Research map / Large language models / Language, safety and interpretability
Multi-Feature Quantized Self-Attention for Fair Large Language Models
Jaeil Park, Sung-Bae 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 |
|---|---|---|---|
| add MQAR quantized adversarial attention regularization | Llama 3.2 · StereoSet gender | ICAT: 72.5 -> 86 | Improved |
| add MQAR | Llama 3.2 · StereoSet race | ICAT: 68.3 -> 84 | Improved |
| add MQAR | Llama 3.2 · CrowS-Pairs gender (All) | bias score: 61.9 -> 48.1 | Improved |
| add MQAR on four other backbones | BERT, T5, GPT-Neo, Mixtral · StereoSet gender | ICAT: 71.2 -> 83.3 (BERT); 67.4 -> 79.9 (T5) | Improved |
| add MQAR for abusive language detection | GPT-Neo · Founta abusive language (original data) | AUC: 93.8 -> 93.7 | No clear effect |
| add vector quantization to adversarial regularization (ablation) | MQAR · StereoSet gender | ICAT: 67.8 -> 86.0 | Improved |
| add adversarial regularizer to the quantized bottleneck (ablation) | MQAR · StereoSet gender | ICAT: 63.9 -> 86.0 | Improved |
| keep the reconstruction decoder (ablation) | MQAR · Founta abusive language (original data) | AUC: 50.3 -> 93.9 | Improved |
| keep the discriminator (ablation) | MQAR · Founta abusive language (generated data) | FPED: 0.392 -> 0.0654 | Improved |
| use codebook size K=3 rather than 2, 4 or 8 (ablation) | MQAR · StereoSet gender | ICAT: 82.1, 84.7, 80.3 -> 86.0 | Improved |
| use MQAR instead of fine-tuning-based FineDeb | FineDeb · Llama 3.2 StereoSet | latency per 100 tokens (ms): 189.8 -> 132.7 | Improved |
| add MQAR on top of frozen backbone | Llama 3.2 · Llama 3.2 StereoSet | latency per 100 tokens (ms): 97.8 -> 132.7 | Hurt |
Closest papers by meaning
- Interpretable Debiasing of Vision-Language Models for Social Fairness (CVPR 2026 · similarity 0.87)
- MABR: Multilayer Adversarial Bias Removal Without Prior Bias Knowledge (AAAI 2025 · similarity 0.86)
- The Devil is in the Neurons: Interpreting and Mitigating Social Biases in Language Models (ICLR 2024 · similarity 0.86)
- Investigating and Mitigating Undesirable Biases in Large Language Models (AAAI 2025 · similarity 0.86)
- Language Bias in LVLMs: From In-Depth Analysis to Simple and Effective Mitigation (ICML 2026 · similarity 0.86)
- Towards Understanding the Word Sensitivity of Attention Layers: A Study via Random Features (ICML 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.
