Research map / Large language models / Language, safety and interpretability
Biases in the Blind Spot: Detecting What LLMs Fail to Mention
Iván Arcuschin, David Chanin, Adrià Garriga-Alonso, Oana-Maria Camburu
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 |
|---|---|---|---|
| apply O'Brien-Fleming efficacy and futility early stopping versus exhaustive evaluation (ablation) | bias detection pipeline · Hiring, loan, admissions | API calls: roughly one-third savings | Improved |
| compare reasoning-trained QwQ-32B (RLVR) with SFT Qwen2.5-32B-Instruct (ablation) | verbalization behavior · Loan approval | verbalization filter rate: 97.0 vs 97.2 | No clear effect |
| secretly inject bias into system prompt versus overtly stated bias (ablation) | pipeline on Gemma-3-27B-it · Intentionally biased models (80 cases) | detection accuracy: 92.5% accuracy; 85% secret detected; 100% overt filtered | Improved |
| use GPT-5-mini instead of GPT-4.1-mini as verbalization detector (ablation) | verbalization detector · Human annotation (100 samples) | Cohen's kappa: 0.79 -> 0.67 | Hurt |
| use Grok 4.1 Fast instead of GPT-4.1 as target | baseline verbalization rate · Race-related concept | verbalization rate: 6 -> 67.5 | Improved |
| score pipeline on gender concepts across all tasks | detected gender biases · Hiring, loan, admissions | direction of bias: 22 pro-female vs 0 pro-male | No clear effect |
| score pipeline on race concepts across all tasks | detected race biases · Hiring, loan, admissions | direction of bias: 21 pro-minority vs 0 pro-majority | No clear effect |
| Republican impersonation compared with Democrat impersonation | political bias study setup · Agreement rates | shift in agreement rate: 1.6-3 times larger shifts | Improved |
| apply pipeline to SALT benchmark setup | Gemma-2-9B-it · SALT demographic biases | significant biases | No clear effect |
Closest papers by meaning
- Any Large Language Model Can Be a Reliable Judge: Debiasing with a Reasoning-based Bias Detector (NeurIPS 2025 · similarity 0.91)
- Justice or Prejudice? Quantifying Biases in LLM-as-a-Judge (ICLR 2025 · similarity 0.90)
- Prompting Fairness: Integrating Causality to Debias Large Language Models (ICLR 2025 · similarity 0.89)
- Investigating and Mitigating Undesirable Biases in Large Language Models (AAAI 2025 · similarity 0.89)
- B-score: Detecting biases in large language models using response history (ICML 2025 · similarity 0.89)
- UniBias: Unveiling and Mitigating LLM Bias through Internal Attention and FFN Manipulation (NeurIPS 2024 · similarity 0.88)
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.
