Blog by Ines Almeida

When it comes to logical reasoning, LLMs still struggle with simple math. Researchers from Carnegie Mellon University propose combining neural networks with symbolic programming to address this.

Machine translation has improved immensely thanks to AI, but handling multiple languages remains tricky. Researchers from Tel Aviv University and Meta studied this challenge. Through systematic experiments, they uncovered what really causes the most interference.

Researchers from the National University of Singapore and Noah's Ark Lab (Huawei) have developed a new training method that slashes memory requirements with no loss in performance. Their method, called CAME, can match state-of-the-art systems like BERT while using over 40% less memory.

Backpack models have an internal structure that is more interpretable and controllable compared to existing models like BERT and GPT-3.

Weakly supervised learning is a popular technique. The goal is to alleviate the costly data annotation bottleneck. But new research suggests these methods may be significantly overstating their capabilities.

Recent advances in AI language models like GPT-4 and Claude 2 have enabled impressively fluent text generation. However, new research reveals these models may perpetuate harmful stereotypes and assumptions through the narratives they construct.

Researchers at Carnegie Mellon University and University of Lisbon systematically studied when context is needed for high-quality translation across 14 languages.

Measuring the political leaning of various pretrained LMs.

Emily Bender, Alex Hanna, and Hannah Zeavin discuss misguided uses of AI in the mental health area.

While AI can now generate passable jokes, truly understanding humor remains difficult. Key challenges include perceiving incongruous imagery, resolving indirect connections to captions, and mastering the complexity of human culture/experience.

A key challenge in AI is enabling systems to learn from just a few examples, like humans can. One technique that helps is showing the AI systems answered examples to guide its reasoning, called demonstration learning.

Summarization is a key capability for many Generative AI systems. New research explores how better selection of training data can enhance the relevance and accuracy of summaries.

New research explores how to make AI conversational agents more polite using a technique called "hedging."

A new survey from Gartner has found that the availability of generative AI systems like ChatGPT is quickly becoming a top concern for enterprise risk management.

Prosecraft, a data-driven project designed to analyze word usage and provide statistics on various writing-style markers, has recently come under fire from authors and subsequently shut down its operations.

Researchers at Anthropic recently published a paper demonstrating a constitutional AI technique. Their goal was to make the assistants helpful, while avoiding harmful, dangerous, or unethical content.

A new study published in the peer-reviewed Conference on Fairness, Accountability, and Transparency (FAccT 2023) analyzed the presence and impact of Big Tech companies in AI research, specifically focusing on natural language processing (NLP) which powers many language-based AI like chatbots.

Researchers from Tsinghua University and Wuhan University propose teaching AI systems to mimic authorial styles.

In a paper titled "What the DAAM: Interpreting Stable Diffusion Using Cross Attention", researchers propose a method called DAAM (Diffusion Attentive Attribution Maps) to analyze how words in a prompt influence different parts of the generated image.

What is RLHF?