Research map / Architectures and efficiency / Attention and transformers
Influence Patterns for Explaining Information Flow in BERT
Kaiji Lu, Zifan Wang, Piotr Mardziel, Anupam Datta
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 |
|---|---|---|---|
| keep only embedding-level influence-pattern nodes (GPR) vs. random nodes of the same count (ablation) | BERT ablated to pattern nodes · SVA/RA/SA, all subtasks | ablated accuracy: SVA Obj. .55 -> .96; SA .52 -> .97 | Improved |
| keep GPR embedding patterns instead of attention-weight patterns (ablation) | BERT ablated to pattern nodes · SVA/RA/SA | ablated accuracy: SVA WSC .49 -> 1.0; SA .47 -> .97 | Improved |
| keep GPR embedding patterns instead of internal-influence patterns (ablation) | BERT ablated to pattern nodes · SVA/RA/SA | ablated accuracy: SVA WSC .50 -> 1.0; SA .49 -> .97 | Improved |
| keep GPR embedding patterns instead of conductance-based patterns (ablation) | BERT ablated to pattern nodes · SVA/RA/SA | ablated accuracy: RA NA .71 -> .98; SVA Obj. .93 -> .96 | Improved |
| conductance baseline is strong, close to GPR on SA (ablation) | conductance patterns vs GPR · SA SST-2 | ablated accuracy: SA .94 -> .97 | No clear effect |
| replace each skip node in the attention-level pattern with all A attention heads (ablation) | Pi^a_+ · SVA/RA/SA | ablated accuracy: SVA Obj. .70 -> .50; SA .86 -> .48 | Hurt |
| attention-level GPR pattern versus random attention-level nodes (ablation) | BERT ablated to pattern nodes · SVA/RA/SA | ablated accuracy: RA GA .50 -> .89; SVA Obj. .50 -> .70 | Improved |
| ablating to Pi^e_+ nodes versus the full unablated model (ablation) | BERT-small and BERT-base · RA NA / RA GA | accuracy: RA NA .83 -> .98; RA GA .73 -> .99 | Improved |
| ablating to Pi^e_+ nodes versus the full unablated model on SVA (ablation) | BERT-small · SVA Obj./Subj./WSC/APP | accuracy: SVA Obj. .96 -> .96; WSC 1.0 -> 1.0 | No clear effect |
| random-node baseline variability over 50 runs (ablation) | Pi^e_rand · SA SST-2 | ablated accuracy: SA .52 +/- .030 | No clear effect |
| skip connections versus attention heads in attention-level patterns | Pi^a · all tasks | share of nodes in patterns: 75.4% of nodes in Pi_a are skip | Improved |
| word like versus love/hate as clausal verb in SVA-Obj (SP) | BERT-small · SVA-Obj (SP) | accuracy: like .82; hate .99; love 1.0 | Hurt |
Closest papers by meaning
- Talking Heads: Understanding Inter-Layer Communication in Transformer Language Models (NeurIPS 2024 · similarity 0.84)
- Training Dynamics of Transformers to Recognize Word Co-occurrence via Gradient Flow Analysis (NeurIPS 2024 · similarity 0.83)
- A Phase Transition between Positional and Semantic Learning in a Solvable Model of Dot-Product Attention (NeurIPS 2024 · similarity 0.83)
- Pathways of Visual Information Flow in Vision-Language Models (NeurIPS 2026 · similarity 0.83)
- Abrupt Learning in Transformers: A Case Study on Matrix Completion (NeurIPS 2024 · similarity 0.82)
- Unveiling Induction Heads: Provable Training Dynamics and Feature Learning in Transformers (NeurIPS 2024 · similarity 0.82)
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.
