Research map / Architectures and efficiency / RNNs and normalization
Global Information Thresholding for Sufficient and Necessary Circuits
Jegyeong Cho
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 EAP-IG-inputs top-k budget with retention-calibrated global threshold | EAP-IG circuit extraction · GPT-2 IOI | Delta NLL sufficiency (lower is better): 1.009 -> 0.7579 | Improved |
| replace EAP-IG-inputs top-k budget with retention-calibrated global threshold | EAP-IG circuit extraction · GPT-2 IOI | Delta NLL necessity: 8.328 -> 8.431 | Improved |
| replace EAP-IG-inputs top-k budget with retention-calibrated global threshold | EAP-IG circuit extraction · GPT-2 IOI | Delta Brier / ECE sufficiency: 0.1550 -> 0.1005 (Brier), 0.2010 -> 0.1415 (ECE) | Improved |
| shrink circuit to 75 edges with a very high threshold (ablation) | Ours · GPT-2 IOI | Delta NLL sufficiency (lower is better): 0.7579 -> 2.718 | Hurt |
| shrink circuit to 75 edges with a very high threshold (ablation) | Ours · GPT-2 IOI | Delta NLL necessity: 8.431 -> 5.836 | Hurt |
| enlarge circuit to 1500 edges with a lower threshold (ablation) | Ours · GPT-2 IOI | Delta NLL sufficiency (lower is better): 0.7579 -> 1.138 | Hurt |
| enlarge circuit from 750 to 1500 edges (ablation) | Ours · GPT-2 IOI | Delta NLL necessity: 8.431 -> 8.065 | Hurt |
| vary retention target from 0.90 to 0.99 (ablation) | Ours · GPT-2 IOI | CPR: 2.0041 -> 2.0156 (0.012 total range) | No clear effect |
| add boundary-reinforcement buffer ratio (0.2 or 0.4) (ablation) | Ours · GPT-2 IOI | CPR across two perturbations: 2.0112/2.0205 at beta 0.0 -> 2.0117/2.0115 at beta 0.2 | Improved |
| use a single global threshold instead of manually fixed top-k budget on MCQA (Qwen-2.5) | Ours · MIB MCQA Qwen-2.5 | CPR and CMD: CPR 0.73, CMD 0.14 | Hurt |
| use retention-calibrated threshold across tasks and models | Ours · MIB CPR/CMD across tasks and models | CPR and CMD: best or second-best on most task-model pairs | Mixed |
| retention-calibrated threshold vs EAP-IG-activations on InterpBench | Ours · InterpBench | AUROC: 0.81 -> 0.79 | No clear effect |
Closest papers by meaning
- The Computational Complexity of Circuit Discovery for Inner Interpretability (ICLR 2025 · similarity 0.85)
- Optimal ablation for interpretability (NeurIPS 2024 · similarity 0.83)
- Rethinking Circuit Completeness in Language Models: AND, OR, and ADDER Gates (NeurIPS 2025 · similarity 0.83)
- Transcoders find interpretable LLM feature circuits (NeurIPS 2024 · similarity 0.83)
- EAP-GP: Mitigating Saturation Effect in Gradient-based Automated Circuit Identification (NeurIPS 2025 · similarity 0.83)
- Efficient Automated Circuit Discovery in Transformers using Contextual Decomposition (ICLR 2025 · similarity 0.83)
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.
