时光荏苒

妈妈“现在的孩子怎么这么多病呢”
儿子“不是病多了，是能治的病多了，搁以前都是夭折的”

好精辟的回答，受教了！

来源：https://www.zhihu.com/question/1926923396882621109/answer/1970943451643224638 评论区

入职两年多，第二次到北京总部。

推荐论文：

Adaptive Memory Momentum via a Model-Based Framework for Deep Learning Optimization

AlphaFlow_ Understanding and Improving MeanFlow Models

Arithmetic-Mean μP for Modern Architectures_ A Unified Learning-Rate Scale for CNNs and ResNets

Equilibrium Matching_ Generative Modeling with Implicit Energy-Based Models

From Condensation to Rank Collapse_ A Two-Stage Analysis of Transformer Training Dynamics

On residual network depth

On the Optimal Construction of Unbiased Gradient Estimators for Zeroth-Order Optimization

Optimal Scaling Needs Optimal Norm

Understanding the Generalization of Stochastic Gradient Adam in Learning Neural Networks

Who Said Neural Networks Aren't Linear？

Why Low-Precision Transformer Training Fails_ An Analysis on Flash Attention

https://papers.cool/arxiv/2510.04988,2510.20771,2510.04327,2510.02300,2510.06954,2510.03470,2510.19953,2510.03871,2510.11354,2510.08570,2510.04212

疯狗逻辑：虽然某些人跟疯狗有很大区别，但我只要我认为这些区别不重要，那么某些人就是疯狗。

我是一个比较蠢的人，只会按部就班地进行推导，同时也没啥直觉，通常无法理解能推导出来以外的内容。

推荐论文：

Conda_ Column-Normalized Adam for Training Large Language Models Faster

DiVeQ_ Differentiable Vector Quantization Using the Reparameterization Trick

Efficient Hyperparameter Tuning via Trajectory Invariance Principle

Muon Outperforms Adam in Tail-End Associative Memory Learning

Power Lines_ Scaling Laws for Weight Decay and Batch Size in LLM Pre-training

Unveiling the Role of Learning Rate Schedules via Functional Scaling Laws

https://papers.cool/arxiv/2509.24218,2509.26469,2509.25049,2509.26030,2505.13738,2509.19189

推荐论文：

Are We Really Learning the Score Function？ Reinterpreting Diffusion Models Through Wasserstein Gradient Flow Matching

Attention as an Adaptive Filter

Causal Attention with Lookahead Keys

Depth-Aware Initialization for Stable and Efficient Neural Network Training

Dynamic Low-rank Approximation of Full-Matrix Preconditioner for Training Generalized Linear Models

Flow Straight and Fast in Hilbert Space_ Functional Rectified Flow

Limitations of Normalization in Attention Mechanism

Predicting the Order of Upcoming Tokens Improves Language Modeling

Rotational Equilibrium_ How Weight Decay Balances Learning Across Neural Networks

Scaled-Dot-Product Attention as One-Sided Entropic Optimal Transport

The Optimiser Hidden in Plain Sight_ Training with the Loss Landscape's Induced Metric

Transition Models_ Rethinking the Generative Learning Objective

UltraMemV2_ Memory Networks Scaling to 120B Parameters with Superior Long-Context Learning

Understanding Transformers through the Lens of Pavlovian Conditioning

https://papers.cool/arxiv/2509.00336,2509.04154,2509.07301,2509.05018,2508.21106,2509.10384,2508.17821,2508.19228,2305.17212,2508.08369,2509.03594,2509.04394,2508.18756,2508.08289

推荐论文：

Accelerating Newton-Schulz Iteration for Orthogonalization via Chebyshev-type Polynomials

Zero-Variance Gradients for Variational Autoencoders

https://papers.cool/arxiv/2506.10935,2508.03587

mark：知乎关注8万。

推荐论文：

AbbIE_ Autoregressive Block-Based Iterative Encoder for Efficient Sequence Modeling

Analysis of Muon's Convergence and Critical Batch Size

Conformal Transformations for Symmetric Power Transformers

GPAS_ Accelerating Convergence of LLM Pretraining via Gradient-Preserving Activation Scaling

https://papers.cool/arxiv/2507.08567,2507.01598,2503.03269,2506.22049