autoregressive synthesis decoding

Traditional retrosynthetic planning suffers from exponential search-space explosion and poor generalizability due to its reliance on single-step iterative decomposition. To address this, we propose an end-to-end multi-step synthesis pathway generation paradigm, formulating multi-step retrosynthesis for the first time as a controllable conditional sequence generation task—enabling hard constraints such as prescribed step count and specified starting materials. Methodologically, we design a Transformer-based seq2seq model that jointly encodes molecular graphs and SMILES sequences, enabling molecule-level conditional pathway prediction. On the PaRoutes benchmark, our approach achieves a Top-1 accuracy 2.2–3.3× higher than state-of-the-art baselines. Moreover, it successfully generates chemically feasible, multi-step routes for numerous unseen FDA-approved drugs, demonstrating substantial improvements in both planning efficiency and cross-molecule generalization.

Graph-structured data (e.g., molecular graphs) lack a natural autoregressive generation order, hindering sequential generative modeling. Method: This paper proposes a learnable, dynamic ordering strategy that models the generation sequence as a state-dependent probability distribution, jointly optimized with the graph generation process. It innovatively formulates the autoregressive ordering as a trainable policy and establishes an end-to-end optimization framework grounded in the variational lower bound, integrating variational inference, gradient estimation (e.g., REINFORCE or Gumbel-Softmax), and graph neural networks to enable co-learning of ordering and structure. Contribution/Results: The approach achieves state-of-the-art performance on molecular generation benchmarks QM9 and ZINC250k, significantly surpassing prior methods in Fréchet Chemical Distance (FCD). It is the first method to realize fully data-driven, jointly optimized learning of generation order—without relying on handcrafted heuristics or fixed traversal rules.

This study addresses the trajectory bias problem in constrained decoding for masked diffusion models by proposing the TWISTER decoder. This method introduces automaton twisting into a Sequential Monte Carlo (SMC) framework for the first time, integrating finite-state automata with Feynman-Kac particle filtering to derive and correct step-level sampling tilts. It formally proves that the resulting target distribution is equivalent to an unbiased Doob h-transform path measure. The proposed approach enables exact computation and efficient sampling under regular language constraints, ensuring that generated sequences strictly satisfy syntactic structures while preserving the model’s original probability distribution. Consequently, this work significantly improves the statistical consistency of constrained generation.

This work addresses the one-to-many mapping challenge in planar path synthesis—where a single target trajectory corresponds to multiple linkage mechanisms (e.g., four-bar, six-bar, eight-bar)—by proposing a unified framework based on conditional autoregressive sequence generation. The approach formulates mechanism synthesis as a discrete sequence generation task, leveraging a decoder-only Transformer integrated with a variational autoencoder (VAE). By quantizing joint coordinate sequences, incorporating explicit mechanism-type tokens, and employing bounded latent-space noise scheduling, the model enables retrieval-free generation of diverse, high-fidelity designs. A novel ordered structure-aware Gaussian smoothing auxiliary loss is introduced to enhance geometric coherence, while dynamic time warping (DTW), Chamfer distance, and forward kinematics are jointly utilized for geometric alignment and evaluation. Experiments demonstrate strong performance on held-out test sets, achieving average Chamfer distance of 0.0132 and DTW of 0.153; further gains are realized via a VAE latent-space k-nearest-neighbor topology matching baseline, yielding Chamfer distance of 0.0071 and DTW of 0.117.

Existing generative models for drug design frequently produce molecules that are chemically unsynthesizable. Method: We propose a GFlowNet-based generative framework grounded in forward-synthesis pathway modeling, which—uniquely—explicitly embeds chemical reaction templates and a purchasable reagent library into the action space to enable end-to-end learning of synthetic feasibility constraints. To mitigate reaction-encoding bias and support multi-constraint MDP formulation, we introduce an inverse-policy learning mechanism. Our approach integrates reaction graph encoding, SA Score-guided optimization, and independent retrosynthetic validation. Results: Experiments demonstrate substantial improvements over baselines: generated molecules exhibit significantly higher structural diversity, average SA Score decreases by 12.3%, retrosynthetic success rate increases by 18.7%, and the model reliably infers feasible synthesis pathways for novel molecules.

This study addresses the practical unsynthesizability of generated molecules in goal-directed molecular design, alongside the low search efficiency and goal deviation inherent in existing discrete-space methods. To overcome these limitations, this work proposes RouteFlow, a novel framework that reformulates synthesizable molecular design as a search problem within a continuous route latent space. The method introduces a cycle-consistency mechanism to prevent optimization from drifting off the true synthetic manifold and employs a reward-guided flow matching sampler for efficient directed optimization. Experimental evaluations across 16 tasks demonstrate that RouteFlow achieves state-of-the-art sample efficiency, synthetic accessibility, and retrosynthetic success rates, enabling efficient and intrinsically synthesizable molecular discovery.

This study addresses the challenge of balancing reaction pathway coverage and instruction controllability in retrosynthesis models by proposing the RIGS framework. Introducing a novel "coverage-before-control" paradigm, RIGS achieves instruction-guided retrosynthesis through a two-stage training procedure. First, a language projector learns preference projections to construct nested multi-support sets. Subsequently, with the generator frozen, lightweight instruction tuning is performed via residual adapters, effectively decoupling coverage capacity from control logic. Experimental results demonstrate that this approach significantly enhances both pathway diversity and instruction-following accuracy. Furthermore, the study reveals a non-monotonic relationship between model scale and coverage-control performance.

This study addresses the challenge of predicting generation quality for hybrid diffusion-autoregressive models under a fixed number of decoding steps by proposing a unified theoretical framework for decoding schedule performance prediction. Methodologically, the model is formulated as a path on a corruption lattice, where schedule cost is defined as the dependencies discarded during parallel steps. Leveraging graph arboricity analysis and pairwise dependency kernel estimation, schedule rankings are predicted directly from pretrained weights. Theoretically, we prove that the minimum number of steps required for zero-cost schedules is determined by data geometry. Empirically, the effectiveness of this predictive framework is validated across text, image, and video generation tasks. This work establishes both a theoretical foundation and practical principles for future decoding design in generative models.

This work addresses the challenge of implementing probabilistic computations such as Bayesian inference in biochemical systems, where conventional chemical reaction networks (CRNs) are often prohibitively large for practical use. The authors introduce, for the first time, factor graph reduction theory into CRN design by uncovering the implicit factor graph structure embedded within the Napp–Adams compilation framework and applying graph reduction algorithms. This approach significantly compresses the network size while preserving the fixed points of belief propagation for key variables. By doing so, it overcomes a critical limitation of existing CRN simplification techniques, which are unable to handle probabilistic models, thereby enabling efficient and exact compression of probabilistic CRNs.

Existing 3D molecular generation methods struggle to simultaneously support unconditional and fragment-based conditional generation while often requiring predefined molecular sizes. This work proposes KRONOS, a unified framework that integrates autoregressive and diffusion mechanisms within the latent space of a pretrained molecular autoencoder to jointly model molecular graph topology and 3D geometry. Inspired by Fill-in-the-Middle, KRONOS employs a hybrid training strategy that enables efficient support for both generation modes within a single architecture without performance trade-offs. Experimental results demonstrate that KRONOS achieves superior unconditional generation performance on QM9 and GEOM-Drugs compared to existing autoregressive models and matches that of diffusion-based approaches, while preserving near-identical unconditional generation quality even when trained for fragment-conditioned synthesis.