AI Paper Atlas

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

Multi-Feature Quantized Self-Attention for Fair Large Language Models

Jaeil Park, Sung-Bae Cho

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Task vectors and steering

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
add MQAR quantized adversarial attention regularizationLlama 3.2 · StereoSet genderICAT: 72.5 -> 86Improved
add MQARLlama 3.2 · StereoSet raceICAT: 68.3 -> 84Improved
add MQARLlama 3.2 · CrowS-Pairs gender (All)bias score: 61.9 -> 48.1Improved
add MQAR on four other backbonesBERT, T5, GPT-Neo, Mixtral · StereoSet genderICAT: 71.2 -> 83.3 (BERT); 67.4 -> 79.9 (T5)Improved
add MQAR for abusive language detectionGPT-Neo · Founta abusive language (original data)AUC: 93.8 -> 93.7No clear effect
add vector quantization to adversarial regularization (ablation)MQAR · StereoSet genderICAT: 67.8 -> 86.0Improved
add adversarial regularizer to the quantized bottleneck (ablation)MQAR · StereoSet genderICAT: 63.9 -> 86.0Improved
keep the reconstruction decoder (ablation)MQAR · Founta abusive language (original data)AUC: 50.3 -> 93.9Improved
keep the discriminator (ablation)MQAR · Founta abusive language (generated data)FPED: 0.392 -> 0.0654Improved
use codebook size K=3 rather than 2, 4 or 8 (ablation)MQAR · StereoSet genderICAT: 82.1, 84.7, 80.3 -> 86.0Improved
use MQAR instead of fine-tuning-based FineDebFineDeb · Llama 3.2 StereoSetlatency per 100 tokens (ms): 189.8 -> 132.7Improved
add MQAR on top of frozen backboneLlama 3.2 · Llama 3.2 StereoSetlatency per 100 tokens (ms): 97.8 -> 132.7Hurt

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