AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Secure Outlier-Aware Large Language Model Inference

Lifan Zhao, Zhixuan Fang

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Parallel and early-exit decoding

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 CrypTen non-linear protocols with SOAL (prefixed special tokens, new RMSNorm/SiLU/Softmax protocols)CrypTen · Llama2-7B, 512-token prompttotal time (s): 428.95 -> 193.59Improved
replace CrypTen non-linear protocols with SOALCrypTen · GPT-2, 512-token prompttotal time (s): 48.86 -> 23.30Improved
replace CrypTen non-linear protocols with SOALCrypTen · Mixtral 8x7B, 512-token prompttotal time (s): 1104.46 -> 668.23Improved
add conformant maxima to replace exact maximum finding in Softmax (ablation)SOAL Softmax · Llama2-7B, 512-token promptSoftMax time (s): 199.99 -> 26.62Improved
add scaling secret and outlier removal to normalization (two Newton-Raphson iterations instead of 11) (ablation)SOAL RMSNorm · Llama2-7B, 512-token promptNormalization time (s): 27.86 -> 14.87Improved
add special-token prefixing with new sigmoid protocol for activation (ablation)SOAL SiLU · Llama2-7B, 512-token promptActivation time (s): 87.30 -> 37.13Improved
replace CrypTen normalization with SOAL on GPT-2SOAL RMSNorm · GPT-2, 512-token promptNormalization time (s): 3.07 -> 2.99No clear effect
replace CrypTen normalization with SOAL on Mixtral 8x7B communicationSOAL RMSNorm · Mixtral 8x7B, 512-token promptNormalization communication (GB): 24.31 -> 32.25Hurt
replace Sigma protocols with SOAL under function secret sharingSigma · GPT-2, 512 tokenskey size (GB): 86.686 -> 61.830Improved
replace Sigma protocols with SOAL under function secret sharingSigma · GPT-2, 512 tokensonline time (s): 10.048 -> 7.577Improved
add SOAL prefix tokens and protocols to Llama2-7BLlama2-7B · WikiText-2perplexity: 5.55 -> 5.58No clear effect
add SOAL to GPT-2GPT-2 · WikiText-2perplexity: 36.25 -> 37.50Hurt

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