A new arXiv preprint introduces VGFM, a technique that guides flow-matching generative models with dense value signals to produce more expressive robot control policies. The approach targets robot learning from large offline datasets, where multimodal action representations are needed to capture varied behaviors. It aims to improve policy expressiveness within this offline learning paradigm.
A new arXiv paper proposes CyFM, a technique that applies optimal transport on a cylindrical geometry to generate complex-valued signals in few sampling steps. Rather than treating data such as MRI scans and audio spectrograms as flat two-channel Euclidean inputs, the method models amplitude and phase separately. The authors argue this representation better matches the structure of complex signals for generative modelling.
A new arXiv preprint argues that continuous and discrete flow matching, typically developed as separate frameworks, are dual formulations of the same underlying construction. The authors show that applying a position-wise argmax to continuous convex-interpolant flows with one-hot targets recovers the discrete counterpart, and they examine how the choice of source distribution shapes categorical generation. A revised version of the submission has replaced the initial preprint.
A new arXiv paper introduces Refinement-based Flow Policy Optimization, a method for using flow-based policies in online reinforcement learning. Standard flow matching needs samples from the target distribution, which is unavailable when the desired action distribution is only implicitly defined. The approach is presented as a refinement procedure that sidesteps this requirement in online RL settings.
A new arXiv preprint proposes a restricted initialization scheme for flow-matching diffusion models based on sparse matrix decomposition, aimed at reducing the cost of adapting these models to downstream tasks. The work builds on the observation that fine-tuning flow-matching models is expensive, and that low-rank adaptation combined with timestep-aware choices may help. The abstract suggests the method constrains initialization at a principal timestep to improve training efficiency.
A preprint introduces AbFlow, a generative approach that models full-atom antibody structures end-to-end while focusing on the paratope, the region that contacts the antigen. The method uses flow matching augmented with interaction information to guide design. The authors position it as addressing gaps in existing antibody design pipelines, which have lacked a unified generative framework at this structural resolution.
The paper looks at a widely used training recipe for diffusion and flow-matching models, where the network predicts a clean sample that is then converted into a velocity for the loss. That conversion amplifies errors near the endpoints of the noise schedule, making training unstable and tied to a particular sampler. The authors propose aligning the prediction objective with the loss objective so that training stays robust regardless of which sampler is used at inference.
Researchers introduce uFlowCSP, a generative approach to crystal structure prediction built on mean flow models. The method addresses a key drawback of diffusion- and flow-matching-based systems, which typically require many iterative sampling steps during inference. The paper situates the work next to existing generative CSP models such as CDVAE, DiffCSP, FlowMM, and CrystalFlow.