According to arxiv.org, researchers have introduced Flow-OPD, the first unified post-training framework that integrates on-policy distillation into Flow Matching models for text-to-image generation. The framework addresses two critical bottlenecks: reward sparsity from scalar-valued rewards and gradient interference from optimizing heterogeneous objectives. Flow-OPD employs a two-stage alignment strategy that first cultivates domain-specialized teacher models via single-reward GRPO fine-tuning, then consolidates heterogeneous expertise into a single student model through on-policy sampling, task-routing labeling, and dense trajectory-level supervision.
In related work, arxiv.org reports that researchers have developed REPR-ALIGN, a representation alignment objective for adapting autoregressive language models to diffusion language models without full retraining. According to the paper, the approach “aligns the hidden states of the DLM to the frozen AR model at every layer using cosine similarity, while optimizing the standard masked denoising objective.” This technique yields up to 4x training acceleration and is particularly effective in low-data regimes, with no adapters or architectural changes beyond the attention mask required.
Additionally, arxiv.org describes research on the Modality Gap, a geometric anomaly where embeddings of distinct modalities expressing identical semantics occupy systematically offset regions. The work proposes the Fixed-frame Modality Gap Theory to address limitations of prior approaches in large-scale scenarios.