Retrosynthetic planning is a fundamental task in organic chemistry, yet remains challenging due to its combinatorial complexity. To address this, conventional approaches typically rely on hybrid frameworks that combine single-step predictions with external search heuristics, inevitably fracturing the logical coherence between local molecular transformations and global planning objectives. To bridge this gap and embed sophisticated strategic foresight directly into the model's chemical reasoning, we introduce ReTriP, an end-to-end generative framework that reformulates retrosynthesis as a direct Chain-of-Thought reasoning task. We establish a path-coherent molecular representation and employ a progressive training curriculum that transitions from reasoning distillation to reinforcement learning with verifiable rewards, effectively aligning stepwise generation with practical route utility. Empirical evaluation on RetroBench demonstrates that ReTriP achieves state-of-the-art performance, exhibiting superior robustness in long-horizon planning compared to hybrid baselines.
Figure 2: The ReTriP framework. (a) Path-coherent data construction: (a1) The target molecule undergoes notational augmentation. (a2) Iterative alignment ensures fragment traceability. (b) Progressive training curriculum: (b1) A three-stage SFT process transitions from trajectory modeling to reasoning distillation and loss-rebalancing calibration. (b2) RLVR aligns stepwise logic with precursor availability and synthetic efficiency. (c) Inference and scaling: A ranking-guided TTA selects optimal input notations, followed by a consensus-based voting mechanism.
Generative Model
Search Module
Top-1
Top-2
Top-3
Top-4
Top-5
Template-based
Retrosim ( Coley et al., 2017 )
Retro*
35.1
40.5
42.9
44.0
44.6
Neuralsym ( Segler and Waller, 2017 )
Retro*-0
42.0
49.3
52.0
53.6
54.3
GLN ( Dai et al., 2019 )
Retro*
39.6
48.9
52.7
54.6
55.7
Semi-template-based
G2Gs ( Shi et al., 2020 )
Retro*
5.4
8.3
9.9
10.9
11.7
Table 1: Exact match accuracy (%) on RetroBench. Baseline results are as reported by Liu et al. (2023b) , Liu et al. (2024b) , and Kang et al. (2025) .
Figure 4
Components
Accuracy (%) @ TTA Size
Path Alignment
CoT
RLVR
Pre-ranking
1
3
4
5
30.3
32.8
34.0
35.0
✓
33.5
36.1
37.6
38.5
✓
✓
33.9
37.3
39.2
40.6
✓
✓
✓
36.6
39.3
40.6
41.7
✓
✓
✓
✓
39.7
41.9
42.4
43.5
Table 2: Component-wise ablation study. Without pre-ranking, inference inputs are randomly sampled. Stochastic results report the mean of 10 runs.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Model Category
Model
Parameters
Single-pass Acc (%)
General LLMs
GLM-5
744B
1.8
DeepSeek-V3.2
671B
1.9
Qwen3.5-397B-A17B
397B
1.6
Kimi-K2.5
1.04T
1.0
Chemistry LLMs
ChemLLM-20B-Chat-DPO
20B
0.0
ChemDFM-R-14B
14B
0.5
Appendix
Table 3: Out-of-the-box performance of foundation models on RetroBench.
Figure 5: Qualitative case study. A generated 9-step retrosynthetic plan, featuring integrated CoT reasoning for critical steps.
Hyperparameter
Epochs
Batch Size
Learning Rate
LoRA Rank
Values
1.5
256
1e-4
16
Appendix
Table 4: Hyperparameters for each supervised fine-tuning stage.
Hyperparameter
Group Size
KL β
Batch Size
Learning Rate
Temperature
Top- p
Values
8
0.04
512
1e-6
1.0
0.95
Appendix
Table 5: Hyperparameters for the reinforcement learning stage.
Figure 6: Training reward curve of ReTriP during the RLVR phase. The evolution of the mean retrosynthetic reward demonstrates rapid convergence and effective objective alignment.
Figure 7: Trend of normalized Levenshtein distance (NLD) across retrosynthetic depths. The consistently lower NLD of our method demonstrates that fragment traceability is preserved throughout the planning horizon, minimizing representational noise.