Research map / Architectures and efficiency / Attention and transformers
Language Through a Prism: A Spectral Approach for Multiscale Language Representations
Alex Tamkin, Dan Jurafsky, Noah Goodman
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 |
|---|---|---|---|
| Add a prism layer after the last BERT layer | BERT · 20 Newsgroups | accuracy: 32.21 -> 51.01 | Improved |
| Add a prism layer after the last BERT layer | BERT · Switchboard Dialog Speech Acts | accuracy: 47.09 -> 54.02 | Improved |
| Add a prism layer after the last BERT layer | BERT · Penn Treebank | accuracy: 95.86 -> 94.41 | Hurt |
| Use only the lowest-frequency sector (ablation) | BERT + Prism · 20 Newsgroups | accuracy: 45.1% versus 51.0% for the full model | Hurt |
| Use only the highest-frequency sector (ablation) | BERT + Prism · 20 Newsgroups | accuracy: 5.3% versus 51.0% for the full model | Hurt |
| Use only the highest-frequency sector (ablation) | BERT + Prism · Penn Treebank | accuracy: 84.1% versus 94.4% for the full model | Hurt |
| Use only the lowest-frequency sector (ablation) | BERT + Prism · Penn Treebank | accuracy: 16.8% versus 94.4% for the full model | Hurt |
| Place prism layers after every BERT layer instead of only after the last layer (ablation) | BERT + Prism · 20 Newsgroups | accuracy: 45.2% (-5.8%) | Hurt |
| Place prism layers after every BERT layer instead of only after the last layer (ablation) | BERT + Prism · Switchboard Dialog Speech Acts | accuracy: 51.8% (-2.2%) | Hurt |
| Place prism layers after every BERT layer instead of only after the last layer (ablation) | BERT + Prism · Penn Treebank | accuracy: 94.0% (-0.4%) | Hurt |
| Use a low-pass spectral filter | Pretrained BERT representations · 20 Newsgroups | probing accuracy | Improved |
| Use a high-pass spectral filter | Pretrained BERT representations · Penn Treebank | probing accuracy | Mixed |
Closest papers by meaning
- The Representation Landscape of Few-Shot Learning and Fine-Tuning in Large Language Models (NeurIPS 2024 · similarity 0.86)
- Layer by Layer: Uncovering Hidden Representations in Language Models (ICML 2025 · similarity 0.86)
- Less is More: Local Intrinsic Dimensions of Contextual Language Models (NeurIPS 2025 · similarity 0.86)
- Scaling Embedding Layers in Language Models (NeurIPS 2025 · similarity 0.85)
- How do Large Language Models Handle Multilingualism? (NeurIPS 2024 · similarity 0.85)
- OWLS: Scaling Laws for Multilingual Speech Recognition and Translation Models (ICML 2025 · similarity 0.85)
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.
