Flow-Based Sequential Generator

• Flow-based sequential generators are models that use invertible transformations to construct output sequences while maintaining tractable likelihoods.
• They integrate techniques such as flow matching and trajectory balance to effectively train on sequential data in applications like recommendation systems and molecule synthesis.
• Empirical results show improvements in performance and diversity across modalities, including natural language, audio, and molecular design, by robustly modeling sequential dependencies.

A sequential generator with flow-based synthesis refers to a class of generative models in which the construction of output sequences occurs through the integration or composition of invertible or approximately invertible transformations (i.e., flows), parameterized to synthesize target data in discrete or continuous time, while preserving tractable calculation of data likelihoods or satisfying principled sampling guarantees. These methods unify the modeling of both sequential dependencies and high-dimensional target distributions, and have been deployed across tasks including sequential recommendation, conditional sequence generation, molecule synthesis, and time-series/audio generation.

1. Core Principles of Flow-Based Sequential Generators

Flow-based generative models utilize a sequence of invertible mappings, known as flows, to transform a simple base distribution (commonly Gaussian) into the data distribution of interest. When extended to sequential generation, these transformations model the dependencies across ordered data such as tokens, actions, or time steps, providing both expressivity and efficient likelihood computation.

A typical flow-based sequential generator employs either:

• Continuous flows: defined by an ODE parameterized by a velocity field, facilitating smooth interpolation between noise and data.
• Discrete flows: stacking invertible affine or coupling transformations, permitting parallel or autoregressive computation across steps.

Sequential construction can involve:

• Directly modeling dependencies over sequential positions (e.g., positions in text, item lists, atoms in molecules).
• Conditioning each flow step on the historical context, generated tokens or actions, or additional input modalities.

2. Flow Matching and Conditional Losses

Flow matching introduces a learning objective that aligns the generator’s vector field with an analytically tractable coupling between the base and target distributions. This is realized through integrating the ODE: $$\frac{d\mathbf{z}_t}{dt} = v_\theta(\mathbf{z}_t, t)$$ or, for discrete trajectories, via enforceable flow-consistency equations over transitions.

In FMRec for sequential recommendation, flow matching is realized by linear interpolation in embedding space,

$$\mathbf{z}_t = (1-t)\mathbf{x}_c + t\mathbf{x}_n, \quad t\in[0,1]$$

with a vector field

$$u_t(\mathbf{z}_t | \mathbf{x}_n) = -\mathbf{x}_c + \mathbf{x}_n$$

The model is trained to predict the clean embedding from any interpolation point using the loss: $$\mathcal{L}_{FM} = \mathbb{E}\left[ \| f_\Theta(\mathbf{z}_t, t) - \mathbf{x}_c \|_2^2 \right]$$ Auxiliary losses, such as cross-entropy over item prediction and MSE on reconstructed user-history, further anchor the hidden representations, enhancing robustness to noise perturbations (Liu et al., 22 May 2025).

In GFlowNet-based approaches, such as S3-GFN, training hinges on pathwise balance constraints (Trajectory Balance), ensuring the model samples final states in exact proportion to a designated reward function. The loss forms, e.g.,

$$L_{TB}(\tau) = \left(\log \frac{Z_\theta \cdot P_F(\tau;\theta)}{R(x) \cdot P_B(\tau|x;\theta)} \right)^2$$

ensure that, globally, the learned policy satisfies the required sampling proportionality properties (Bengio et al., 2021, Kim et al., 4 Feb 2026).

3. Architectural Variants Across Modalities

Recommender Systems

FMRec uses:

• A "straight-line" flow in embedding space to interpolate between target item embeddings and noise.
• Two Transformer decoders for context fusion and embedding prediction.
• A deterministic ODE sampler for inference, eschewing stochastic noise to produce recommendations closely aligned with user preferences (Liu et al., 22 May 2025).

