Research map / Architectures and efficiency / RNNs and normalization
Certified Circuits: Stability Guarantees for Mechanistic Circuits
Alaa Anani, Tobias Lorenz, Bernt Schiele, Mario Fritz, Jonas Fischer
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 |
|---|---|---|---|
| add certification with abstention on unstable components | top-K relevance circuits · ImageNet ID (ResNet-101) | peak cACC / size K: 0.83 -> 0.95 cACC; K 0.700 -> 0.336 | Improved |
| add certification with abstention | top-K relevance circuits · ImageNet-A (ResNet-101) | peak cACC: 0.60 -> 0.94 | Improved |
| add certification with abstention | top-K relevance circuits · OOD-CV (ResNet-101) | peak cACC: 0.73 -> 0.93 | Improved |
| add certification with abstention | EAP-IG circuits · IOI, IOI-Hard, Greater-Than (ID) | peak cACC / edges: 58%, 38%, 75% fewer edges at 1.00 cACC | Improved |
| add certification with abstention | EAP-IG circuits · IOI-Hard (OOD) | peak cACC: 0.85 -> 0.91 | Improved |
| add certification with abstention | EAP-IG circuits · ID language tasks | peak cACC: 1.00 -> 1.00 | No clear effect |
| use activation scorer instead of relevance for certified circuits (ablation) | Certified Circuits · ImageNet-A (ResNet-101) | peak cACC: 0.94 -> 0.36 | Hurt |
| use rank scorer instead of relevance for certified circuits (ablation) | Certified Circuits · ImageNet (ResNet-101) | circuit size reduction: down 52% -> down 17% | Hurt |
| increase certified radius to r=42, 50, 59 on language tasks (ablation) | Certified EAP-IG · GPT-2 Small IOI, IOI-Hard, Greater-Than | cACC | No clear effect |
| increase deletion probability for larger radius | Certified Circuits · small concept datasets | base algorithm quality | Hurt |
| add certification to top-K circuits, rediscovering on shifted data | Certified Circuits · ImageNet vs OOD datasets | IoU of circuit structure | Improved |
| certification across random seeds | Certified Circuits · ImageNet 100 classes | pairwise IoU: mean IoU 0.973 | Improved |
Closest papers by meaning
- The Computational Complexity of Circuit Discovery for Inner Interpretability (ICLR 2025 · similarity 0.90)
- Efficient Automated Circuit Discovery in Transformers using Contextual Decomposition (ICLR 2025 · similarity 0.89)
- Many Circuits, One Mechanism: Input Variation and Evaluation Granularity in Circuit Discovery (NeurIPS 2026 · similarity 0.87)
- EAP-GP: Mitigating Saturation Effect in Gradient-based Automated Circuit Identification (NeurIPS 2025 · similarity 0.87)
- Rethinking Circuit Completeness in Language Models: AND, OR, and ADDER Gates (NeurIPS 2025 · similarity 0.86)
- Transcoders find interpretable LLM feature circuits (NeurIPS 2024 · 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.
