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× 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.
Figures & tables
Figure 1: At 8B parameters, Solvent substantially outperforms mid-scale LLMs and rivals frontier trillion-parameter models. Proprietary models with undisclosed parameter counts are arbitrarily positioned in the shaded region.
Figure 2: Overview of Solvent. Left: Two rounds of post-training prepare Solvent for effective RL. Round 1 develops a seed policy with retrosynthetic knowledge and stable reasoning. Round 2 bootstraps a stronger starting policy using the seed policy, enabling final-stage RL to substantially boost performance. Right: Solvent performs long-horizon reasoning through iterative planning, external verification, and adaptive revision.
Method
Params.
Pistachio Hard
Pistachio Reachable
USPTO-190
ChEMBL- 1000
Average (%)
Frontier LLMs (proprietary)
Claude Opus 4.8 †
—
38.57
77.50
54.44
55.63
56.54
Gemini 3.5 Flash †
—
42.86
72.50
52.22
65.63
58.30
Grok 4.5 †
—
52.86
71.25
50.00
62.50
59.15
GPT-5.5 †
—
58.57
82.50
54.44
58.75
63.57
Frontier LLMs (open-weight)
Table 1: At 8B parameters, Solvent remains competitive with frontier LLMs across four multi-step retrosynthesis benchmarks. We report parameter counts, benchmark-specific success rates, and their average. To enable cost-controlled, target-matched evaluation, selected models are evaluated on a shared 400-target subset and marked with † . Em dashes denote undisclosed parameter counts. Gray shading and boldface mark the best frontier and mid-scale results, respectively.
Performance on USPTO-190
Method
N=50
N=100
N=200
N=300
N=400
N=500
Greedy DFS
19.47
19.47
19.47
19.47
19.47
19.47
retro*
38.95
50.00
68.95
74.74
77.89
79.47
retro*-0
27.37
37.89
55.79
62.63
72.11
74.21
retro*+
55.79
68.42
79.47
83.16
84.21
84.21
retro*+-0
47.89
61.05
76.84
82.11
85.79
86.32
Table 2: Comparison with prior search-based retrosynthesis systems across search budgets. Here, N tracks how much explicit reaction-level construction each method performs before finding a successful route, with the cross-method accounting defined in Appendix C.2 . Bold and underlined values denote the best and second-best results, respectively.
Training stage
Pistachio Hard
Pistachio Reachable
USPTO- 190
ChEMBL- 1000
Average
Initial Backbone
0.00
0.00
0.53
0.20
0.18
Round 1
Teacher-guided cold start
2.00
12.67
3.16
7.20
6.26
First-round RL
19.00
40.00
9.47
22.80
22.82
Round 2
Reasoning bootstrapping
23.00
41.33
22.63
29.80
29.19
Final-stage RL
62.00
78.00
48.95
58.50
61.86
Table 3: Stage-wise performance across the two-round post-training curriculum. We report performance for four training stages described in Sec. 4.1 and Sec. 4.2 . Average success improves across the reported stages, with the largest gain occurring after final-stage RL.
Figure 3: The benefit of additional interaction turns increases throughout post-training. (a) Average success rate under turn limits of 2–5. (b) Success gain per additional turn across stages.
Figure 4: Final-stage RL successively improves chemical validity and stock completion. (a) Performance on validation set during final-stage RL. (b) Evolution of chemical validity and stock completion. (c) Mean response length during training.
Appendix figures & tables36 assets
Supplementary material from the paper’s appendix.
Appendix
Benchmark
Targets
Avg. reactions
Avg. depth
USPTO-190
190
6.7
6.0
ChEMBL-1000
1,000
–
–
Pistachio Reachable
150
5.5
5.4
Pistachio Hard
100
7.5
7.2
Appendix
Table 4: Summary of the multi-step retrosynthesis benchmarks. Reaction counts and depths refer to the provided reference routes; these statistics are unavailable for ChEMBL-1000.
