Research map / Architectures and efficiency / Attention and transformers
Decoupling Positional and Symbolic Attention in Transformers
Felipe Urrutia, Jorge Salas, Alexander Kozachinskiy, Cristian Buc Calderon, Hector Pasten, Cristobal Rojas
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 |
|---|---|---|---|
| give a 1-head RoPE model only low frequencies on the positional Index task (ablation) | 1-layer RoPE toy transformer · Index task (positional) | accuracy | Hurt |
| give a 1-head RoPE model only high frequencies on the symbolic Information Retrieval task (ablation) | 1-layer RoPE toy transformer · Information Retrieval task (symbolic) | accuracy | Hurt |
| add a second RoPE angle (theta_1=0 plus a flexible theta_2) to a single head (ablation) | 1-head toy transformer · Partial induction task | accuracy | Improved |
| increase the second angle theta_2 too much in the two-angle head (ablation) | 1-head toy transformer with two angles · Partial induction task | accuracy | Hurt |
Closest papers by meaning
- Round and Round We Go! What makes Rotary Positional Encodings useful? (ICLR 2025 · similarity 0.93)
- On the Emergence of Position Bias in Transformers (ICML 2025 · similarity 0.91)
- PaTH Attention: Position Encoding via Accumulating Householder Transformations (NeurIPS 2025 · similarity 0.90)
- Talking Heads: Understanding Inter-Layer Communication in Transformer Language Models (NeurIPS 2024 · similarity 0.89)
- Dynamical Properties of Tokens in Self-Attention and Effects of Positional Encoding (NeurIPS 2025 · similarity 0.89)
- Analyzing Feed-Forward Blocks in Transformers through the Lens of Attention Maps (ICLR 2024 · 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.
