Research map / Label-efficient and robust learning / Recognition and long-tail
Shared Semantics, Divergent Mechanisms: Unsupervised Feature Discovery by Aligning Semantics and Mechanisms
Hyunjin Cho, Youngji Roh, Jaehyung Kim
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 |
|---|---|---|---|
| cluster jointly in semantic and attribution views (RD) instead of K-means in one view | RD-clustering · AmbigQA, Qwen3-8B | mixed distortion Dγ: lowest in 94% of settings | Improved |
| use semantic-only K-means (KM-S) | K-means · AmbigQA, Qwen3-8B | mechanistic distortion | Hurt |
| use attribution-only K-means (KM-A) | K-means · AmbigQA, Qwen3-8B | semantic coherence | Hurt |
| steer with RD cluster medoid direction instead of semantic K-means medoid | cluster steering · AmbigQA, Qwen3-8B | Spearman correlation of steering: -0.02 -> 0.30 | Improved |
| steer with RD medoid instead of a random single continuation's attribution | cluster steering · AmbigQA, Qwen3-8B | Spearman correlation of steering: 0.13 -> 0.30 | Improved |
| use RD medoid instead of single continuation on held-out production-funding cluster | cluster steering · Table 3 case study | held-out centered logit change (amplification): +0.068 -> +0.598 | Improved |
| increase beta from 0.5 to 0.75 to 1.0 | RD-clustering · 'If God was one of us' prompt | number of clusters: K=3 -> K=5 -> K=9 | Mixed |
| lower gamma to 0.1 (weight attribution view) | RD-clustering · AmbigQA qualitative | cluster structure | Mixed |
| raise gamma to 0.9 (weight semantic view) | RD-clustering · AmbigQA qualitative | number of clusters | Mixed |
| add Split operation | RD-clustering · AmbigQA example | rate-distortion objective L_RD: 4.61 -> 4.57 | Improved |
| add Junk operation | RD-clustering · AmbigQA example | rate-distortion objective L_RD: 5.54 -> 5.42 | Improved |
Closest papers by meaning
- Towards Global-level Mechanistic Interpretability: A Perspective of Modular Circuits of Large Language Models (ICML 2025 · similarity 0.88)
- Extracting Interpretable Task-Specific Circuits from Large Language Models for Faster Inference (AAAI 2025 · similarity 0.87)
- Analyze Feature Flow to Enhance Interpretation and Steering in Language Models (ICML 2025 · similarity 0.87)
- Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models (ICLR 2025 · similarity 0.87)
- Concept-Guided Interpretability via Neural Chunking (NeurIPS 2025 · similarity 0.86)
- LLM Layers Immediately Correct Each Other (NeurIPS 2025 · 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.
