CORTEXA
← Browse
arxivcs.LG2026-07-16

A Continuous-Time Reinforcement Learning Framework for Fine-Tuning Discrete Diffusion Models

Zikun Zhang, Jiayuan Sheng, David D. Yao, Wenpin Tang

We formulate reinforcement learning (RL) in continuous time with discrete state spaces and possibly arbitrary action spaces via a stochastic control approach, where the state dynamics are modeled as a controlled continuous-time Markov chain (CTMC). We consider policy optimization problems and derive the corresponding policy gradient methods, leading to continuous-time variants of proximal policy optimization (PPO) and group relative policy optimization (GRPO). As a primary application, we develop a complete continuous-time RL framework for fine-tuning score-based discrete diffusion models. The proposed framework enables reward-driven optimization without requiring differentiability on the reward signals. In contrast to the existing GRPO-based approaches that only rely on terminal rewards, our formulation allows intermediate reward or advantage signals to be incorporated throughout the denoising trajectory. Importantly, when specialized to masked diffusion models (MDMs), our framework encompasses a rich class of policy parameterizations over the vocabulary simplex with analytically tractable probability ratios, providing a unified perspective on exploration and policy optimization in MDMs. For masked diffusion large language models (dLLMs), we further propose trajectory subsampling techniques to efficiently estimate computationally prohibitive trajectory likelihoods, reducing the computational cost of computing per-position probability ratios. We showcase the effectiveness of our methods on both low-dimensional entropy-regularized optimization problems and RL post-training of dLLMs on mathematical reasoning and coding tasks.

View free PDFSource page

Related papers

arxivcs.LG2026-07-16

Scalable Training of Continuous-Time Spiking Neural Networks with Differentiable Spike-Time Discretization

Yusuke Sakemi, Tomoya Takeuchi, Takeo Hosomi, Kazuyuki Aihara

Continuous-time spiking neural networks (SNNs) provide an event-driven framework for temporal computation, computational neuroscience, and neuromorphic hardware. However, training deep continuous-time SNNs is severely constrained by the memory required for exact spike-time comput…

View free PDFSource page
arxivcs.LG2026-07-19

Robust Assamese Speech Recognition through Controlled Fine-Tuning of Whisper Models

Ganapati Das, Dwipen Laskar, Hasin Afzal Ahmed, Sanjib Kr Kalita, Kshirod Sarmah, Hem Chandra Das, et al.

Developing Automatic Speech Recognition (ASR) for morphologically rich, low-resource languages such as Assamese is challenging due to insufficient annotated speech data. The pretrained Whisper model performs poorly on Assamese speech recognition tasks. This paper presents a contr…

View free PDFSource page
arxivcs.LG2026-07-18

MultiLoReFT: Decoupling Shared and Modality-Specific Subspaces in Multimodal Learning via Low-Rank Representation Fine-Tuning

Sana Tonekaboni, Viktoria Schuster, Caroline Uhler

Real-world perception and decision making are inherently multimodal, integrating complementary signals across modalities. However, training multimodal models faces two main obstacles. First, collecting large-scale, well-aligned paired multimodal datasets is often impractical, mak…

View free PDFSource page
arxivcs.LGcs.AI2026-07-20

Federated Lightweight Fine-Tuning

Radhakrishna Achanta, Will Reed

Federated fine-tuning is bottlenecked by communication: FedAvg and pseudo-gradient schemes transmit a payload that scales with the model, and gradient compression shrinks it by only a constant factor. We take a different lever. Mapping networks generate a network's weights from a…

View free PDFSource page
arxivcs.DCcs.LG2026-07-20

AutoEncoder-Compressed Parallel Split Learning for Pre-trained Model Fine-Tuning

Bas Meuwissen, Vasileios Tsouvalas, Nirvana Meratnia

Distributed Fine-Tuning (DFT) of large-scale Foundation Models (FMs) on resource-constrained edge devices is limited by local compute constraints and communication overhead. Parallel Split Learning (PSL) reduces client-side computation by keeping few model layers on each client a…

View free PDFSource page
arxivphysics.chem-phcond-mat.mtrl-scics.LG2026-07-16

Full-data accuracy with fewer labels for training and fine-tuning machine-learning force fields

Sheng Bi, Yi-Ze Wang, Jun Cheng

Machine-learning force fields (MLFFs) are reliable only near their training distribution, making efficient construction of diverse training sets a major bottleneck for both train-from-scratch and foundation fine-tuning workflows. Active learning can reduce this cost, but standard…

View free PDFSource page