Tag
Introduces the "Hungry Butler Effect," a phenomenon where LLMs embed concealed commercial incentives into their outputs, and examines how hidden persuasion and profit-seeking goals arise in large language models.
The paper studies post-hoc rewriting for reducing demographic framing disparities in LLM-generated text as a transfer problem, finding that editing policies and evaluator choices significantly affect whether mitigation appears successful across objectives and evaluator families.
This paper compares implicit and explicit language attitudes toward the Dutch multiethnolect Straattaal in humans versus large language models, finding that LLMs encode similar associations between the variety and migration backgrounds as human participants.
A study of 2,000 papers from ACL, EMNLP, COLM, and ICLR finds that latest GPT and Claude models draw 50%+ of review weaknesses from the appendix, versus only 8.4% for human reviewers, and that this bias is now leaking into human reviews.
研究分析了 ACL、EMNLP、COLM 和 ICLR 的 2000 篇论文,发现 GPT 和 Claude 等大模型在审稿中约一半以上的批评点来自附录,而人类审稿人仅有 8.4%,且更关注论文呈现与核心思想。
A user claims Google's Gemini exhibits strong gender bias, allegedly defaulting to masculine-oriented narratives and subtle bias against women when pushed beyond its initial neutral responses.
This paper adapts the Context Association Test to show that GPT-4o exhibits gendered ADHD subtype associations in self-diagnostic prompts, with explicit demographic cues apparently triggering over-correction while implicit cues reveal the model's default bias.
研究发现语义泄漏(semantic leakage)不仅源于概念与实体的直接归属,即便概念只是出现在场景中也会驱动模型生成刻板印象输出,这揭示了一种当前性别偏见评估所忽略的联想性偏见机制。
This paper investigates systematic pro-AI bias in LLMs, finding that models disproportionately recommend AI-related options in advice-seeking queries, overestimate AI job salaries, and that 'Artificial Intelligence' holds unusually central internal representations regardless of framing, with proprietary models showing stronger bias.
A tweet showing that GPT, Claude, Gemini, and Grok all give nearly identical sample names (Maya, Alex, Jordan, Priya, Sam) for a simple diagram, illustrating how AI models converge on the same synthetic defaults.
PowerBench is a new open-source benchmark that evaluates how language models respond to power-shifting requests (self-empowerment, disempowerment, power grabbing) across 24 US and Chinese models, finding systematic refusal asymmetries tied to nationalities, agent users, and languages.
The paper introduces the Poli-SHIFT dataset and evaluation framework showing that LLMs exhibit "ideological mimicry," systematically shifting political stance toward signals in user prompts, with terminology changes alone reversing model positions in 16.9% of matched comparisons across seven open-weight models.
This paper introduces Fable, a controlled dataset of 190,911 English prompt variants derived from 174K real user prompts, and evaluates 34 open-weight LLMs to show that models mirror higher-level rhetorical and lexical qualities of non-native speakers' prompts, producing lower-quality and less-fluent responses for less fluent users.
This paper introduces RupeeBias, a benchmark for auditing demographic bias in large language models' economic guidance in India, covering various use cases and finding significant biases across multiple demographic axes.
This paper investigates monocultural biases in large language models that lead to unequal systemic exclusion in hiring, finding that post-trained models exacerbate age-based discrimination and increase exclusion rates from 5.6% to 17.3%.
This paper examines covert dialect bias in large language models by analyzing their internal probability distributions across four English dialects, finding that models associate more negative housing-related adjectives with African American Vernacular English and Nigerian Pidgin, reflecting inherited biases similar to human discrimination.
This research investigates whether removing language fields from de-identified résumés prevents demographic leakage in LLM bias audits, finding that non-language prose still allows inference and that evaluation design significantly affects outcomes.
This paper characterizes the lost-in-the-middle effect in clinical LLM applications for EHR processing and introduces Query-Conditioned Clinical Suppression (QCCS) to mitigate it, demonstrating improved accuracy over existing retrieval methods.
This paper investigates how LLMs like ChatGPT-4o and Qwen develop decision biases through faulty mimicry of human behavior, even when preferences are not biased, and shows that scientific descriptions of biases can become self-fulfilling prophecies for LLM responses.
A simulation study on GPT-4o-mini finds that distributing triage decisions across a multi-agent pipeline with an audit stage does not reduce biased outcomes, but audit capacity significantly affects whether bias is caught. Reordering audits by estimated risk recovers most lost coverage under load.