduces per-molecule turns and output tokens.Parallel uses 1.71 turns and 80.2 output tokens per molecule, compared with 2.30 turns and 106.9 tokens for Sequential.Source: Figure 4 · p. 8, §4.2 · p. 8

4.3 SINGLE-STEP AND ROUND-TRIP MODEL VARIATIONS

• 53.16% Better-Than-Ref for RETROPLANNER versus 43.16% for the strongest baseline with LocalRetro; replacing ReactionT5v2 with RXNGraphormer preserves its advantage on BTR and RT.The LocalRetro test uses USPTO-190 with NB = 50; the RXNGraphormer evaluation reports changes of only 2.63 and 2.52 percentage points on the two metrics.Source: Table 4 · p. 9, §4.3 · p. 9, Table 5 · p. 9

Q3. Which components of the reasoning schema contribute to route diversity and quality?

• 0.769 BTR for MOLECULE (L1), up from 0.720; disconnection and retained-candidate evidence mainly improve diversity, while L3 combines route coverage with distinct first cuts.Molecule-level evidence supports breadth and quality; disconnection-level evidence mainly reduces routes sharing the same first cut.Source: Q3. Which components of the reasoning… · p. 9

5 LIMITATIONS AND CONCLUSION

• Across in-distribution and out-of-distribution benchmarks, RETROPLANNER recovers broader route sets while maintaining high success and improving the quality of retained alternatives.Ablations attribute contributions to both multi-route expansion and the reasoning schema, and the learned policy is reported robust to the choice of single-step and round-trip models.Source: §5 · p. 9
• The evaluation operationalizes synthesizability through reaction plausibility, round-trip consistency, and starting-material price rather than capturing experimental feasibility in full.Future work could optimize these route-level objectives through reinforcement learning and include laboratory factors such as reaction conditions, yield, and safety.Source: §5 · p. 9
• The authors report using generative AI for writing, retrieval, ideation, implementation support, and synthetic training data, with trajectory verification and publicly available data, models, and scoring tools.They state that additional implementation and evaluation details and code are in supplementary material.Source: REPRODUCIBILITY STATEMENT · p. 10
• The proposed routes are computational suggestions and require expert review and experimental validation before laboratory use.Source: ETHICS STATEMENT · p. 10

A.1 RELATED WORK

• Conventional computer-aided synthesis planning combines single-step reaction proposals with multi-step search, commonly using Monte Carlo Tree Search or A*-like methods.These approaches primarily seek at least one route to purchasable building blocks.Source: §A.1 · p. 13
• Multi-objective methods recover non-dominated routes through Pareto search or weight sampling, while the paper frames its approach as reasoning over the current search state to select trade-offs (Hastedt et al., 2026).The cited methods typically specify objectives in advance and often assume route objectives decompose additively over reactions.Source: §A.1 · p. 13
• Prior LLM agents orchestrate retrosynthesis tools through sequential decisions but expand one frontier molecule per turn; this work targets coordinated exploration across multiple frontier states (Zhu et al., 2026).Source: §A.1 · p. 13

A.2 SINGLE-STEP MODELS AND CHEMISTRY TOOLS

• R-SMILES generates candidate reactant sets from product SMILES using a transformer and beam search, and serves as the planner’s template-free single-step model.Its aligned input and output sequences allow local edits rather than regenerating the molecule.Source: §A.2 · p. 13
• LocalRetro scores local reaction-center templates and supplies the top-ranked precursor sets as candidates for the planner.Source: §A.2 · pp. 13–14
• AiZynthfinder’s template expansion policy is repurposed as a reaction-plausibility filter, while ReactionT5v2 predicts products from reactants for round-trip consistency.Source: §A.2 · p. 14
• RXNGraphormer provides an alternative forward model for round-trip consistency, and MolPrice estimates terminal building-block prices as a synthetic-accessibility proxy.MolPrice predicts USD per mmol and is trained on purchasable compounds with contrastive price labels for hard-to-synthesize structures.Source: §A.2 · p. 14