Diagnostic
Value
Evaluation targets
1,440
Exact target overlap
0
Median maximum Morgan similarity
0.466
Targets with maximum similarity ≥0.8
5.28%
Targets with novel Murcko scaffolds
81.49%
Success on novel-scaffold targets
60.26%
Appendix
Table 5: Structural separation between training and evaluation targets. Similarities are computed against the union of retained targets from all training stages. Success uses the same fixed single-trajectory evaluation as Table 1 .
Method
Params.
Pistachio Hard
Pistachio Reachable
USPTO-190
ChEMBL- 1000
Average (%)
Frontier LLMs (proprietary)
Claude Opus 4.8
—
38.57
77.50
54.44
55.63
56.54
Gemini 3.5 Flash
—
42.86
72.50
52.22
65.63
58.30
Grok 4.5
—
52.86
71.25
50.00
62.50
59.15
GPT-5.5
—
58.57
82.50
54.44
58.75
63.57
Frontier LLMs (open-weight)
Appendix
Table 6: LLM performance on the shared 400-target subset. Every model is evaluated on the same frozen targets using the success criterion in Eq. 10 . We report benchmark-specific success rates and their unweighted macro average. Em dashes denote undisclosed parameter counts. Gray shading marks the best result among frontier LLMs, and boldface marks the best result among mid-scale LLMs.
Setting
Value
Annotation model
Intern-S1
Source target–route pairs
104,867
Generations per route
1
Temperature
0.7
Top- p / top- k
1.0 / 50
Reasoning mode
Enabled
Appendix
Table 7: Configuration used to generate retrosynthetic warm-up annotations with Intern-S1.
Setting
Value
Teacher model
DeepSeek-V4-Pro
Target source
USPTO-full
Rollout targets
2,200
Rollouts per target
8
Maximum inventory-query turns
3
Cumulative generation limit
32k tokens
Appendix
Table 8: Configuration used to generate and curate the teacher-guided cold-start trajectories.
Setting
Value
Generating policy
πseed
Target source
USPTO-full
Trajectory sampling
Best-of- N
Temperature / top- p
1.0 / 1.0
Maximum reasoning-and-action turns
5
Cumulative generation limit
32k tokens
Appendix
Table 9: Generation and curation configuration for reasoning bootstrapping.
Setting
Value
Initialization
R3 retrosynthesis-pretrained Qwen3-8B checkpoint
Fine-tuning method
Full-parameter SFT
Examples / validation split
102,212 / 1,000
Training epochs
2
Maximum sequence length
4,096 tokens
Global batch size
64
Appendix
Table 10: Training configuration for the retrosynthetic-intuition warm-up.
Setting
Value
Initialization
πwarmed
Fine-tuning method
Full-parameter SFT
Examples / validation split
700 / 50
Training epochs
2
Maximum sequence length
32,768 tokens
Global batch size
8
Appendix
Table 11: Training configuration for the teacher-guided cold start.
Setting
Value
Initialization
Teacher-guided cold-start policy
Optimization algorithm
GSPO
Training targets
10,000
Optimization updates
50
Rollout targets per batch
128
Rollouts per target
8
Appendix
Table 12: Training configuration for first-round RL.
Setting
Value
Initialization
πwarmed
Fine-tuning method
Full-parameter SFT
Examples / validation split
9,550 / 100
Training epochs
1
Maximum sequence length
32,768 tokens
Global batch size
8
Appendix
Table 13: Training configuration for reasoning bootstrapping.
Setting
Value
Initialization
πboot
Optimization algorithm
GSPO
Training targets
10,000
Optimization updates
780
Rollout targets per batch
64
Rollouts per target
8
Appendix
Table 14: Training configuration for final-stage RL.
Figure 5: Retrosynthetic-intuition warm-up improves the optimization of subsequent reasoning-bootstrapping SFT. Without and with intuition warm-up, training starts from the retrosynthesis-pretrained backbone and πwarmed , respectively. (a) Per-step training loss under the two matched initializations; pale traces show individual updates and solid curves show centered 61-update averages. (b) Validation loss throughout training.
