Research map / Architectures and efficiency / Attention and transformers
Language as a Wave Phenomenon: Semantic Phase Locking and Interference in Neural Networks
Alper YILDIRIM, İbrahim Yücedağ
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 |
|---|---|---|---|
| retain only the phase of the spectral filter (H/|H|) instead of the full learned filter (ablation) | PRISM encoder (frozen) · WikiText-103 filter intervention | degradation factor: 1.41× (constrained), 1.51× (unconstrained) | No clear effect |
| retain only magnitude |H| of the spectral filter (ablation) | PRISM encoder (frozen) · WikiText-103 filter intervention | degradation factor: 153× (constrained), 173× (unconstrained) | Hurt |
| reverse the filter phase (H*) (ablation) | PRISM encoder (frozen) · WikiText-103 filter intervention | degradation factor: 245× (constrained), 139× (unconstrained) | Hurt |
| shuffle the filter phase (ablation) | PRISM encoder (frozen) · WikiText-103 filter intervention | degradation factor: 160× (constrained), 161× (unconstrained) | Hurt |
| remove the spectral filter (H = 1) (ablation) | PRISM encoder (frozen) · WikiText-103 filter intervention | degradation factor: 328× (constrained), 78× (unconstrained) | Hurt |
| scramble phase versus magnitude at a single layer (ablation) | unconstrained PRISM · WikiText-103 layer-wise intervention | degradation factor: 72-205× (phase) vs 1.1× (magnitude) | Hurt |
| use complex PRISM spectral stream instead of real spectral encoder in WPT (ablation) | WPT · WikiText-103 masked LM | perplexity: 5.55 -> 4.94 | Improved |
| use CReLU + LayerNorm instead of ModReLU + PPLN in standalone PRISM (ablation) | PRISM · WikiText-103 masked LM | perplexity: 6.06 -> 5.80 | Improved |
| use CReLU + LayerNorm instead of ModReLU + PPLN in the WPT relational stream (ablation) | WPT · WikiText-103 masked LM | perplexity: 4.94 -> 5.46 | Hurt |
| use complex PRISM instead of real-valued spectral encoder (standalone) | spectral encoder · WikiText-103 masked LM | perplexity: 6.15 -> 6.06 | Improved |
| replace FNet Hybrid with PRISM Hybrid | hybrid spectral encoder · WikiText-103 masked LM | perplexity: 9.87 -> 6.06 | Improved |
| use WPT dual-stream instead of Transformer | encoder · WikiText-103 masked LM | perplexity: 5.28 -> 4.94 | Improved |
Closest papers by meaning
- A Phase Transition between Positional and Semantic Learning in a Solvable Model of Dot-Product Attention (NeurIPS 2024 · similarity 0.84)
- What needs to go right for an induction head? A mechanistic study of in-context learning circuits and their formation (ICML 2024 · similarity 0.84)
- From Condensation to Rank Collapse: A Two-Stage Analysis of Transformer Training Dynamics (NeurIPS 2025 · similarity 0.84)
- Is attention required for ICL? Exploring the Relationship Between Model Architecture and In-Context Learning Ability (ICLR 2024 · similarity 0.83)
- Phase Transitions in Attention: A Bayesian Theory of Copy Head Emergence (NeurIPS 2026 · similarity 0.83)
- Looking Beyond the Top-1: Transformers Determine Top Tokens in Order (ICML 2025 · similarity 0.83)
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.
