New Fine-tuning Method Optimizes Molecular Generation Across Architectures

Goal-directed optimization is crucial for designing molecules with specific desired properties, but current methods, often based on policy-gradient reinforcement learning, are complex and tied to specific model architectures. This limits their reusability across different generative designs. Supervised fine-tuning (SFT) is simpler but traditionally lacks a mechanism to incorporate reward signals directly into its updates. A new approach, Elite-Weighted Supervised Fine-tuning (EW-SFT), addresses these limitations. EW-SFT guides the model update by selecting a high-scoring "elite" set of molecules based on their reward. The model is then fine-tuned using its own pretraining loss on this elite set. This method effectively passes reward information primarily through the selection process, rather than continuous weighting. A key advantage of EW-SFT is its universality: it works across diverse molecular generators (autoregressive, masked-diffusion, discrete-flow) and design tasks (de novo, motif-extension, linker-design) because it only requires scored molecules and the model's native loss. Empirical results show EW-SFT consistently outperforms native optimizers under fixed budgets for 3D shape alignment and 2D similarity oracles, demonstrating its effectiveness and sample efficiency without needing complex reinforcement learning formulations.