Sequence-to-Sequence Modeling

FlowSeq employs conditional normalizing flows over per-time-step latent variables, transforming Gaussian base noise into the latent space, with a parallel softmax decoder over output tokens for fast, non-autoregressive sequence generation (Ma et al., 2019).

Molecule and Pathway Design

CGFlow generalizes flow matching to compositional objects, integrating discrete (sequence of synthesis actions) and continuous (3D conformations) states, and training via a joint objective decomposed into flow-matching on continuous states and GFlowNet trajectory balance for combinatorial steps (Shen et al., 10 Apr 2025).

S3-GFN builds a sequential SMILES generator, leveraging a large language-model prior and reward-proportional trajectory sampling to ensure synthesizability, employing soft-constraint regularization via contrastive buffer replay to penalize non-synthesizable outputs (Kim et al., 4 Feb 2026).

Audio, Speech, and Music Synthesis

Flowtron and WaveFlow model raw audio as a sequential or parallel stack of invertible flows, conditioning on text or mel-spectrograms, and enabling exact maximum likelihood training and tractable sampling (Valle et al., 2020, Ping et al., 2019).

FlowSynth extends flow matching to a fully probabilistic setting, learning a Gaussian-distributed velocity field, which supports principled uncertainty-aware test-time search and yields high-quality virtual instrument synthesis (Yang et al., 24 Oct 2025).

SpeechFlow trains a large non-autoregressive Transformer via flow matching with masked pre-training, supporting downstream tasks via adaptation to task-specific conditions and realizing efficient ODE-based sampling (Liu et al., 2023).

4. Algorithmic Structure and Sampling

Sequential flow-based generators employ a two-phase structure:

• Training (Forward Process):
  • For fixed data (e.g., sequence, graph, audio), sample a time/interpolation parameter, form intermediate states by coupling data and noise, and compute vector field predictions.
  • Minimize losses corresponding to flow-matching, trajectory balance, or similar constraints, often augmented with domain-specific auxiliary terms.
  • Update model parameters via backpropagation through all network layers or flow modules.
• Inference (Reverse or Generative Process):
  • Initialize from the base distribution (e.g., Gaussian noise, empty prefix).
  • Sequentially, or via integration, apply the learned (possibly deterministic) flow or policy to construct the output data.
  • Decoding may involve ODE solvers (for continuous flows), autoregressive or parallel transforms (for discrete flows), and post-processing (e.g., mapping embeddings to items).

A representative pipeline (FMRec) involves, at training, interpolating between noise and target embedding, integrating through Transformer decoders, and applying three losses. At inference, it deterministically integrates the ODE from noise to embedding, then ranks item logits (Liu et al., 22 May 2025).

5. Comparative Performance and Practical Design

Empirical results show:

• FMRec achieves on average a 6.53% improvement in performance over diffusion and baseline methods on sequential recommendation tasks (Liu et al., 22 May 2025).
• S3-GFN achieves ≥95% synthesizable molecule fraction with high reward and diversity in molecule generation tasks (Kim et al., 4 Feb 2026).
• CGFlow attains state-of-the-art 3D molecular docking scores and synthetic feasibility, sampling 5.8× as many diverse high-scoring modes as the best baseline (Shen et al., 10 Apr 2025).
• FlowSynth outperforms state-of-the-art music synthesizers in timbre consistency and adaptive search, with ~85% reduction in FAD and ~40% reduction in pitch MAD versus TokenSynth (Yang et al., 24 Oct 2025).

Design trade-offs include:

• Deterministic vs. stochastic sampling (ODE-based integration versus probabilistic sampling for diversity or uncertainty calibration).
• Augmenting flow-matching with auxiliary cross-entropy, contrastive, or reconstruction losses to regularize model outputs and impose domain constraints.
• Use of pretrained priors to bias the generation space toward validity and feasibility (e.g., SMILES priors in S3-GFN).
• In GFlowNet formulations, care in objective and loss choice is required to prevent mode collapse and ensure correct mass conservation, especially in the presence of cycles (Bengio et al., 2021, Brunswic et al., 2023).

