We systematically ablate core mechanisms of Transformers and RNNs, finding that attention-augmented Recurrent Highway Networks outperform standard Transformers on forecasting high-dimensional chaotic systems.
The Molecular Transformer applies the Transformer architecture to forward reaction prediction, treating it as SMILES-to-SMILES machine translation. It achieves 90.4% top-1 accuracy on USPTO_MIT, outperforms quantum-chemistry baselines on regioselectivity, and provides calibrated uncertainty scores (0.89 AUC-ROC) for ranking synthesis pathways.
This paper benchmarks 24 machine and deep learning methods on activity cliff compounds (structurally similar molecules with large potency differences) across 30 macromolecular targets. Traditional ML with molecular fingerprints consistently outperforms graph neural networks and SMILES-based transformers on these challenging cases, especially in low-data regimes.
MoLFormer is a transformer encoder with linear attention and rotary positional embeddings, pretrained via masked language modeling on 1.1 billion molecules from PubChem and ZINC. MoLFormer-XL outperforms GNN baselines on most MoleculeNet classification and regression tasks, and attention analysis reveals that the model learns interatomic spatial relationships directly from SMILES strings.
SELFormer is a transformer-based chemical language model that uses SELFIES instead of SMILES as input. Pretrained on 2M ChEMBL compounds via masked language modeling, it achieves strong classification performance on MoleculeNet tasks, outperforming ChemBERTa-2 by ~12% on average across BACE, BBBP, and HIV.
AdaptMol combines an end-to-end graph reconstruction model with unsupervised domain adaptation via class-conditional MMD on bond features and SMILES-validated self-training. Achieves 82.6% accuracy on hand-drawn molecules (10.7 points above prior best) while maintaining state-of-the-art results on four literature benchmarks, using only 4,080 real hand-drawn images for adaptation.
This paper introduces consistency models, a new family of generative models that map any point on a Probability Flow ODE trajectory to its origin. They support fast one-step generation by design, while allowing multi-step sampling for improved quality and zero-shot editing tasks like inpainting and colorization.
This paper introduces Discrete Denoising Diffusion Probabilistic Models (D3PMs), which generalize diffusion to discrete state-spaces using structured Markov transition matrices. D3PMs include uniform, absorbing-state, and discretized Gaussian corruption processes, drawing a connection between diffusion and masked language models.
GraSP introduces a general framework for recognizing graphs in images by framing it as sequential subgraph prediction with a binary classifier. A GNN conditions a CNN via FiLM layers to predict whether a candidate graph is a subgraph of the target. Applied to OCSR on QM9, GraSP achieves 67.5% accuracy with no domain-specific modifications.
This paper introduces Latent Diffusion Models (LDMs), which apply denoising diffusion in the latent space of pretrained autoencoders. By separating perceptual compression from generative learning and adding cross-attention conditioning, LDMs achieve FID 1.50 on Places inpainting and FID 3.60 on ImageNet class-conditional synthesis, with competitive text-to-image generation, at a fraction of the compute cost of pixel-space diffusion.
Tallec and Ollivier show that requiring invariance to time transformations in recurrent models leads to gating mechanisms, recovering key LSTM components from first principles. They propose the chrono initialization for gate biases that improves learning of long-term dependencies.
Battaglia et al. argue that combinatorial generalization requires structured representations, systematically analyze the relational inductive biases in standard deep learning architectures (MLPs, CNNs, RNNs), and present the graph network as a unifying framework that generalizes and extends prior graph neural network approaches.