Pass@1 (%)
Pass@8 (%)
SFT epochs
Without warm-up
With warm-up
Δ
Without warm-up
With warm-up
Δ
1
24.99
26.42
+1.43
65.24
69.62
+4.38
2
30.21
32.61
+2.40
70.92
73.10
+2.18
3
28.50
32.94
+4.44
74.70
75.26
+0.56
Appendix
Table 15: Retrosynthetic-intuition warm-up consistently improves success rates under matched reasoning-bootstrapping SFT. Results are macro averages across the four benchmarks.
Epoch
Eval loss
SFT entropy
Macro P@1 (%)
Macro P@8 (%)
Rollout NLL
Final parse (%)
1
0.2850
0.3217
26.42
69.62
0.3151
71.53
2
0.2851
0.2830
32.61
73.10
0.2488
81.32
3
0.3058
0.2586
32.94
75.26
0.1927
77.01
Appendix
Table 16: Validation loss, entropy, and task performance after different numbers of reasoning-bootstrapping epochs. Macro success is averaged across the four benchmarks. Bold marks the column maximum, except for loss, where lower is better. Final parse denotes the rate of parseable final answers.
Figure 6: One epoch of reasoning bootstrapping yields stronger downstream RL. (a) Validation loss and held-out entropy throughout three epochs of SFT. (b) Validation success during matched final-stage RL initialized from the one- and three-epoch policies. RL validation uses the default rollout decoding configuration ( T=1.0 , top- p=1.0 ; different from the benchmark decoding configuration).
Figure 7: Final-stage RL yields large capability gains only after reasoning bootstrapping. (a) Validation success. (b) Mean reward over training rollouts. (c) Mean training-rollout response length. Pale traces in (b) and (c) show individual updates, and solid curves show centered 11-update averages. Both runs use the default rollout decoding configuration ( T=1.0 , top- p=1.0 ); curves are shown through update 600.
Figure 8: Effect of dynamic sampling during final-stage RL. (a) Validation success with and without dynamic sampling. (b) Mean reward over the training rollouts retained for each policy update. Pale traces show individual updates, and solid curves show centered 21-update averages. Validation uses the default rollout decoding configuration ( T=1.0 , top- p=1.0 ).
Feedback
Success (%)
Stock-complete (%)
Chemical validity (%)
Mean response tokens
Mean inventory queries
True
68.20
80.40
73.60
7,020
2.40
Shuffled
39.00
43.80
46.60
9,158
3.27
Masked
17.60
19.20
47.20
7,891
3.63
Appendix
Table 17: Paired intervention on inventory feedback. Results are computed from one trajectory for each of 500 held-out targets. The success rate requires a structurally valid, stock-complete route that passes the chemical-validity assessment.
Performance
Inference behavior
Turn limit
Success rate (%)
Stock-complete routes (%)
Chemically valid routes (%)
Mean tool calls (all)
Mean tool calls (successful)
Mean response tokens
Truncated (%)
2
27.20
28.40
27.80
1.71
1.00
4,524
0.00
3
50.20
56.00
51.60
2.07
1.38
5,743
0.20
4
62.80
70.40
65.40
2.33
1.71
6,563
0.00
5
68.00
77.40
71.80
2.48
1.92
7,063
0.20
6
69.20
83.60
74.00
2.49
2.10
7,294
0.20
Appendix
Table 18: Held-out performance and inference behavior across turn limits. Blue shading marks the selected configuration, and bold marks the maximum in each performance column. Tool calls are averaged over all trajectories or successful trajectories only.
Performance (%)
Generation behavior
T
Success rate
Stock- complete
Chemical validity
Token NLL
Response tokens
Inventory queries
Truncated (%)
0.6
65.80
76.00
69.60
0.1084
7,055
2.52
0.80
0.7
66.80
77.20
71.40
0.1322
6,911
2.48
0.40
0.8
68.00
77.40
71.80
0.1638
7,063
2.48
0.20
0.9
65.00
76.80
70.80
0.2103
7,595
2.47
0.60
1.0
63.00
73.60
70.80
0.2847
7,958
2.43
0.20
Appendix
Table 19: Inference performance and generation behavior across sampling temperatures. Results are computed on the frozen 500-target validation split. Blue shading marks the selected configuration; bold marks the highest task metrics and the lowest truncation rate.
