Research map / Architectures and efficiency / Attention and transformers
Quantifying LLM Attention-Head Stability: Implications for Circuit Universality
Karan Bali, Jack Stanley, Praneet Suresh, Danilo Bzdok
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 |
|---|---|---|---|
| use AdamW instead of Adam for pretraining (ablation) | 8- and 12-layer MLP GPT models and all attention-only models · attention-head seed stability across refits | head stability | Improved |
| use AdamW instead of Adam for pretraining (ablation) | 2- and 4-layer MLP GPT models · attention-head seed stability across refits | head stability | No clear effect |
| use AdamW instead of Adam for pretraining (ablation) | GPT-like models · wikitext-2 validation perplexity | validation perplexity | No clear effect |
| use AdamW instead of Adam for pretraining (ablation) | GPT-like models · stability gap between most and least stable layer | stability gap | Improved |
| use AdamW instead of Adam for pretraining (ablation) | GPT-like models · positions of most and least stable layer | relative depth of stability extremes | No clear effect |
| use AdamW instead of Adam for pretraining (ablation) | GPT-like models · residual stream stability (CKA) | residual stream CKA stability | No clear effect |
| use AdamW instead of Adam for pretraining (ablation) | GPT-like models · attention-head output activation norm | mean l2 norm of head outputs | Improved |
| increase depth of the model (ablation) | GPT-like models · stability gap across layers | stability gap | Hurt |
| increase prompt length from 5 to 50 tokens (ablation) | middle-layer attention heads · layer-wise head stability | head stability | Hurt |
| allow cross-layer matching of heads (ablation) | stability metric · head stability across refits | head stability | Improved |
| larger mean query-weight norm in a layer | layers of GPT-like models · layer-wise head stability | head stability | Hurt |
| compare residual stream instead of attention heads | GPT-like refits · cross-seed stability | stability | Improved |
Closest papers by meaning
- Circuit Component Reuse Across Tasks in Transformer Language Models (ICLR 2024 · similarity 0.89)
- From Condensation to Rank Collapse: A Two-Stage Analysis of Transformer Training Dynamics (NeurIPS 2025 · similarity 0.88)
- LLM Circuit Analyses Are Consistent Across Training and Scale (NeurIPS 2024 · similarity 0.88)
- Differentiation and Specialization of Attention Heads via the Refined Local Learning Coefficient (ICLR 2025 · similarity 0.88)
- Causal Head Gating: A Framework for Interpreting Roles of Attention Heads in Transformers (NeurIPS 2025 · similarity 0.88)
- Transcoders find interpretable LLM feature circuits (NeurIPS 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.
