See Every Block Once: Variance-Reduced Post-Training for Masked Diffusion Language Models
arXiv 于 2611.17754v1 发布论文《See Every Block Once: Variance-Reduced Post-Training for Masked Diffusion Language Models》,指出策略梯度方法虽被广泛用于让大语言模型输出对齐人类偏好,但其对数似然计算代价高昂。该论文提出一种面向掩码扩散语言模型的方差缩减后训练方法,主张每个块只采样一次。同一批次中另有 ReMaRL(2611.22576v1)研究冻结扩散语言模型中的解掩与重掩顺序对最终答案准确率的影响,以及 OpenGNS(2611.24331v1)发布跨视觉、语言与扩散模型的梯度噪声规模开放数据集。三者均为 arXiv 新公告摘要,正文细节尚未公开。
How this story unfolded
Key points5 items Model summary · ordered by when the evidence was published
- Point 01 11/16, 13:00
arXiv 2611.17754v1 发布新论文,主题是掩码扩散语言模型的方差缩减后训练方法
Based on 1 - Point 02 11/16, 13:00
该论文指出策略梯度方法被广泛用于对齐大语言模型输出与人类偏好,但对数似然计算存在代价问题
Based on 1 - Point 03 11/18, 13:00
ReMaRL(arXiv 2611.22576v1)关注冻结扩散语言模型中 token 解掩顺序对最终答案准确率的影响
Based on 1 - Point 04 11/19, 13:00
OpenGNS(arXiv 2611.24331v1)发布覆盖视觉、语言与扩散模型的梯度噪声规模开放数据集
Based on 1 - Point 05 11/19, 13:00
OpenGNS 所依据的 GNS 指标衡量随机梯度下降中梯度的信噪比
Based on 1
ReportsBy publish time · newer further down
- November 16, 2026 · Mon
- Start
See Every Block Once: Variance-Reduced Post-Training for Masked Diffusion Language Models
Original summary arXiv:2611.17754v1 Announce Type: new Abstract: Policy gradient methods are widely used to align large language model outputs with human preferences, but the log-likelihoods they r
arXiv cs.CL Original ↗ - November 18, 2026 · Wed
- +2d 0h
ReMaRL: Learning to Unmask and Remask in Frozen Diffusion Language Models
Original summary arXiv:2611.22576v1 Announce Type: new Abstract: Masked diffusion language models generate tokens in parallel, and the order in which tokens are unmasked affects final-answer accura
arXiv cs.LG Original ↗ - November 19, 2026 · Thu
- +3d 0h
OpenGNS: An Open Dataset of the Gradient Noise Scale across Vision, Language, and Diffusion Models
Original summary arXiv:2611.24331v1 Announce Type: new Abstract: The Gradient Noise Scale (GNS) metric measures the signal-to-noise ratio of gradients in stochastic gradient descent and, as shown b
arXiv cs.LG Original ↗ - Still developing · later reports will keep extending this axis