Research map / Architectures and efficiency / Attention and transformers
How Do Transformers Learn Topic Structure: Towards a Mechanistic Understanding
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Pretrain the transformer on Wikipedia rather than use random initialization | Bert · Wikipedia LDA topics, ambiguity threshold 0.0005 | same-topic/different-topic embedding cosine similarity ratio: 1.00080 -> 1.21 | Helped |
| Pretrain the transformer on Wikipedia rather than use random initialization | Bert · Wikipedia LDA topics, ambiguity threshold 0.0005 | same-topic/different-topic attention weight ratio: 0.99943 -> 1.32 | Mixed |
| Pretrain the transformer on Wikipedia rather than use random initialization | Bert · Wikipedia LDA topics, ambiguity threshold 0.001 | same-topic/different-topic embedding cosine similarity ratio: 0.99974 -> 1.13 | Helped |
| Pretrain the transformer on Wikipedia rather than use random initialization | Bert · Wikipedia LDA topics, ambiguity threshold 0.002 | same-topic/different-topic embedding dot product ratio: 1.00027 -> 1.13 | Helped |
Closest papers by meaning
- Mechanism and Emergence of Stacked Attention Heads in Multi-Layer Transformers (ICLR 2025 · similarity 0.86)
- FASTopic: Pretrained Transformer is a Fast, Adaptive, Stable, and Transferable Topic Model (NeurIPS 2024 · similarity 0.86)
- Geometric Factual Recall in Transformers (NeurIPS 2026 · similarity 0.85)
- Towards Understanding How Transformers Learn In-context Through a Representation Learning Lens (NeurIPS 2024 · similarity 0.85)
- Latent Concept Disentanglement in Transformer-based Language Models (ICLR 2026 · similarity 0.85)
- AttnLRP: Attention-Aware Layer-Wise Relevance Propagation for Transformers (ICML 2024 · 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.