Research map / Graph learning / Graph algorithms and clustering
Skill Path: Unveiling Language Skills from Circuit Graphs
Hang Chen, Xinyu Yang, Jiaying Zhu, Wenya Wang
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 |
|---|---|---|---|
| remove skill paths of the target skill from the circuit graph (ablation) | GPT2-small circuit graph · PVT/IDT/ICL samples | accuracy vs original output: PVT 1.00 -> 0.01; IDT 1.00 -> 0.00; ICL1 1.00 -> 0.00 | Hurt |
| remove ~500 random paths instead of skill paths (ablation) | GPT2-small circuit graph · PVT/IDT/ICL samples | accuracy vs original output: PVT 1.00 -> 0.23 (random 500); IDT 1.00 -> 0.29 | Mixed |
| combine existing pruning methods with skill-path extraction | ACDC/EAP/DiscoGP/etc. · skill overlap | ovlp(IDT, PVT): ACDC 0.19 -> 0.74; DiscoGP 0.18 -> 0.79 | Improved |
| combine pruning with skill-path extraction | ACDC · skill overlap | ovlp(ICL1, IDT): 0.17 -> 0.63 | Improved |
| use skill paths for unrelated greater-than skill | skill paths · GT vs PVT/IDT/ICL1 | edge overlap: ovlp(GT,PVT) ACDC 0.11 -> 0.68; ovlp(GT,IDT) 0.05 -> 0.13 | Improved |
| use ACDC, DiscoGP, Scrubbing instead of E-pruning, EAP as pruning (ablation) | skill path framework · layer distribution of receivers | stratification | Mixed |
Closest papers by meaning
- Circuit Component Reuse Across Tasks in Transformer Language Models (ICLR 2024 · similarity 0.87)
- Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models (ICLR 2025 · similarity 0.87)
- Efficient Automated Circuit Discovery in Transformers using Contextual Decomposition (ICLR 2025 · similarity 0.87)
- Transformers Struggle to Learn to Search (ICLR 2025 · similarity 0.87)
- Towards Understanding Fine-Tuning Mechanisms of LLMs via Circuit Analysis (ICML 2025 · similarity 0.86)
- Knowledge Circuits in Pretrained Transformers (NeurIPS 2024 · similarity 0.86)
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.
