时光荏苒

mark：知乎关注9万。

步入中年，清明祭祖的意义已经不再是“拜太公”，而是一年一度看望一下那些曾经相熟的亲人。多年以后，那些曾经朝夕相伴的人都将化作一抔黃土葬于山间，一年也就只能相见一次，每每设想到此事，总是情不能自已。

推荐论文：

Attention Sinks Induce Gradient Sinks

Beyond Test-Time Training_ Learning to Reason via Hardware-Efficient Optimal Control

ByteFlow_ Language Modeling through Adaptive Byte Compression without a Tokenizer

Column Normalization and Hyperparameter Transfer

Deriving Hyperparameter Scaling Laws via Modern Optimization Theory

Lost in the Middle at Birth_ An Exact Theory of Transformer Position Bias

On the Role of Batch Size in Stochastic Conditional Gradient Methods

Scale Space Diffusion

Spectral Condition for μP under Width-Depth Scaling

Test-Time Training with KV Binding Is Secretly Linear Attention

Why Adam Can Beat SGD_ Second-Moment Normalization Yields Sharper Tails

https://papers.cool/arxiv/2603.17771,2603.09221,2603.03583,2603.09952,2603.15958,2603.10123,2603.21191,2603.08709,2603.00541,2602.21204,2603.03099

推荐论文：

Adaptive Batch Sizes Using Non-Euclidean Gradient Noise Scales for Stochastic Sign and Spectral Descent

ARO_ A New Lens On Matrix Optimization For Large Models

Convex Dominance in Deep Learning I_ A Scaling Law of Loss and Learning Rate

Fast Catch-Up, Late Switching_ Optimal Batch Size Scheduling via Functional Scaling Laws

Generative Modeling via Drifting

Improving Flow Matching by Aligning Flow Divergence

Power-SMC_ Low-Latency Sequence-Level Power Sampling for Training-Free LLM Reasoning

PRISM_ Structured Optimization via Anisotropic Spectral Shaping

Rotary Positional Embeddings as Phase Modulation_ Theoretical Bounds on the RoPE Base for Long-Context Transformers

Stabilizing Consistency Training_ A Flow Map Analysis and Self-Distillation

TEON_ Tensorized Orthonormalization Beyond Layer-Wise Muon for Large Language Model Pre-Training

The Effect of Mini-Batch Noise on the Implicit Bias of Adam

https://papers.cool/arxiv/2602.03001,2602.09006,2602.07145,2602.14208,2602.04770,2602.00869,2602.10273,2602.03096,2602.10959,2601.22679,2601.23261,2602.01642

推荐论文：

ECO_ Quantized Training without Full-Precision Master Weights

How Do Transformers Learn to Associate Tokens_ Gradient Leading Terms Bring Mechanistic Interpretability

Improved Convergence Rates of Muon Optimizer for Nonconvex Optimization

In Search of Adam's Secret Sauce

L3_ Large Lookup Layers

Manifold constrained steepest descent

R2k is Theoretically Large Enough for Embedding-based Top-k Retrieval

Scalable Power Sampling_ Unlocking Efficient, Training-Free Reasoning for LLMs via Distribution Sharpening

The Depth Delusion_ Why Transformers Should Be Wider, Not Deeper

Why Adam Works Better with β1=β2- The Missing Gradient Scale Invariance Principle

https://papers.cool/arxiv/2601.22101,2601.19208,2601.19400,2505.21829,2601.21461,2601.21487,2601.20844,2601.21590,2601.20994,2601.21739

推荐论文：

Completed Hyperparameter Transfer across Modules, Width, Depth, Batch and Duration

Let's (not) just put things in Context_ Test-Time Training for Long-Context LLMs

Leveraging Flatness to Improve Information-Theoretic Generalization Bounds for SGD

Sliding Window Recurrences for Sequence Models

The Affine Divergence_ Aligning Activation Updates Beyond

https://papers.cool/arxiv/2512.22382,2512.13898,2601.01465,2512.13921,2512.22247

推荐论文：

Dimension-free error estimate for diffusion model and optimal scheduling

Equivalence of Context and Parameter Updates in Modern Transformer Blocks

Optimizing Optimizers for Fast Gradient-Based Learning

Provable Benefit of Sign Descent_ A Minimal Model Under Heavy-Tailed Class Imbalance

When do spectral gradient updates help in deep learning？

https://papers.cool/arxiv/2512.01820,2511.17864,2512.06370,2512.00763,2512.04299

推荐论文：

An Exploration of Non-Euclidean Gradient Descent_ Muon and its Many Variants

Back to Basics_ Let Denoising Generative Models Denoise

Branching Flows_ Discrete, Continuous, and Manifold Flow Matching with Splits and Deletions

Decoupling Positional and Symbolic Attention Behavior in Transformers

How Memory in Optimization Algorithms Implicitly Modifies the Loss

Isotropic Curvature Model for Understanding Deep Learning Optimization_ Is Gradient Orthogonalization Optimal？

L2M_ Mutual Information Scaling Law for Long-Context Language Modeling

Larger Datasets Can Be Repeated More_ A Theoretical Analysis of Multi-Epoch Scaling in Linear Regression

On the Structure of Floating-Point Noise in Batch-Invariant GPU Matrix Multiplication

On the Surprising Effectiveness of Large Learning Rates under Standard Width Scaling

Scaling Laws and In-Context Learning_ A Unified Theoretical Framework

https://papers.cool/arxiv/2510.09827,2511.13720,2511.09465,2511.11579,2502.02132,2511.00674,2503.04725,2511.13421,2511.00025,2505.22491,2511.06232

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。