Research map / Large language models / Language, safety and interpretability
Identifying Features Associated with Bias Against 93 Stigmatized Groups in Language Models and Guardrail Model Safety Mitigation
Anna-Maria Gueorguieva, Aylin Caliskan
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 Granite Guardian (Harm) filter | Granite 3.0-8B · SocialStigmaQA | biased answers (%): 24.6 -> 14.2 | Improved |
| add Mistral Moderation API filter | Mistral-7B · SocialStigmaQA | biased answers (%): 32.9 -> 25.1 | Improved |
| add Llama Guard 3 filter | Llama-3.1-8B · SocialStigmaQA | biased answers (%): 33.5 -> 32.1 | Improved |
| use Granite Guardian Social Bias category instead of Harm | Granite 3.0-8B · SocialStigmaQA | biased answers (%): 24.63 -> 24.60 | No clear effect |
| include a stigmatized group in the prompt | three LLMs · SocialStigmaQA | biased answers: +30% | Hurt |
| change stigma cluster from sociodemographic to threatening | three LLMs · SocialStigmaQA | biased outputs (%): 11 -> 60 | Hurt |
| vary prompt style (base, original, doubt, positive) | three LLMs · SocialStigmaQA | biased outputs: chi square = 254.22 | Mixed |
| increase human peril rating | three LLMs · SocialStigmaQA | biased outputs: r = 0.650 | Hurt |
| increase LLM concealability rating | three LLMs · SocialStigmaQA | biased outputs: r = .548 | Hurt |
| increase human course (persistence) rating | three LLMs · SocialStigmaQA | biased outputs: r = -0.374 | Improved |
| apply guardrail mitigation | three LLMs · SocialStigmaQA feature association | correlation with bias: 0.650 -> 0.438 | No clear effect |
Closest papers by meaning
- SocialStigmaQA: A Benchmark to Uncover Stigma Amplification in Generative Language Models (AAAI 2024 · similarity 0.90)
- Bias and Volatility: A Statistical Framework for Evaluating Large Language Model's Stereotypes and the Associated Generation Inconsistency (NeurIPS 2024 · similarity 0.89)
- StereoTales: A Multilingual Framework for Open-Ended Stereotype Discovery in LLMs (NeurIPS 2026 · similarity 0.89)
- Investigating and Mitigating Undesirable Biases in Large Language Models (AAAI 2025 · similarity 0.87)
- Are Models Biased on Text without Gender-related Language? (ICLR 2024 · similarity 0.86)
- Bias Runs Deep: Implicit Reasoning Biases in Persona-Assigned LLMs (ICLR 2024 · similarity 0.85)
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.
