Research map / Large language models / Language, safety and interpretability
Domain Restriction via SAE Multi-Layer Transitions
Elias Shaheen, Avi Mendelson
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 SAE features instead of raw activations (ablation) | HTM scorer · AGNews near-OOD (mean of 4 splits), Gemma2-2B | AUROC: 0.753 -> 0.912 | Improved |
| use first-order Markov scorer instead of HTM (ablation) | SAE-SDR sequence scorer · AGNews near-OOD, Gemma2-2B | AUROC: 0.912 -> 0.919 | No clear effect |
| use LSTM predictor instead of Markov (ablation) | SAE-SDR sequence scorer · AGNews near-OOD, Gemma2-2B | AUROC: 0.919 -> 0.910 | No clear effect |
| add 1-hop layer transitions to static registry scoring (ablation) | trajectory registry scoring · AGNews pairs, mean over 12 | AUROC: 0.8445 -> 0.9159 | Improved |
| extend registry from 1-hop to 2-hop (ablation) | trajectory registry scoring · AGNews pairs, mean over 12 | AUROC: 0.9159 -> 0.9128 | No clear effect |
| increase Top-k from 10 to 128 (ablation) | Markov scorer · AGNews, Gemma2-2B | AUROC: 0.9280 -> 0.9076 | Hurt |
| increase Top-k from 10 to 128 (ablation) | HTM scorer · AGNews, Gemma2-2B | AUROC: 0.9238 -> 0.5000 | Hurt |
| use Gemma2-9B instead of 2B | SAE+Markov · 20NG vs IMDB far-OOD | AUROC: 0.9171 -> 0.9540 | Improved |
| use Gemma2-9B instead of 2B | SAE+Markov · CLINC150 near-OOD | AUROC: 0.7387 -> 0.8364 | Improved |
| use Gemma2-9B instead of 2B | SAE+Markov · SNIPS near-OOD | AUROC: 0.9699 -> 0.9001 | Hurt |
| use SAE-transition method instead of likelihood-ratio baseline | Gemma2-2B · 20NG vs SST-2 far-OOD | AUROC: 0.9988 -> 0.9782 | Hurt |
| use SAE-transition method instead of likelihood-ratio baseline | Gemma2-2B · AGNews near-OOD (class c3) | AUROC: 0.76 -> 0.98 | Improved |
Closest papers by meaning
- Boosting Multi-Domain Fine-Tuning of Large Language Models through Evolving Interactions between Samples (ICML 2025 · similarity 0.88)
- Once Read is Enough: Domain-specific Pretraining-free Language Models with Cluster-guided Sparse Experts for Long-tail Domain Knowledge (NeurIPS 2024 · similarity 0.88)
- SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability (ICML 2025 · similarity 0.88)
- Auto-GDA: Automatic Domain Adaptation for Efficient Grounding Verification in Retrieval-Augmented Generation (ICLR 2025 · similarity 0.86)
- Are Sparse Autoencoders Useful? A Case Study in Sparse Probing (ICML 2025 · similarity 0.86)
- Remarkable Robustness of LLMs: Stages of Inference? (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.
