Policy‑Mask is a budget‑aware policy that jointly selects a timestep, a diffusion block, and a subset of channels from a frozen diffusion model to optimize downstream discriminative tasks. By normalizing channel scores and training on the downstream objective, it outperforms fixed‑extraction baselines on segmentation, classification, and semantic correspondence, achieving state‑of‑the‑art results on ADE20K, CIFAR‑100, and SPair‑71k. The method also extends to diffusion transformers by mapping U‑Net channels to image‑token hidden dimensions.
arXiv:2609.37974v1 Announce Type: cross Abstract: Masked diffusion models (MDMs) generate text by unmasking several tokens per step, but they are trained and sampled under different conditions. The m...
The paper introduces GaugeLasso, a method that applies symmetric group‑lasso penalties to transformer channels during training, enabling entire tensor slices to be zeroed out while maintaining dense tensors for GPU efficiency. By calibrating channel penalties based on inference utility per compute, the network self‑organizes into depth‑dependent structural profiles that can be dramatically smaller than the original architecture, achieving up to 255‑fold compression on a polynomial division task and outperforming hand‑designed baselines on language modeling and autoencoding benchmarks. The approach also accelerates training and reveals over‑provisioned axes that guide subsequent design iterations.
The paper introduces Learned Relay Representations (Relay), a technique for Masked Diffusion Models (MDMs) that preserves and forwards internal latent information across denoising steps via a differentiable per-token channel trained with truncated backpropagation through time. Relay enables MDMs to be forward‑thinking, avoiding costly recomputation of internal representations and improving efficiency. Applied to Fast‑dLLM v2, Relay outperforms standard supervised fine‑tuning on coding tasks and reduces inference latency by up to 32%, demonstrating a clear performance‑latency advantage for diffusion language models.
arXiv:2607. 03103v1 Announce Type: cross Abstract: Clinical cardiac imaging pipelines currently deploy separate models for each dataset and modality, incurring redundant training costs and precluding knowledge sharing across anatomically related tasks.
arXiv:2604. 26985v2 Announce Type: replace-cross Abstract: Masked diffusion models (MDMs) generate discrete sequences by iterative denoising under an absorbing masking process.
arXiv:2608. 03929v1 Announce Type: new Abstract: Aligning diffusion models with human preferences usually relies on a sparse terminal reward evaluated on the final generated samples, presenting a severe temporal credit-assignment challenge across the multi-step denoising process.
arXiv:2609.33803v2 Announce Type: replace-cross Abstract: Reward models underpin the alignment of large language models, yet the dominant designs reduce each prompt--response pair to a point estimate...
arXiv:2610.13997v1 Announce Type: cross Abstract: Concept representations in diffusion models drift across network depth and denoising time, yet existing sparse autoencoders offer only per-block or p...
arXiv:2609.15643v1 Announce Type: new Abstract: Flow-matching diffusion models have recently emerged as a strong paradigm for high-fidelity visual generation. However, their prohibitively high fine-t...
arXiv:2607. 25157v1 Announce Type: new Abstract: Discrete masked diffusion language models support bidirectional generation and infilling, but adapting pretrained autoregressive (AR) transformers requires reconciling causal pretraining with bidirectional denoising.
On-Policy Residual Learning for Diffusion Language Models (ORL) introduces a framework that learns residual context directly from model-generated trajectories, using an entropy-contraction objective to sharpen predictions. The approach co-designs algorithmic and system components—such as asynchronous residual capture, block-diffusion attention, and packed compute-balanced training—to enable practical scaling. Experiments on a 4B DLM converted from Qwen3-4B show up to 6.2× faster tokens per forward and 6.9× higher end-to-end throughput, while also improving speed-quality trade-offs over autoregressive models and transferring effectively to other DLMs like LLaDA-2.1-mini and DiffusionGemma.
The paper introduces TIPS, a tiny inference-time policy that controls which tokens are visible during each denoising pass of masked diffusion language models. TIPS, a 600‑K‑parameter shallow transformer, learns to unmask and remask tokens while keeping the language model frozen, using visibility‑aware confidence features and a fixed schedule for expected visibility. Trained via supervised warmup and reinforcement learning, TIPS achieves over 58% accuracy on GSM8K and MATH‑500 with only 32 decoder evaluations, surpassing strong baselines by more than 11 percentage points and remaining competitive at higher budgets without retraining.