Method
Params.
Success (%)
Output tokens
Inventory queries
Frontier LLMs (proprietary)
Claude Opus 4.8 †
—
56.54
3.57k
2.73
Gemini 3.5 Flash †
—
58.30
20.33k
2.70
Grok 4.5 †
—
59.15
14.30k
1.72
GPT-5.5 †
—
63.57
5.32k
2.57
Frontier LLMs (open-weight)
Appendix
Table 20: Inference cost of the evaluated LLMs. Success and usage are macro-averaged; daggers mark the shared 400-target subset, and em dashes denote undisclosed parameter counts.
Synthetic accessibility
ChemCensor
Round trip
Method
Params.
Stock routes (/400)
Target–leaf Δ SA
Reducing steps (%)
RC routes /100 steps
FG routes /100 steps
Route RT@20 (%)
Frontier LLMs (proprietary)
Claude Opus 4.8
—
304
0.611
70.10
8.35
3.49
14.42
Gemini 3.5 Flash
—
300
0.607
68.60
9.89
6.02
15.75
Grok 4.5
—
295
0.618
63.89
8.48
4.48
13.06
GPT-5.5
—
289
0.539
66.57
9.98
5.33
10.90
Appendix
Table 21: Complementary in silico assessments on the shared 400-target subset. Stock routes are stock-complete. Δ SA is the decrease from target SA to mean leaf SA; reducing steps have a lower mean precursor SA than product SA. ChemCensor columns report all-step-supported stock-complete routes per 100 evaluated reaction steps, macro-averaged across benchmarks. Route RT is the percentage of all 400 targets whose routes pass exact top-20 forward recovery at every step.
Orthogonal diagnostic
Primary pass (%)
Primary non-pass (%)
Difference (95% CI)
ChemCensor RC step support
59.72
45.92
+13.80 ( 6.65 , 21.15 )
ChemCensor FG step support
45.14
29.90
+15.24 ( 7.67 , 22.76 )
Exact route RT@20
14.98
0.00
+14.98 ( 10.67 , 19.50 )
Appendix
Table 22: Orthogonal assessments grouped by primary chemical validity. We split 293 stock-complete Solvent routes into routes that pass and routes that do not pass the primary chemical-validity assessment. Each row reports the mean score of an independent diagnostic in the two groups. Difference is the mean for primary-pass routes minus the mean for primary-non-pass routes; 95% confidence intervals use 10,000 benchmark-stratified bootstrap resamples.
Figure 9: Round 1 RL stabilizes long-horizon interaction. (a) Training reward, (b) mean response length, and (c) the fraction of truncated trajectories. Faint lines show per-update measurements and colored lines show centered seven-update moving averages. The dashed line marks the selected πseed ; the shaded region denotes the exploratory continuation beyond this checkpoint.
Figure 10: Reward and entropy during final-stage RL. (a) Mean rollout reward and (b) policy entropy. Faint lines show per-update measurements and colored lines show centered 25-update moving averages.
Figure 11: Target molecule for the convergent-route case study.
Query
Proposed route boundary
Available leaves
1
Aniline and aryl isocyanate
0/2
2
Two anilines and CDI
1/3
3
Both aniline branches expanded to simpler precursors
6/6
Appendix
Table 23: Inventory feedback during route revision.
Figure 12: Inventory-verified convergent route rendered in forward synthetic order.
Figure 13: Target molecule for the Boc-protection case study.
Figure 14: Inventory-verified protecting-group route rendered in forward synthetic order.
Figure 15: Target molecule shared across independently sampled trajectories.
Figure 16: Stock-complete route based on benzyl substitution, rendered in forward synthetic order.
Figure 17: Stock-complete route centered on reductive amination, rendered in forward synthetic order.
Policy
Truncated (%)
Parseable final answer (%)
Malformed or incomplete interaction (%)
Teacher-guided SFT
78.27
12.93
71.32
πseed
0.79
70.91
39.69
Appendix
Table 24: Rollout reliability before and after Round 1 RL under matched decoding. Malformed or incomplete interactions fail the tool-trace validity check.
