Research map / Architectures and efficiency / RNNs and normalization
Language Model Circuits Are Sparse in the Neuron Basis
Aryaman Arora, Zhengxuan Wu, Jacob Steinhardt, Sarah Schwettmann
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 pre-down-projection MLP activations instead of MLP outputs as the neuron basis (ablation) | IG circuit tracing on Llama 3.1 8B · SVA (4 tasks avg) | circuit size at given faithfulness: about 100x smaller circuits | Improved |
| use pre-down-projection MLP activations instead of MLP outputs (ablation) | IG circuit tracing on Gemma-2-2B · SVA (4 tasks avg) | circuit size: approximately 100x reduction | Improved |
| replace Integrated Gradients with RelP attribution (ablation) | MLP-activation neuron circuits · SVA paired | faithfulness/completeness vs circuit size: near-perfect with ~200 neurons | Improved |
| replace IG with RelP attribution (ablation) | residual-stream and MLP-output neuron bases · SVA paired | faithfulness | Improved |
| replace IG with RelP attribution (ablation) | SAE bases · SVA paired | faithfulness/completeness | Mixed |
| use RelP instead of IG with a zero baseline (unpaired data) (ablation) | MLP-activation neuron circuits · SVA unpaired, mean ablation | faithfulness/completeness | Improved |
| use RelP with stop-gradients instead of IG-inp for edge attribution (ablation) | MLP-activation neuron circuits · SVA edge-based circuits | faithfulness: over 80% faithfulness with ~10^5 edges | Improved |
| remove stop-gradients on intermediate MLPs in RelP edge attribution (ablation) | RelP edge attribution · SVA edge-based circuits | faithfulness/completeness | Hurt |
| use RelP instead of IG for gradients in EAP/EAP-IG (ablation) | MIB circuit localization, Llama 3.1 · MIB (5 tasks) | CMD: best on three tasks, second-best on two | Improved |
| increase SAE width (16k to 65k) (ablation) | SAE bases · SVA paired, Gemma-2-2B | circuit sparsity | No clear effect |
| steer single neuron L23/N8079 (say a capital) | Llama 3.1 8B Instruct · state-capitals multi-hop (50 questions) | top output flip | Improved |
| steer each of six neuron clusters in Texas circuit | Llama 3.1 8B Instruct · Texas example | top prediction | Mixed |
Closest papers by meaning
- Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models (ICLR 2025 · similarity 0.91)
- Transcoders find interpretable LLM feature circuits (NeurIPS 2024 · similarity 0.91)
- Automatically Interpreting Millions of Features in Large Language Models (ICML 2025 · similarity 0.89)
- MIB: A Mechanistic Interpretability Benchmark (ICML 2025 · similarity 0.88)
- Extracting Interpretable Task-Specific Circuits from Large Language Models for Faster Inference (AAAI 2025 · similarity 0.87)
- Towards Interpretability Without Sacrifice: Faithful Dense Layer Decomposition with Mixture of Decoders (NeurIPS 2025 · similarity 0.87)
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.
