Looped transformers with recall and outer normalization produce reachable, input-dependent fixed points with stable gradients, enabling generalization, while those without recall cannot; a new internal recall variant performs competitively or better.
hub
On Layer Normalization in the Transformer Architecture
17 Pith papers cite this work. Polarity classification is still indexing.
17
Pith papers citing it
abstract
The Transformer is widely used in natural language processing tasks. To train a Transformer however, one usually needs a carefully designed learning rate warm-up stage, which is shown to be crucial to the final performance but will slow down the optimization and bring more hyper-parameter tunings. In this paper, we first study theoretically why the learning rate warm-up stage is essential and show that the location of layer normalization matters. Specifically, we prove with mean field theory that at initialization, for the original-designed Post-LN Transformer, which places the layer normalization between the residual blocks, the expected gradients of the parameters near the output layer are large. Therefore, using a large learning rate on those gradients makes the training unstable. The warm-up stage is practically helpful for avoiding this problem. On the other hand, our theory also shows that if the layer normalization is put inside the residual blocks (recently proposed as Pre-LN Transformer), the gradients are well-behaved at initialization. This motivates us to remove the warm-up stage for the training of Pre-LN Transformers. We show in our experiments that Pre-LN Transformers without the warm-up stage can reach comparable results with baselines while requiring significantly less training time and hyper-parameter tuning on a wide range of applications.
hub tools
citation-role summary
background 1
method 1
citation-polarity summary
representative citing papers
Every fixed finite feedforward neural network definable in an o-minimal structure has finite sample complexity in the agnostic PAC setting.
A graph transformer with RL stabilizations is the first to exceed benchmarks for dynamic RMSA, supporting up to 13% more traffic load on networks up to 143 nodes.
Longformer uses local windowed attention plus task-specific global attention to achieve linear scaling and state-of-the-art results on long-document language modeling, QA, and summarization after pretraining.
Cell-level Transformers classify collimated ALP photon-jets versus single photons with AUC 0.98 and regress diphoton mass to ~64 MeV, beating shower-shape and other ML baselines in an ATLAS-like GEANT4 simulation.
MICViT outperforms CNN and transformer baselines on brain age prediction from multimodal 3D MRI by combining modality-specific and cross-modal local/global attention across three heterogeneous datasets.
Quantized reasoning models produce longer chains of thought, inflating token usage and negating per-token speedups from low-bit quantization across multiple benchmarks.
Block Attention Residuals make routing observable as a tensor, but causal probes on trained versus baseline 0.6B models show routing mass often fails to predict causal contribution and structured motifs require training.
Sign-flip perturbations produce π/(π-2) ≈ 2.75 times more transverse output energy than equal-norm sign-preserving perturbations in a ReLU + RMSNorm block because ReLU creates directional asymmetry that RMSNorm's transverse projection exposes.
TaperNorm gradually removes internal normalization in pre-norm transformers via learned gates that reach zero, revealing final norm as a scale anchor and enabling up to 1.18x faster KV-cached decoding with small loss increases.
A JEPA-style EEG foundation model with shallow EMA targets plus light reconstruction reaches strong multi-task transfer and 3.06-year validation age MAE on a large multi-site corpus.
Review Residuals add an update-conditioned gate to transformer residual connections, yielding depth-stable training and performance gains that emerge and grow with model size from 590M parameters upward.
Splitting weight matrices into a fixed-norm direction and learnable per-row/column magnitudes improves LLM training over AdamW/Muon, removes weight decay and warmup, and transfers the optimal LR across width.
A visual transformer model trained on IRIS inversions predicts chromospheric temperature and density from SDO data with correlations around 0.8 on 80% of test cases.
Attention Residuals replaces fixed residual summation with input-dependent softmax attention over preceding layers, and a blocked variant is shown to improve uniformity and downstream performance in a 48B-parameter model pre-trained on 1.4T tokens.
Multi-Gate Residuals stabilizes activation scales in deep residual networks via multi-stream gating and attention pooling without added communication overhead.
A survey taxonomy of LLMs identifies three scaling crises and six efficiency paradigms while tracing the shift from generation to tool-using agents.
citing papers explorer
-
Stability and Generalization in Looped Transformers
Looped transformers with recall and outer normalization produce reachable, input-dependent fixed points with stable gradients, enabling generalization, while those without recall cannot; a new internal recall variant performs competitively or better.