Policy
Tokens
Steps
Stock- complete
Chemical validity
Round 2 bootstrapping
6,449
4
Yes
No
Final-stage RL
3,935
3
Yes
Yes
Appendix
Table 25: Protection-group strategy before and after final-stage RL.
Figure 18: Routes proposed before and after final-stage RL, with steps shown in forward synthetic order. (a) The Round 2 bootstrapping route is stock-complete but fails the chemical-validity check because its initial methylation does not account for the exocyclic nitrogen. (b) The final-stage RL route carries Boc protection through the selective bond formation and removes it only in the final step.
Multi-step retrosynthesis planning seeks to decompose a target molecule into commercially available building blocks through a sequence of feasible reactions. The vast combinatorial search space makes this task challenging even for expert chemists. Traditional methods combine tree search with offline-trained value networks that score candidates in isolation, without reasoning about complete multi-step routes. Recent work leverages Large Language Models (LLMs) for this task, but relies on simple interfaces that limit exploration of the full search space. We introduce RetroAgent, an LLM agent that bridges symbolic search and neural reasoning through a harness with structured memory. Through memory and chemistry tools, the agent observes the full search state, including explored routes, available alternatives, and properties of intermediates, enabling informed decisions grounded in both global progress and domain knowledge. Experiments on in-distribution and out-of-distribution benchmarks demonstrate that RetroAgent delivers strong performance and generalization.
Retrosynthesis prediction aims to identify reactants that can synthesize a given product molecule. Although molecular large language models (LLMs) have recently shown promising results, most existing methods either generate reactants directly or provide only generic product-level analysis, without explicitly reasoning about bond-disconnection strategies that justify specific reactant choices. This paper proposes RetroReasoner, a retrosynthetic reasoning model that captures chemists' strategic disconnection-based thinking. RetroReasoner is trained with supervised fine-tuning and reinforcement learning. For supervised fine-tuning, SyntheticRetro generates structured disconnection rationales paired with reactant predictions. For reinforcement learning, a round-trip reward evaluates predicted reactants by passing them through a forward synthesis model and rewarding predictions that reconstruct the original product. RetroReasoner can also be applied to multi-step retrosynthetic planning by incorporating it into a parallelized Monte Carlo tree search framework, reducing search time while increasing the number and diversity of valid synthetic pathways. Experimental results show that RetroReasoner outperforms prior baselines, including not only molecular LLMs but also retrosynthesis-specific expert models, and generates a broader range of feasible reactant proposals, especially for challenging reaction instances. The code is available at https://github.com/KU-AGI/RetroReasoner.
Hanbum Ko, Chanhui Lee, Ye Rin Kim +4
Department of Artificial Intelligence, Korea University · Materials Intelligence Lab, LG AI Research · Department of Statistics, Korea University
Single-step retrosynthesis is a central component of computer-aided synthesis planning, yet its intrinsically one-to-many nature is poorly captured by single-answer evaluation and benchmarking protocols. To address this, we introduce Top-K prompting as a robust training and inference paradigm to better capture diverse, plausible reaction predictions. We compile CREED-CCV-2+USPTO-XL, an ultra-large-scale dataset of ~45.6 million verified reactions to train the C3LM (Chemistry Constraint-Consistent Language Model). By integrating fine-tuning with ChemCensor-based and novelty-oriented rewards, our model achieves state-of-the-art performance on the OOD URSA-expert-2026 benchmark. Further analysis of reaction uniqueness shows that LLMs and conventional models explore complementary reaction spaces, motivating ensemble-based retrosynthesis systems. Overall, our results establish Top-K, plausibility-aware training as a practical new direction for robust future LLM-based synthesis planning.
Bogdan Zagribelnyy, Ivan Ilin, Nikita Bondarev +5
Insilico Medicine AI Limited, Abu Dhabi, UAE · Insilico Medicine Canada Inc., Montreal, Quebec, Canada · Insilico Medicine Hong Kong Ltd., Hong Kong SAR, China