Research map / Architectures and efficiency / RNNs and normalization
Efficient Automated Circuit Discovery in Transformers using Contextual Decomposition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces contextual decomposition for transformers (CD-T), an analytic method that decomposes activations into relevant and irrelevant parts to find attention-head circuits in GPT-2 and a toy transformer much faster than ACDC and EAP.
Problem
Automated circuit discovery via activation patching or its approximations is slow, approximate, and needs non-zero gradients, and cannot easily find fine-grained circuits at specific sequence positions.
Approach
Extends contextual decomposition (previously for LSTMs/RNNs/CNNs) to transformer modules including self-attention, giving exact relevant/irrelevant decompositions of node activations; an iterative algorithm ranks source attention heads (optionally per sequence position) by relevance to target nodes and prunes greedily, with no training or hand-crafted examples.
Key results
- CD-T achieves an average of 97% ROC AUC against manual circuits on IOI, Greater-than and Docstring, outperforming ACDC and EAP under low runtime.
- CD-T circuits are reported to be 80% more faithful than random circuits of up to about 60% of the original model size for IOI and Greater-than.
- CD-T reaches faithfulness 1 with fewer nodes than EAP on all three tasks, and reduces runtime from hours (ACDC) to seconds.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CD-T instead of ACDC or EAP for circuit discovery | Transformer circuit discovery · IOI, Greater-than, Docstring | ROC AUC: 97\% average | Helped |
| Use CD-T instead of ACDC | Transformer circuit discovery · IOI, Greater-than, Docstring | Runtime: hours to seconds | Helped |
| Use CD-T circuits rather than random circuits | Circuit faithfulness · IOI and Greater-than | Faithfulness relative to random circuits: 80\% more faithful; up to 60\% of the original model size | Helped |
| Use CD-T circuits rather than random circuits | Circuit faithfulness · Docstring | Faithfulness relative to random circuits: 80\% more faithful; up to 80\% of the original model's size | Helped |
| Use CD-T circuits instead of EAP circuits | Circuit discovery · IOI, Greater-than, Docstring | Nodes needed to reach faithfulness = 1 | Helped |
| Use mean-ablation instead of zero-ablation (ablation) | CD-T circuits · Greater-than and Docstring | Faithfulness | Helped |
| Use mean-ablation instead of zero-ablation (ablation) | CD-T circuits · IOI | Faithfulness | No clear effect |
| Add three heads to their circuit (ablation) | Docstring circuit · Docstring | Task-specific metric: 2.884 -> 3.524 | Helped |
| Add three heads to their circuit (ablation) | Docstring circuit · Docstring | Correct guess rate: 0.54 -> 0.66 | Helped |
Surprising findings
- Mean-ablation yields higher faithfulness than zero-ablation on Greater-than and Docstring, though IOI results are comparable.
- Manual circuits outperform both CD-T and EAP at equal circuit size, indicating room for improvement.
Limitations
- Evaluated only on small models and three tasks with attention-head circuits; MLPs not included in discovery.
- Relevance metric ignores sign of contribution, so it can select negative heads without distinguishing direction.
- Heuristic choices (number of heads per iteration, percentile thresholds) require tuning.
Experimental setup
Datasets: IOI (indirect object identification) Greater-than Docstring
Benchmarks: IOI Greater-than Docstring
Metrics: ROC AUC runtime faithfulness hypothesis-test probability of being more faithful than rand
Baselines: ACDC EAP
Builds on: Contextual Decomposition ACDC Edge Attribution Patching Path patching
Closest papers by meaning
- Finding Transformer Circuits With Edge Pruning (NeurIPS 2024 · similarity 0.91)
- Transcoders find interpretable LLM feature circuits (NeurIPS 2024 · similarity 0.90)
- The Computational Complexity of Circuit Discovery for Inner Interpretability (ICLR 2025 · similarity 0.89)
- EAP-GP: Mitigating Saturation Effect in Gradient-based Automated Circuit Identification (NeurIPS 2025 · similarity 0.88)
- Circuit Component Reuse Across Tasks in Transformer Language Models (ICLR 2024 · similarity 0.88)
- Beyond Components: Singular Vector-Based Interpretability of Transformer Circuits (NeurIPS 2025 · similarity 0.88)
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.