BBQ is a new benchmark dataset showing that QA models often default to social stereotypes, achieving up to 3.4 points higher accuracy when the correct answer aligns with bias.
hub
Bryson, and Arvind Narayanan
16 Pith papers cite this work, alongside 2,800 external citations. Polarity classification is still indexing.
16
Pith papers citing it
2,800
external citations · OpenAlex
hub tools
citation-role summary
other 1
citation-polarity summary
roles
other 1polarities
unclear 1representative citing papers
H-SAL erases latent concepts from text profiles using self-descriptions as implicit debiasing signals and shows competitive performance on a new multi-domain Stack Exchange helpfulness benchmark.
Under semantic underdetermination, LLMs cannot always guarantee strong correctness, strict non-bias, and high utility at once.
LLMs contain identifiable COCO neurons that enable implicit self-correction against stereotypes; targeted editing of these neurons improves fairness and robustness to jailbreaks while preserving generation quality.
Counterfactual prompting effects on LLMs are often indistinguishable from those caused by meaning-preserving paraphrases, causing most previously reported demographic sensitivities to disappear under proper statistical comparison.
A methodological framework detects subtle group-associated linguistic biases in LLM outputs by generating controlled synthetic minimal pairs, abstracting n-grams, and ranking high-signal fragments with a PMI variant for expert review.
Reasoning models expend more tokens on association-incompatible tasks than compatible ones, indicating greater effort on counter-stereotypical information, except for Claude 3.7 Sonnet which shows the reverse pattern linked to its bias-focused reasoning.
The authors provide a detailed taxonomy of 21 risks associated with language models, covering discrimination, information leaks, misinformation, malicious applications, interaction harms, and societal impacts like job loss and environmental costs.
Authors release a new 800-sentence gender-balanced profession dataset and use it to test occupational gender stereotypes in three sentiment analysis models.
Contextual embeddings are propagated through WordNet to produce full-coverage sense representations that let a simple k-NN classifier outperform prior neural WSD models.
Relative Probability Association Metric (RPAM) measures LM associations via softmax-normalized continuation probabilities and correlates strongly with human associations and downstream LM behavior across three models.
The paper guides ML use in economic history, identifies systematic prediction bias that distorts coefficients, and shows debiasing via small expert-labeled samples can correct it while preserving scale.
Human-written screenplays pass the Bechdel test more often than those generated by GPT-5, Gemini 3 Pro, and Claude Sonnet 4.5, though network analyses show mixed bias patterns across all script types.
AI alignment should target objective floors of competence, accuracy, honesty, and lawfulness rather than aggregated human preferences.
ReBias-Lens shows LLM self-reflection produces layer-wise smoothing of global valence fluctuations that reduces behavioral bias overall, yet selectively locks in and amplifies certain category-specific biases.
citing papers explorer
-
A CAP-like Trilemma for Large Language Models: Correctness, Non-bias, and Utility under Semantic Underdetermination
Under semantic underdetermination, LLMs cannot always guarantee strong correctness, strict non-bias, and high utility at once.