6. Limitations, Extensions, and Theoretical Guarantees

Flow-based sequential generators are subject to several constraints:

• The invertibility of each step can pose architectural and computational challenges, particularly for long sequences or high-dimensional data.
• Cycle handling in action/state graphs requires careful loss design, as flow-matching losses uninformed by cycle-invariance can result in pathologies (e.g., unbounded walk lengths, incorrect sampling distributions). Stable loss modifications have been developed to address this (Brunswic et al., 2023).
• In image and high-dimensional settings, one-shot flow matching often fails to fully match the target distribution; iterative correction via path refinement or gradual subinterval advances mitigates drift and improves sample fidelity (Haber et al., 23 Feb 2025).
• Scalability and sampling efficiency depend on mixing the number of flow steps, step size, and architectural choices (e.g., the trade-off between height of squeezing in WaveFlow and the number of sequential steps) (Ping et al., 2019).

Theoretical proofs guarantee, under the respective learning objectives (flow-matching, trajectory balance, etc.), that minimizers of these losses yield policies or velocity fields that sample in direct proportion to the target distribution or reward (Bengio et al., 2021).

7. Applications and Emerging Directions

The sequential generator with flow-based synthesis paradigm has enabled:

• Next-item and user sequence modeling in recommender systems, robust to noise and stochasticity (Liu et al., 22 May 2025).
• Reward-proportional generation of molecular structures, paths, and synthesis routes with explicit control over synthesizability and multiobjective design (Bengio et al., 2021, Kim et al., 4 Feb 2026, Shen et al., 10 Apr 2025).
• High-fidelity, controllable synthesis in text-to-speech, speech enhancement, instrument modeling, and general time-series data (Valle et al., 2020, Ping et al., 2019, Yang et al., 24 Oct 2025, Liu et al., 2023).
• Non-autoregressive and scalable sequence generation for natural language and translation tasks (Ma et al., 2019).

Future research trends include iterative flow matching to refine sample quality (Haber et al., 23 Feb 2025), further integration of uncertainty-aware control at test-time (Yang et al., 24 Oct 2025), generalized treatment of continuous and cyclic state spaces (Brunswic et al., 2023), and joint modeling of discrete-compositional and continuous-structural generative processes (Shen et al., 10 Apr 2025).

---

References:

• FMRec: "Flow Matching based Sequential Recommender Model" (Liu et al., 22 May 2025)
• GFlowNet: "Flow Network based Generative Models for Non-Iterative Diverse Candidate Generation" (Bengio et al., 2021)
• S3-GFN: "Synthesizable Molecular Generation via Soft-constrained GFlowNets with Rich Chemical Priors" (Kim et al., 4 Feb 2026)
• CGFlow: "Compositional Flows for 3D Molecule and Synthesis Pathway Co-design" (Shen et al., 10 Apr 2025)
• Flowtron: "Flowtron: an Autoregressive Flow-based Generative Network for Text-to-Speech Synthesis" (Valle et al., 2020)
• WaveFlow: "WaveFlow: A Compact Flow-based Model for Raw Audio" (Ping et al., 2019)
• FlowSynth: "FlowSynth: Instrument Generation Through Distributional Flow Matching and Test-Time Search" (Yang et al., 24 Oct 2025)
• SpeechFlow: "Generative Pre-training for Speech with Flow Matching" (Liu et al., 2023)
• FlowSeq: "FlowSeq: Non-Autoregressive Conditional Sequence Generation with Generative Flow" (Ma et al., 2019)
• Iterative FM: "Iterative Flow Matching -- Path Correction and Gradual Refinement for Enhanced Generative Modeling" (Haber et al., 23 Feb 2025)
• Non-acyclic GFlowNet: "A Theory of Non-Acyclic Generative Flow Networks" (Brunswic et al., 2023)