-
Every Feedforward Neural Network Definable in an o-Minimal Structure Has Finite Sample Complexity
Every fixed finite feedforward neural network definable in an o-minimal structure has finite sample complexity in the agnostic PAC setting.
-
Graph Transformers and Stabilized Reinforcement Learning for Large-Scale Dynamic Routing Modulation and Spectrum Allocation in Elastic Optical Networks
A graph transformer with RL stabilizations is the first to exceed benchmarks for dynamic RMSA, supporting up to 13% more traffic load on networks up to 143 nodes.
-
Longformer: The Long-Document Transformer
Longformer uses local windowed attention plus task-specific global attention to achieve linear scaling and state-of-the-art results on long-document language modeling, QA, and summarization after pretraining.
-
Transformer-based machine learning using low-level calorimeter signals for collimated photon identification at collider experiments
Cell-level Transformers classify collimated ALP photon-jets versus single photons with AUC 0.98 and regress diphoton mass to ~64 MeV, beating shower-shape and other ML baselines in an ATLAS-like GEANT4 simulation.
-
Modeling Local, Global, and Cross-Modal Context in Multimodal 3D MRI
MICViT outperforms CNN and transformer baselines on brain age prediction from multimodal 3D MRI by combining modality-specific and cross-modal local/global attention across three heterogeneous datasets.
-
Quantization Inflates Reasoning: Token Inflation as a Hidden Cost of Low-Bit Reasoning Models
Quantized reasoning models produce longer chains of thought, inflating token usage and negating per-token speedups from low-bit quantization across multiple benchmarks.
-
When Does Routing Become Interpretable? Causal Probes on Block Attention Residuals
Block Attention Residuals make routing observable as a tensor, but causal probes on trained versus baseline 0.6B models show routing mass often fails to predict causal contribution and structured motifs require training.
-
A Geometric Analysis of Sign-Magnitude Asymmetry in a ReLU + RMSNorm Block under Ternary Quantization
Sign-flip perturbations produce π/(π-2) ≈ 2.75 times more transverse output energy than equal-norm sign-preserving perturbations in a ReLU + RMSNorm block because ReLU creates directional asymmetry that RMSNorm's transverse projection exposes.
-
Gated Normalization Removal and Scale Anchoring in Pre-Norm Transformers
TaperNorm gradually removes internal normalization in pre-norm transformers via learned gates that reach zero, revealing final norm as a scale anchor and enabling up to 1.18x faster KV-cached decoding with small loss increases.
-
STST-JEPA: Shallow-Target Spatio-Temporal Joint Embedding Prediction Architecture For EEG Self-Supervised Learning
A JEPA-style EEG foundation model with shallow EMA targets plus light reconstruction reaches strong multi-task transfer and 3.06-year validation age MAE on a large multi-site corpus.
-
Review Residuals: Update-Conditioned Residual Gating for Transformers
Review Residuals add an update-conditioned gate to transformer residual connections, yielding depth-stable training and performance gains that emerge and grow with model size from 590M parameters upward.
-
Improving Neural Network Training by Decoupling the Magnitude and Direction of Weight Vectors
Splitting weight matrices into a fixed-norm direction and learnable per-row/column magnitudes improves LLM training over AdamW/Muon, removes weight decay and warmup, and transfers the optimal LR across width.
-
Predicting the thermodynamics in the chromosphere from the translation of SDO data into the IRIS$^{2}$ inversion results using a visual transformer model
A visual transformer model trained on IRIS inversions predicts chromospheric temperature and density from SDO data with correlations around 0.8 on 80% of test cases.
-
Attention Residuals
Attention Residuals replaces fixed residual summation with input-dependent softmax attention over preceding layers, and a blocked variant is shown to improve uniformity and downstream performance in a 48B-parameter model pre-trained on 1.4T tokens.
-
Multi-Gate Residuals
Multi-Gate Residuals stabilizes activation scales in deep residual networks via multi-stream gating and attention pooling without added communication overhead.
-
LLMOrbit: A Circular Taxonomy of Large Language Models -From Scaling Walls to Agentic AI Systems
A survey taxonomy of LLMs identifies three scaling crises and six efficiency paradigms while tracing the shift from generation to tool-using agents.