Summary: New AI and ML models introduced with improved efficiency, performance, and capabilities in various tasks.

From Medium: 2024-07-22 04:50:17

1. RankRAG: A new LLM model significantly outperforms existing ones on knowledge-intensive benchmarks with a fine-tuning framework for effective context ranking and answering generation, using a small ranking dataset.

2. Mixture of A Million Experts: A novel approach efficiently routes to tiny experts using a learned index structure, showing superior efficiency compared to other methods with a parameter-efficient expert retrieval mechanism.

3. Reasoning in Large Language Models: Analysis shows that the density of self-attention graphs in LLMs defines the models’ intrinsic dimension, which impacts their expressive capacity.

4. Lookback Lens: A new approach detects and reduces contextual hallucinations in LLMs, reducing them by 10% in a summarization task through a detection model and decoding strategy.

5. RouteLLM: Improves performance and reduces costs in training LLMs by suggesting router models to balance cost and performance dynamically during inference.

6. Learning to (Learn at Test Time): Proposes new layers for sequence modeling with expressive hidden states that match or outperform existing models while being faster in wall-clock time.

7. Physicochemical graph neural network: Develops a model using physicochemical constraints that predicts protein-ligand interactions directly from sequence data with state-of-the-art performance without costly 3D structures.

8. Meta Platforms to release largest Llama 3 model with 405 billion parameters capable of understanding and generating images and text on July 23.

9. Quora’s Poe now allows users to create interactive web apps directly in chats with AI chatbots, enhancing user engagement and experience in the platform.

10. SmolLM: Introduces a family of state-of-the-art small models with various parameters trained on a high-quality dataset with improved data curation, model evaluation, and usage.



Read more at Medium: AI & ML news: Week 15–21 July. OpenAI and Mistral new models, Andrej… | by Salvatore Raieli | Jul, 2024