Flow-GRPO: Training Flow Matching Models via Online RL

HiCoGen introduces a Hierarchical Compositional Generative framework that enhances text-to-image generation in diffusion models by utilizing a Chain of Synthesis paradigm. This method decomposes complex prompts into semantic units, synthesizing them iteratively to improve compositional accuracy and visual context in generated images.

OmniRefiner has been introduced as a detail-aware refinement framework aimed at improving reference-guided image generation. This framework addresses the limitations of current diffusion models, which often fail to retain fine-grained visual details during image refinement due to inherent VAE-based latent compression issues. By employing a two-stage correction process, OmniRefiner enhances pixel-level consistency and structural fidelity in generated images.

A new benchmark has been introduced to advance research in reinforcement learning (RL) for continuous control, featuring 200 diverse tasks with language instructions and demonstrations. The study presents Newt, a language-conditioned multitask world model that is pretrained on demonstrations and optimized through online interaction across all tasks.

A new study has introduced differential smoothing as a method to mitigate diversity collapse in large language models (LLMs) during reinforcement learning (RL) fine-tuning. This approach provides a formal proof of the selection and reinforcement bias leading to reduced output variety and proposes a solution that enhances both correctness and diversity in model outputs.

A recent study has introduced a quantum-enhanced reinforcement learning (RL) approach to optimize the initialization of the Newton-Raphson method, which is critical for solving power flow equations. This method aims to improve convergence rates, particularly in scenarios with high renewable energy penetration where traditional methods struggle.

A recent study investigates the limitations of applying Markov Chain Monte Carlo (MCMC) methods to arbitrary objective functions, particularly through a two-block MCMC framework that alternates between Metropolis-Hastings and Gibbs sampling. The research highlights that the performance of these methods is significantly influenced by the sharpness of the likelihood form used, introducing a sharpness parameter to explore its effects on regularization and in-sample performance.

A new approach to combinatorial optimization has emerged with the introduction of Plan-and-Branch-and-Bound (PlanB&B), a model-based reinforcement learning (MBRL) agent designed to enhance the efficiency of branch-and-bound (B&B) solvers in Mixed-Integer Linear Programming (MILP). This method aims to learn optimal branching strategies tailored to specific MILP distributions, moving beyond traditional static heuristics.

A recent study introduces a novel approach to latent safety filters that enhance Hamilton-Jacobi reachability, enabling safe visuomotor control under complex constraints. The research highlights the limitations of current methods that rely on discrete policy switching, which may compromise performance in high-dimensional environments.