Research map / Architectures and efficiency / Attention and transformers
How Transformers Represent Hierarchies: A Local-to-Global Mechanism
Zhiling Zhou, Tianhao Wang, Zhuoran Yang
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 attention at offset 1 in layer 1 (ablation) | 5-layer Transformer · Balanced full-binary PCFG (depth 4) | generation accuracy: 0.6596 vs 0.9995 | Improved |
| Add attention at offset 2 in layer 2 (ablation) | 5-layer Transformer · Balanced full-binary PCFG (depth 4) | generation accuracy: 0.8220 vs 0.9995 | Improved |
| Add attention at offset 4 in layer 3 (ablation) | 5-layer Transformer · Balanced full-binary PCFG (depth 4) | generation accuracy: 0.9849 vs 0.9995 | Improved |
| Add attention at offset 8 in layer 4 (ablation) | 5-layer Transformer · Balanced full-binary PCFG (depth 4) | generation accuracy: 0.9995 vs 0.9995 | No clear effect |
| Mask off-diagonal offsets in layers 1-3 (ablation) | 5-layer Transformer · Balanced full-binary PCFG (depth 4) | generation accuracy: 0.9995 unchanged | No clear effect |
| Use MLP probe instead of linear probe | Earley-state probing · PCFG probing | probing accuracy | Improved |
| Increase Transformer layer from 1 to 4 | Earley-state probe · PCFG probing (50 grammars) | probing accuracy: exceeding 90% at peak | Improved |
| Go from layer 4 to final layer 5 | Earley-state probe · PCFG probing (50 grammars) | probing accuracy: decreases at final layer | Hurt |
| Increase span window from nonterminal layer 1 to 4 | Span-restricted CYK probe · PCFG probing | probing accuracy at Transformer layer 1: 97.1% -> 63.5% | Hurt |
| Go from layer 4 to layer 5 in span probe (window 1) | Span-restricted CYK probe · PCFG probing | probing accuracy: 99.9% -> 98.6% | Hurt |
Closest papers by meaning
- Hierarchical Autoregressive Transformers: Combining Byte- and Word-Level Processing for Robust, Adaptable Language Models (ICLR 2025 · similarity 0.89)
- Effective Context in Transformers: An Analysis of Fragmentation and Tokenization (NeurIPS 2026 · similarity 0.89)
- Bigram Subnetworks: Mapping to Next Tokens in Transformer Language Models (NeurIPS 2025 · similarity 0.89)
- Distilling Sequential Computation in Transformer Language Models (NeurIPS 2026 · similarity 0.89)
- Towards Understanding the Universality of Transformers for Next-Token Prediction (ICLR 2025 · similarity 0.88)
- Compositional Capabilities of Autoregressive Transformers: A Study on Synthetic, Interpretable Tasks (ICML 2024 · similarity 0.88)
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.
