Research map / Probabilistic and scientific modeling / Molecules and proteins
From Directions to Regions: Decomposing Activations in Language Models via Local Geometry
Or David Shafran, Shaked Ronen, Omri Fahn, Shauli Ravfogel, Atticus Geiger, Mor Geva
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 MFA components instead of SAE features for DBM localization | DBM · RAVEL (Gemma-2-2B, Llama-3.1-8B) | accuracy: 3-16 points | Improved |
| use MFA components instead of SAE features for DBM localization | DBM · MCQA | accuracy: Gemma 64.9 -> 80.2; Llama 65.6 -> 70.5 | Mixed |
| restrict MFA to centroids only (remove loadings) (ablation) | MFA + DBM · RAVEL (Gemma-2-2B) | accuracy: Continent 85.7 -> 86 | No clear effect |
| restrict MFA to centroids only (remove loadings) (ablation) | MFA + DBM · MCQA (Gemma-2-2B) | accuracy: 80 -> 39 | Hurt |
| steer with MFA centroids instead of SAE features | Gemma-2-2B · steering (concept + fluency) | harmonic mean score: roughly doubles median | Improved |
| steer with MFA centroids instead of SAE features | Llama-3.1-8B · steering (concept + fluency) | harmonic mean score: about one third median gain | Improved |
| increase MFA components from 1K to 32K (ablation) | MFA · steering | harmonic mean score | No clear effect |
| increase number of MFA components K (ablation) | MFA · region and loading annotation | fraction of narrow regions | Mixed |
| use MFA instead of SAE for decomposition | MFA vs SAE · Pile activations | interpretability fraction: 0.29 -> 0.96 | Improved |
Closest papers by meaning
- Exploring Activation Patterns of Parameters in Language Models (AAAI 2025 · similarity 0.86)
- Invariant Features in Language Models: Geometric Characterization and Model Attribution (NeurIPS 2026 · similarity 0.86)
- When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search (NeurIPS 2026 · similarity 0.86)
- Towards Best Practices of Activation Patching in Language Models: Metrics and Methods (ICLR 2024 · similarity 0.86)
- Less is More: Local Intrinsic Dimensions of Contextual Language Models (NeurIPS 2025 · similarity 0.85)
- Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior (NeurIPS 2026 · similarity 0.85)
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.
