Research map / Large language models / Language, safety and interpretability
What Is One Grain of Sand in the Desert? Analyzing Individual Neurons in Deep NLP Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Rank neurons by linguistic-correlation classifier weights and retain the top 10% rather than the bottom 10% (ablation) | NMT English POS classifier · English POS | accuracy: 15.8 -> 69.8 | Helped |
| Rank neurons by linguistic-correlation classifier weights and retain the top 10% rather than the bottom 10% (ablation) | NLM English POS classifier · English POS | accuracy: 18.4 -> 54.2 | Helped |
| Retrain using the top 10% ranked neurons instead of the bottom 10% (ablation) | NMT English POS classifier · English POS | accuracy: 80.6 -> 89.1 | Helped |
| Use the cross-model correlation ranking rather than single-model variance or distance-from-mean rankings (ablation) | NMT · Translation | BLEU | Helped |
| Ablate the top 50 cross-model-ranked neurons rather than the bottom 50 (ablation) | NMT · Translation | BLEU: 15-20 BLEU points versus 0.5 BLEU points | Hurt |
| Use linguistic-correlation rankings instead of top-to-bottom cross-model rankings (ablation) | NMT · Translation | BLEU | Mixed |
| Use semantic-tag neuron ranking rather than POS neuron ranking (ablation) | NMT · Translation | BLEU | Hurt |
| Remove the top-ranked cross-model neurons (ablation) | NLM · Language modeling | perplexity | Hurt |
| Increase elastic-net regularization below the reported 1e^{-4} range (ablation) | English POS linguistic-correlation classifier · English POS | accuracy | No clear effect |
| Retain the top-ranked neurons instead of the bottom-ranked neurons when retraining the classifier (ablation) | NMT and NLM linguistic-correlation classifiers · POS, morphology, and semantic tagging | accuracy | Helped |
| Reduce the representation to the top 10% of ranked neurons (ablation) | NMT and NLM linguistic-correlation classifiers · POS, morphology, and semantic tagging | accuracy | Hurt |
Closest papers by meaning
- Identifying and Controlling Important Neurons in Neural Machine Translation (ICLR 2019 · similarity 0.89)
- Neurons to Words: A Novel Method for Automated Neural Network Interpretability and Alignment (AAAI 2025 · similarity 0.87)
- Automated Natural Language Explanation of Deep Visual Neurons with Large Models (Student Abstract) (AAAI 2024 · similarity 0.86)
- The Emergence of Abstract Thought in Large Language Models Beyond Any Language (NeurIPS 2025 · similarity 0.86)
- Automatically Interpreting Millions of Features in Large Language Models (ICML 2025 · similarity 0.86)
- Capability Localization: Capabilities Can be Localized rather than Individual Knowledge (ICLR 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.