Research map / Large language models / Language, safety and interpretability
Semantic Regexes: Auto-Interpreting LLM Features with a Structured Language
Angie Boggust, Donghao Ren, Yannick Assogba, Dominik Moritz, Arvind Satyanarayan, Fred Hohman
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 |
|---|---|---|---|
| replace GPT-4o-mini with GPT-4o as explainer/evaluator (ablation) | semantic-regex · GPT-2-RES-25k (520 features) | purity: 0.68 ± 0.32 -> 0.74 ± 0.30 | No clear effect |
| replace GPT-4o-mini with GPT-4o as explainer/evaluator (ablation) | semantic-regex · GPT-2-RES-25k (520 features) | clarity: 0.69 ± 0.37 -> 0.68 ± 0.36 | No clear effect |
| replace GPT-4o-mini with GPT-4o as explainer/evaluator (ablation) | token-act-pair · GPT-2-RES-25k (520 features) | clarity: 0.45 ± 0.37 -> 0.59 ± 0.37 | Improved |
| replace GPT-4o-mini with GPT-4o as explainer/evaluator (ablation) | token-act-pair · GPT-2-RES-25k (520 features) | purity: 0.73 ± 0.26 -> 0.84 ± 0.21 | Improved |
| use semantic regex instead of token-act-pair descriptions | token-act-pair · GPT-2-RES-25k | clarity: 0.45 ± 0.36 -> 0.68 ± 0.36 | Improved |
| use semantic regex instead of max-acts descriptions | max-acts · GPT-2-RES-25k | faithfulness: 0.52 ± 0.46 -> 0.37 ± 0.45 | Hurt |
| use semantic regex instead of max-acts descriptions | max-acts · GPT-2-RES-25k | purity: 0.78 ± 0.26 -> 0.66 ± 0.32 | Hurt |
| use semantic regex instead of max-acts descriptions | max-acts · GPT-2-RES-25k | clarity: 0.70 ± 0.35 -> 0.68 ± 0.36 | No clear effect |
| use semantic regex instead of max-acts descriptions | max-acts · Gemma-2-2B-RES-65k (100 features/layer) | clarity: 0.45 ± 0.40 -> 0.47 ± 0.42 | No clear effect |
| give participants semantic regex instead of natural language descriptions | max-acts · user study (24 participants, 12 features) | decision boundary understanding: as well or higher on 9 of 12 features | Mixed |
Closest papers by meaning
- LatentLens: Revealing Highly Interpretable Visual Tokens in LLMs (ICML 2026 · similarity 0.87)
- Capturing Polysemanticity with PRISM: A Multi-Concept Feature Description Framework (NeurIPS 2025 · similarity 0.86)
- LLM-wrapper: Black-Box Semantic-Aware Adaptation of Vision-Language Models for Referring Expression Comprehension (ICLR 2025 · similarity 0.86)
- Sparsity-Guided Holistic Explanation for LLMs with Interpretable Inference-Time Intervention (AAAI 2024 · similarity 0.86)
- Concept Bottleneck Large Language Models (ICLR 2025 · similarity 0.85)
- LLM Layers Immediately Correct Each Other (NeurIPS 2025 · 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.