• The training set contains 3,206 filtered PaRoutes targets, each searched once with Retro*-0 under a budget of 300 single-step calls.The targets have gold-route depths of 2–14 steps, with a median of 5 and mean of 4.89.Source: §A.3 · p. 14
• Retaining up to eight routes from the first two Pareto fronts broadens starting-material diversity without substantially relaxing route quality.Routes are selected with distinct starting-material sets; training uses top-1 round-trip recovery for Pareto ranking, stricter than the top-5 evaluation criterion.Source: §A.3 · p. 14
• Reaction evidence combines mapped bond changes, affected atoms and functional groups, reaction classes, template matching, and molecular descriptors.When atom mapping fails, the result is marked not measured rather than treated as no bond change.Source: §A.3 · p. 14
• Training uses 4 NVIDIA B200 GPUs with FSDP and bf16, while inference uses 8 NVIDIA H100 GPUs.Table 6 is identified as reporting training and decoding hyperparameters.Source: §A.3 · p. 14, Table 6 · p. 15

• Candidate reactions are scored for model confidence, plausibility, round-trip rank, and aggregate precursor price, with lower price indicating cheaper precursors.A reaction counts as plausible at p ≥ 0.05; round-trip consistency records the forward model’s recovery rank, and precursor price is normalized to [0, 1].Source: §A.4 · p. 15
• The route objectives average the fraction of plausible reactions and the fraction with round-trip rank at most 5, while aggregating terminal-molecule prices.Targets without a returned route score 0 on plausibility and round-trip consistency and receive a price cap of $1000.Source: §A.4 · p. 15
• Route price is capped and logarithmically normalized because starting-material prices are heavy-tailed, then reversed so higher values indicate better scores.The representative route is selected by equal-weight averaging of the three objectives.Source: §A.4 · pp. 15–16
• Better-Than-Ref counts targets for which a planner returns a route scoring better than the benchmark’s published route.An unsolved target or a route tied with the reference counts as a loss.Source: §A.4 · p. 16
• 1stCutDiff compares recovered routes by whether their initial disconnections yield different precursor sets, regardless of later steps.It is computed only for targets with at least two routes; 1stCuts separately counts distinct first-cut precursor sets.Source: §A.4 · p. 16

• At NB = 500, RETROPLANNER reaches 98.95% success on USPTO-190 and 96.50% on ChEMBL-1000, while recovering more higher-quality routes than prior LLM planners.RetroAgent attains the highest success rates overall, but RETROPLANNER remains close.Source: §A.5 · p. 16, Table 7 · p. 17
• Action-token attention to search-graph route rows is highest with the instance’s own analysis, and diverging-from tokens attend most to the route they reference.Meaningless text or another instance’s analysis raises attention only slightly.Source: §A.5 · p. 16, Figure 5 · p. 17
• In the illustrated five-step comparison, Pareto-optimal route A is plausible and round-trips at every step for $104, while dominated route B does so on 60% and 40% of steps at $233.The routes share their first disconnection, and different objectives separate leading candidates at different later steps.Source: §A.5 · p. 17, Figure 6 · p. 18

• The prompt material is headed “System / Developer / User.”Source: §A.6 · p. 18
• The embedded system text identifies ChatGPT, gives a 2024-06 knowledge cutoff and current date, and sets reasoning to medium.Source: §A.6 · p. 18
• The embedded instructions list analysis, commentary, and final as valid channels.Source: §A.6 · p. 18

Developer

• A molecule closes through any one reaction, while every piece under a reaction must reach purchasable material; completed routes are compared on three objectives without fixed priority.A route beats another only when it is at least equal on all three axes and better on one.Source: Developer · pp. 18–19

# Tools

• A single call can open several frontier molecules, returning their disconnections at the same depth for comparison on the next turn.At most one action applies per molecule in a call.Source: # Tools · p. 19

Turn 3 — tool

• Opening ba and d5 together keeps two uy-branch cuts alive at one depth so they can be ranked side by side.d5 is shared by two reactions, r4 and r5.Source: Turn 3 — tool · p. 20

Turn 4 — tool

• The d5 candidates illustrate trade-offs across the signals: c2 has p.933, rt1, and $.576, while c5 has p.999, rt-, and $.646.These candidate values describe different cuts, not a single route-level score.Source: Turn 4 — tool · p. 21

Turn 8 — assistant (analysis)

• Route reasoning compares structural changes and divergences; for example, the oxa-Michael cut closes the pyran and reads back to product, unlike the amide cuts.The competing amide cuts can tie on scored axes, while the single-step model's confidence separates them.Source: Turn 8 — assistant (analysis) · p. 22

DISCONNECTIONS of d5

• The selected d5 cut clears the plausibility filter at .999 and is the only untried cut here that reaches the free acid.The passage describes this as the plausibility-based choice among the available untried cuts.Source: DISCONNECTIONS OF d5 · p. 23