Topic: bias

A new method called Chain-of-Models aims to audit large language models (LLMs) used as judges to reduce bias, addressing limitations of current mitigation techniques that are either ineffective against various biases or impractical at scale due to reliance on human evaluation.

arxiv.org↗

The article discusses instances where artificial intelligence reaches correct conclusions using flawed reasoning, highlighting concerns about the reliability of AI decisions. This matters because such issues could lead to biased or unethical outcomes in critical applications like healthcare and law enforcement.

Researchers are examining whether large language models understand and convey emotional nuances through different sociodemographic frames, a critical aspect as these models increasingly influence public opinion. This study addresses concerns beyond bias, focusing on how LLMs align with sympathetic or empathetic framing across diverse groups.

arxiv.org↗

The study explores how large language models associate gender with musical instruments, revealing potential biases that could reinforce societal stereotypes. This matters because such biases in widely-used AI systems can inadvertently promote gender stereotypes.

arxiv.org↗

The ggml/llama.cpp project released a new version addressing security issues by moving suppress_tokens handling to common/sampling and removing has_logit_bias, emphasizing improved safety in the latest update.