🧠 Large Language Models

News about foundation models, LLMs, multimodal models, benchmarks, and model releases from OpenAI, Anthropic, Google, Meta, Mistral, and others.

1,644 articles

Who can veto an enterprise LLM? For 87% of leaders, it's not the budget holder

Many people weigh in on a company's purchase of a large language model (LLM), but far fewer hold the power to reject it. A survey finds that for 87% of leaders, that veto does not belong to whoever controls the budget. Source: learn.g2.com

Brain-inspired AI model tracks hidden cognitive states as people read, word by word

Researchers have built a brain-inspired model that uses a large language model to follow five hidden cognitive states while readers work through complex English text. It traces how comprehension unfolds word by word. Source: bioengineer.org

LLMs lean on Western moral values and overlook other cultures, study finds

Research indicates that AI models tend to favor Western moral frameworks when judging people's actions. This raises concerns about how well they reflect the ethical views of other cultures. Source: wisconsinmuslimjournal.org

Liquid AI launches d1-ultra, a multimodal decision model scoring 65 on AA Decision Index

Liquid AI has released d1-ultra, a decision model that scores 65 on the Artificial Analysis Decision Index. It returns calibrated probabilities without generating any output tokens. Source: marktechpost.com

Guide to LLM training data compliance: copyright, personal data and trade secrets

A comprehensive study examines the intellectual property compliance challenges and legal vulnerabilities involved in training large language models. It also lays out practical steps for mitigating risks tied to copyright, personal information and trade secrets. Source: eu.36kr.com

Chinese team uses homegrown LLM to help crack a major math problem

For the first time, Chinese researchers have relied on a domestically developed LLM to help solve a world-class mathematical problem. The result is being presented as a milestone for China's own AI models. Source: globaltimes.cn

Researchers flag high-risk security flaws in popular LLM gateway tool

Security researchers have discovered serious vulnerabilities in a widely deployed LLM gateway tool that organizations use to connect their systems to language models. The flaws are rated high-risk and put the companies relying on it at potential exposure. Source: infosecurity-magazine.com

NASA builds AI assistant on verified aviation models to curb hallucinations

Researchers at NASA Ames created an LLM-based decision-support system for aviation that routes natural-language questions through nine validated expert models. Grounding answers in verified tools is meant to reduce the risk of hallucinated output. Source: bioengineer.org

Service members to face off against AI in 'Mindgame' tournament

Military service members will compete against AI in a tournament called 'Mindgame'. Participants will also get an after-action review tool built on a large language model. Source: defensescoop.com

Apple's iPhone filing in China lists Apple Intelligence and Tongyi Qianwen

Telecom equipment access permit information published by China's Ministry of Industry and Information Technology on October 14 shows a completed filing for a new iPhone. The listing features Apple Intelligence and Alibaba's Tongyi Qianwen, among other details. Source: eu.36kr.com

Intel launches Inference Microservices for tuned LLM serving on its CPUs

Intel has introduced Inference Microservices, a way to get optimized LLM performance on existing Intel processors. The offering aims to remove days of complicated manual hardware tuning. Source: intel.com

Elon survey: Americans fear AI will drain work of its meaning

A national survey finds worries about fewer jobs, inadequate pay and lost sense of purpose. Even so, workers are embracing AI and expect they will need to adapt. Source: elon.edu

Study tests LLMs for analysing hospital replies to patient safety reports online

Researchers examined whether a large language model can reliably code hospital responses to patient-reported safety incidents. The coding targeted two candidate indicators of a corrective culture. Source: pslhub.org

LLMTransRec blends transformers and LLM embeddings for more accurate, private recommendations

A new hybrid framework, LLMTransRec, pairs transformer-based sequence modeling with LLM embeddings to improve recommendation quality. It is designed to boost accuracy while keeping user data private. Source: bioengineer.org

Vela uses user behavior as a training signal to boost LLM intelligence

Large language models are increasingly used in real-world settings, such as personal assistants that act on a user's long-term behavior. The Vela work explores treating that behavior as a source of supervision to make LLMs smarter. Source: alphaxiv.org

LLMs tested for generating multiple-choice questions for epilepsy education

LLMs are increasingly used to create medical assessment items, but their reliability for disease-specific questions is unclear. This study examines how well they can generate multiple-choice questions for epilepsy education. Source: frontiersin.org

Systematic review: clinical trials of LLM interventions often report incompletely

A systematic review set out to characterise how completely clinical trials of large language model interventions in medicine are reported. The assessment relies on established reporting guidelines. Source: bmjgroup.com

Six LLMs compared on scientific writing in ophthalmology subspecialties

As LLM adoption in healthcare speeds up, evaluating their accuracy has to keep pace. This study compares six large language models on how well they generate scientific and technical writing across ophthalmologic subspecialties. Source: cureus.com

Meta shares rise as Zuckerberg pitches Muse Wonder as a frontier AI model

Meta's stock moved higher after CEO Mark Zuckerberg unveiled the Muse Wonder AI model. He claims the model places the company at the cutting edge of AI development. Source: investors.com

LLM-simulated survey answers poorly mimic human economic preferences in 74 countries

Large-scale surveys are key to studying economic preferences, but they are expensive and slow. The study tests whether responses simulated by large language models can stand in for real respondents, and finds they are weak proxies across 74 countries. Source: nature.com