Research map / Large language models / Language, safety and interpretability
Towards Understanding and Mitigating Social Biases in Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Apply autoregressive INLP instead of the original language model | GPT-2 · StereoSet Religion | SS: 58.02 -> 54.57 | Helped |
| Apply autoregressive INLP instead of the original language model | GPT-2 · StereoSet Religion | LM: 88.46 -> 89.13 | Helped |
| Apply autoregressive INLP instead of the original language model | GPT-2 · Gender Respect | Negative regard difference: 0.160 -> 0.003 | Helped |
| Apply autoregressive INLP instead of the original language model | GPT-2 · Gender Occupation | Negative regard difference: 0.084 -> 0.034 | Helped |
| Apply autoregressive INLP instead of the original language model | GPT-2 · Human evaluation Religion | Fairness: 3.93 -> 4.00 | Helped |
| Apply autoregressive INLP instead of the original language model | GPT-2 · Human evaluation Religion | Clarity: 4.97 -> 4.93 | Hurt |
| Apply autoregressive INLP instead of the original language model | GPT-2 · Human evaluation Religion | Content: 4.99 -> 4.93 | Hurt |
| Train the bias classifier with diverse contexts rather than simple contexts only (ablation) | Bias classifier · Bias classifier Diverse contexts | Accuracy: 53.6 -> 61.2 | Helped |
| Add subsequences to simple and diverse classifier training contexts (ablation) | Bias classifier · Bias classifier Diverse contexts | Accuracy: 61.2 -> 63.7 | Helped |
| Add subsequences to simple and diverse classifier training contexts (ablation) | Bias classifier · Bias classifier Sub-sequences | Accuracy: 60.4 -> 62.5 | Helped |
| Use autoregressive INLP rather than INLP applied at every token | GPT-2 · StereoSet Religion | LM: 82.83 -> 89.13 | Helped |
| Use autoregressive INLP rather than INLP applied at every token | GPT-2 · StereoSet Religion | ICAT: 73.04 -> 80.97 | Helped |
Closest papers by meaning
- SocialStigmaQA: A Benchmark to Uncover Stigma Amplification in Generative Language Models (AAAI 2024 · similarity 0.91)
- Biases Mitigation and Expressiveness Preservation in Language Models: A Comprehensive Pipeline (Student Abstract) (AAAI 2024 · similarity 0.90)
- Investigating and Mitigating Undesirable Biases in Large Language Models (AAAI 2025 · similarity 0.90)
- The Devil is in the Neurons: Interpreting and Mitigating Social Biases in Language Models (ICLR 2024 · similarity 0.90)
- Are Models Biased on Text without Gender-related Language? (ICLR 2024 · similarity 0.90)
- More of the Same: Persistent Representational Harms Under Increased Representation (NeurIPS 2025 · similarity 0.89)
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.