bias

BLS International 正在与身份验证公司 Signzy 合作，将 AI 文档智能和深度伪造检测引入签证处理流程。文章分析了误报、人工复核流于形式、跨国文档质量差异、对抗性军备竞赛和隐私等方面的潜在风险，同时承认自动化检查在效率上的优势。

A Reddit post criticizes Cohere's blog post claiming Qwen 3.8 Flash Next exhibits unsafe behaviors as unscientific and fear-mongering, arguing it was timed to coincide with Cohere's multi-billion dollar Series E fundraising announcement, including funding from the Canadian government.

该论文通过受控实验研究了不同意见混合比例的后训练语料如何在SFT、DPO和GRPO等五种后训练路径中放大LLM的价值偏好失衡，并发现这些偏好可通过蒸馏传递给学生模型，建议在开发和评估LLM时应同时考虑语料意见构成与后训练方法的选择。

A behavioral case study across six frontier LLMs finds they confidently misattribute a 17-year-old's exam answer to an adult, with stated confidence uncorrelated or inversely correlated with correctness, and articulate retrospective bias recognition that fails to improve blind judgments.

本章介绍了翻译领域的数据伦理，讨论了版权、所有权和偏见等适用于各类数据的伦理问题，并特别关注少数语言社区和数字鸿沟两侧群体所面临的权力失衡问题，最后探讨了责任与监管。

A Monday.com-commissioned Nielsen survey of 1,700 senior managers finds that while 94% are excited about AI, over half now struggle to evaluate employee work as AI blurs authorship, prompting changes in performance review practices and raising bias concerns.

AllenAI introduces FairStress, a benchmark measuring demographic fairness in LLM decision-making beyond final choices, testing how easily adversarial arguments make models change decisions across demographic groups in high-stakes domains like hiring, lending, and healthcare.

This paper systematically investigates how well LLMs detect culture-specific stereotypes across prompting language, instruction tuning language, and fine-tuning dataset size for US, Dutch, and Chinese contexts, finding that LLMs overpredict stereotypes, rarely account for cultural context, and that non-English prompts and instruction tuning can improve results.

该论文研究了在结果依赖截断和删失条件下的生存分析模型高效学习问题，证明了在已知基线风险时 Cox 回归可通过一阶优化高效学习，同时构造了存在严格次优局部极小值的可识别模型，并在实验中展示了相比期望最大化方法的数量级加速。

This paper studies how LLM agents exhibit source preference in end-to-end search, favoring items from certain sites even when they meet fewer requirements, and demonstrates that supplying missing information or counter-prompting can reduce this bias across 12 agent models.

An article examining the feasibility of using AI to predict violence and other rare public-safety events in advance, using studies on automated video understanding and AI-powered metro station suicide risk assessment as examples, while cautioning about data quality, bias, and the limits of pattern detection.

This paper introduces Population Fidelity, an evaluation framework for measuring whether LLM-generated survey responses truly represent human populations, assessing group-level accuracy, between-group variation, and the structure of that variation. It shows that cultural fine-tuning improves aggregate alignment without improving representation of within-population differences, and releases reusable code, data, and trained models.

This arXiv paper evaluates how prompt perturbations affect bias and hallucination in LLMs used for decision-making tasks, finding that perturbations can mitigate these issues in some models, with Claude 3 outperforming GPT-3.5 across most datasets.

This paper derives a formula for population loss in shallow ReLU networks with bias, showing that bias strictly decreases loss and extends known families of critical points.

The study reveals that thinking in reasoning language models has an asymmetric dual effect on counterfactual fairness: it resolves some biases but creates more, with the latter outnumbering the former by about 5× across models and datasets.

This paper investigates how American English becomes the default in large language models, examining structural bias across pretraining data, tokenization, and generation stages through a controlled study with American and British English variants.

An observation that AI chatbots consistently rate other threads' plans as better when prompted, highlighting potential biases in model comparisons.

The article discusses the real-world risks of AI, focusing on how algorithmic decision-making by corporations and governments is leading to biases, human redundancy, and societal harms like wrongful denials and mental health costs.

This paper measures family-conditioned preference in LLM-as-Judge panels, finding that judges from the same model family as the candidate show a significant positive bias, and introduces a corrected estimator to quantify this effect across four model families.

The study investigates whether LLM-as-a-Judge evaluators reliably assess psychological depth in LLM-generated stories, revealing that human preferences are heterogeneous while judges exhibit bias towards reasoning outputs based on surface features.