Research map / Architectures and efficiency / Attention and transformers
Signal Propagation in Transformers: Theoretical Perspectives and the Role of Rank Collapse
Lorenzo Noci, Sotiris Anagnostidis, Luca Biggio, Antonio Orvieto, Sidak Pal Singh, Aurelien Lucchi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase token alignment toward rank collapse | Transformer self-attention · Transformer initialization | query/key gradient norm | Hurt |
| scale residual branches with depth as α₁², α₂² = O(1/L) | Transformer · Infinite-depth signal propagation | token correlation | Improved |
| use nonzero residual strengths that are constant with depth instead of depth-dependent scaling (ablation) | Transformer · Infinite-depth signal propagation | token correlation | Hurt |
| remove depth-dependent residual scaling (ablation) | Transformer · Finite-depth signal propagation | token correlation | Hurt |
| add layer normalization (ablation) | Transformer with residual scaling · Initialization | token correlation | No clear effect |
| use PRE-LN instead of no layer normalization (ablation) | Transformer without residual scaling · Initialization | token correlation | Improved |
| standardize input variance with layer normalization | Transformer self-attention · Initialization | query/key gradient norm | Improved |
| train with residual scaling and inverse-temperature parameters using SGD instead of Adam | Transformer encoder-decoder · IWSLT14 German-to-English | BLEU | Mixed |
| increase input variance | Transformer self-attention · Initialization | query/key gradient norm | Improved |
| increase input variance | Transformer self-attention · Initialization | value gradient norm | Improved |
Closest papers by meaning
- On the Role of Attention Masks and LayerNorm in Transformers (NeurIPS 2024 · similarity 0.92)
- Dynamical Properties of Tokens in Self-Attention and Effects of Positional Encoding (NeurIPS 2025 · similarity 0.90)
- Mind the Gap: a Spectral Analysis of Rank Collapse and Signal Propagation in Attention Layers (ICML 2025 · similarity 0.90)
- Unraveling the Gradient Descent Dynamics of Transformers (NeurIPS 2024 · similarity 0.89)
- Transformers Learn Faster with Semantic Focus (NeurIPS 2025 · similarity 0.89)
- From Condensation to Rank Collapse: A Two-Stage Analysis of Transformer Training Dynamics (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
