denoising

该论文研究扩散语言模型智能体之间能否通过压缩的连续隐藏向量进行因果性通信，发现去噪状态的传递能显著影响接收方推理，且单次晚时传递的效果优于多次重复交换。

VLA-Rolling is a lightweight inference framework that reuses discarded action tails from flow-matching VLA models with adaptive denoising budgets, achieving up to 1.48x speedup and up to 15.08 percentage-point success rate gains across 16 model-benchmark pairs on LIBERO, MetaWorld, LIBERO-Plus, and RoboTwin 2.0.

CleanRM 使用少量干净反馈，通过最优传输校准逐样本噪声比例，并以噪声校正的代理损失训练奖励模型，在 RLHF 中实现无偏的干净风险估计并超越现有去噪基线。

ForcingBooster 提出一种自适应逐 patch 去噪框架，根据视频块难度动态分配去噪步数，在保持画质的同时将总去噪步数减少约 50%，加速自回归视频扩散模型的生成。

The paper introduces RACE (Reliable Action-Chunk Extension), a framework that predicts subskill transition timing via an auxiliary one-step denoising pass to enable reliable long action-chunk execution in Vision-Language-Action models, reducing stop-and-go idle time about 5x on real robots while improving success rates over fine-tuning baselines.

This paper finds that latent world action models fail to generalize and proposes Simple-WAM, which improves generalization performance while maintaining efficiency by simplifying future modeling into a single forward pass.

Rolling-WAM introduces a method to distribute denoising across replanning cycles in world action models for robotic manipulation, achieving a 4.5x speedup in replanning while maintaining competitive performance.

A new approach called 'Thinking with Looped Flows' trains recurrent reasoning with local denoising objectives, achieving state-of-the-art performance on ARC-AGI benchmarks among looped models.

This paper introduces Prefix-Denoising Consistency (PDC), a test-time verification method for Diffusion Language Models that improves performance on reasoning tasks by using prefix-conditioned regeneration and majority voting.

Semantic Reasoning Denoising (SRD) introduces an operatorized Markov denoising method to correct semantic errors in language model reasoning trajectories, using executable error operators and iterative refinement, with demonstrated improvements across six benchmarks and competitive performance on transfer tasks.

This paper introduces Entropy-Valley, a training-free length selector for masked diffusion machine translation that uses predictive entropy to improve adequacy, showing that length choice matters more than unmasking order.

TT-Net introduces a quantum-inspired tensor network denoising block for conditional GANs that accesses cross-channel information, outperforming SVD-Net and other methods on PSNR and SSIM metrics across various noise types.

This paper challenges the assumption that iterative denoising always improves forecasts in diffusion-based time series forecasting, proposing a global stopping criterion and a Bernoulli sampler to enhance accuracy and speed.

Introduces Continuous Interaction Diffusion (CID), a diffusion-native model–runtime architecture that integrates tool interaction into iterative denoising, enabling asynchronous external reads while generation continues without waiting for discrete tool calls.

Introduces ZUNA1.1, a 380M-parameter diffusion autoencoder for flexible EEG signal reconstruction, capable of handling variable-length sequences, arbitrary channels, and temporal intervals, while outperforming standard interpolation methods. The model is released open source under the Apache 2.0 license.

Introduces ReDe, a framework that denoises reasoning traces by filtering irrelevant and repetitive steps to improve hallucination detection in large reasoning models, achieving up to 87.32 AUROC on TruthfulQA.

This paper presents GAIA, a geometry-aware learning framework for UWB denoising and work-zone reconstruction that couples temporal range modeling with latent anchor-layout estimation. Evaluated on real-world outdoor data, GAIA reduces range MSE by 18.4% and improves polygon IoU by 15.5% over baselines, demonstrating effective boundary-level reconstruction under NLOS conditions.

This paper introduces dOPSD, an on-policy self-distillation method for diffusion language models that leverages internal denoising trajectories to improve mathematical reasoning and code generation.

PhotoQuilt is a training-free framework that generates high-resolution photomosaics by combining global layout composition with separate tile generation in latent space, overcoming diffusion models' limitations in balancing local detail and global structure.

The paper proposes DIF, a model-agnostic method for denoising implicit feedback in cold-start recommendation by using pseudo-labels from content-similar warm items and uncertainty estimation, achieving significant improvements in a billion-user video app.