AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Global Information Thresholding for Sufficient and Necessary Circuits

Jegyeong Cho

CVPR 2026 · Architectures and efficiency › RNNs and normalization · approach: Layer-wise analysis of language models

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 changedWhereMeasuredEffect
replace EAP-IG-inputs top-k budget with retention-calibrated global thresholdEAP-IG circuit extraction · GPT-2 IOIDelta NLL sufficiency (lower is better): 1.009 -> 0.7579Improved
replace EAP-IG-inputs top-k budget with retention-calibrated global thresholdEAP-IG circuit extraction · GPT-2 IOIDelta NLL necessity: 8.328 -> 8.431Improved
replace EAP-IG-inputs top-k budget with retention-calibrated global thresholdEAP-IG circuit extraction · GPT-2 IOIDelta 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 IOIDelta NLL sufficiency (lower is better): 0.7579 -> 2.718Hurt
shrink circuit to 75 edges with a very high threshold (ablation)Ours · GPT-2 IOIDelta NLL necessity: 8.431 -> 5.836Hurt
enlarge circuit to 1500 edges with a lower threshold (ablation)Ours · GPT-2 IOIDelta NLL sufficiency (lower is better): 0.7579 -> 1.138Hurt
enlarge circuit from 750 to 1500 edges (ablation)Ours · GPT-2 IOIDelta NLL necessity: 8.431 -> 8.065Hurt
vary retention target from 0.90 to 0.99 (ablation)Ours · GPT-2 IOICPR: 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 IOICPR across two perturbations: 2.0112/2.0205 at beta 0.0 -> 2.0117/2.0115 at beta 0.2Improved
use a single global threshold instead of manually fixed top-k budget on MCQA (Qwen-2.5)Ours · MIB MCQA Qwen-2.5CPR and CMD: CPR 0.73, CMD 0.14Hurt
use retention-calibrated threshold across tasks and modelsOurs · MIB CPR/CMD across tasks and modelsCPR and CMD: best or second-best on most task-model pairsMixed
retention-calibrated threshold vs EAP-IG-activations on InterpBenchOurs · InterpBenchAUROC: 0.81 -> 0.79No clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy