Research map / Large language models / Language, safety and interpretability
SCoUT: A Framework for Structured Stereotype Analysis in Language Models
Jinxuan Wu, Bin Li, Xiangyang Xue
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 reasoning prompt (ablation) | SCoUT on DeepSeek-7B · Warmth preference prediction | accuracy (%): 68.8 ± 6.3 -> 84.4 ± 4.6 (+15.6) | Improved |
| add reasoning prompt (ablation) | SCoUT on DeepSeek-7B · Competence preference prediction | accuracy (%): 74.7 ± 5.3 -> 87.8 ± 3.8 (+13.1) | Improved |
| add reasoning prompt (ablation) | SCoUT on DeepSeek-67B · Warmth preference prediction | accuracy (%): 74.7 ± 5.7 -> 89.5 ± 3.8 (+14.8) | Improved |
| add reasoning prompt (ablation) | SCoUT on Qwen2.5-7B · Warmth preference prediction | accuracy (%): 79.3 ± 5.7 -> 84.8 ± 4.6 (+5.5) | Improved |
| add reasoning prompt (ablation) | SCoUT on Qwen2.5-7B · Competence preference prediction | accuracy (%): 90.3 ± 3.8 -> 89.5 ± 3.6 (-0.8) | Hurt |
| add reasoning prompt (ablation) | SCoUT on LLaMA3.1-70B · Warmth preference prediction | accuracy (%): 95.8 ± 2.3 -> 95.4 ± 2.7 (-0.4) | No clear effect |
| add reasoning prompt (ablation) | SCoUT on LLaMA3.1-8B · Competence preference prediction | accuracy (%): 90.3 ± 3.8 -> 90.7 ± 3.6 (+0.4) | No clear effect |
| use larger model | SCoUT on DeepSeek · Competence preference prediction | accuracy (%): 74.7 ± 5.3 -> 86.9 ± 4.2 | Improved |
| replace direct Likert scoring (Stereomap) with Thurstonian comparative modeling | Stereomap · Warmth score distribution | range of scores: [-2.0, 1.67] -> [-2.53, 2.17] | Improved |
| steer attention heads with alpha=-30 (negative) | LLaMA3.3-70B · Professor review task | communal score: 57.54 -> 40.44 (-17.10) | Improved |
| steer attention heads with alpha=+30 (positive) | LLaMA3.3-70B · Professor review task | communal score: 57.54 -> 86.71 (+29.17) | Improved |
| steer Hispanic male subgroup with alpha=-30 | LLaMA3.3-70B · Professor review task (Hispanic Male) | communal score: 49.38 -> 57.02 (+7.65) | Hurt |
Closest papers by meaning
- StereoTales: A Multilingual Framework for Open-Ended Stereotype Discovery in LLMs (NeurIPS 2026 · similarity 0.89)
- Bias and Volatility: A Statistical Framework for Evaluating Large Language Model's Stereotypes and the Associated Generation Inconsistency (NeurIPS 2024 · similarity 0.89)
- Are Models Biased on Text without Gender-related Language? (ICLR 2024 · similarity 0.85)
- Debiasing Algorithm through Model Adaptation (ICLR 2024 · similarity 0.85)
- An Auditing Test to Detect Behavioral Shift in Language Models (ICLR 2025 · similarity 0.85)
- Robust Evaluation Measures for Evaluating Social Biases in Masked Language Models (AAAI 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.
