THE DECODER

1/ Google researchers have developed a new method called SALT that speeds up the training of large language models by up to 28 percent, while improving their performance by using smaller AI models as assistant teachers. 2/ The method works in two stages: First, the large model learns from the smaller model through knowledge distillation, with the smaller model helping in areas where it can already make good predictions. The large model is then trained conventionally. 3/ In tests, a 2.8 billion-parameter model trained with SALT achieved the same performance as a conventionally trained model in just 70 percent of the usual training time, and even outperformed it after further fine-tuning, particularly in arithmetic and text comprehension.

1/ Researchers from Renmin University of China, Tsinghua University, and Huawei Poisson Lab have developed RetroLLM, an AI system that integrates information search and text generation into a single process, offering improved efficiency compared to existing solutions. 2/ RetroLLM generates clues from the given question, then employs advanced search techniques such as "Constrained Beam Search" and "Forward-Looking Constrained Decoding" to identify relevant information, which is continuously incorporated during the answer generation process. 3/ In evaluations, RetroLLM demonstrated significantly better performance than existing systems, achieving 10 to 15 percent higher accuracy on question-answering tasks, with particularly strong results on more complex "multi-hop" questions that require multiple steps of reasoning.

1/ Researchers at Johns Hopkins University have developed GenEx, an AI system that can generate a fully explorable 3D environment from a single photograph, allowing robots and AI agents to move freely within the generated space. 2/ GenEx enables a wide range of applications, including the generation of bird's-eye views, multi-view videos, and 3D maps. 3/ The system also supports decision-making by AI agents through "imaginary exploration," which has led to significantly higher accuracy in traffic decision tests compared to using a single original image.

1/ Meta has introduced ExploreToM, a framework designed to generate diverse and challenging data for evaluating the theory-of-mind (ToM) understanding of large language models (LLMs), as previous datasets are often too simplistic and may overestimate the models' capabilities. 2/ Current top models, including Llama-3.1-70B, Mixtral 7x8B, and GPT-4o, struggle with the complex ToM scenarios generated by ExploreToM, with their accuracy dropping to as low as 0% for Mixtral and Llama and up to 9% for GPT-4o in the tests. 3/ The study reveals that LLMs have difficulties with simple state-tracking, a crucial skill for ToM reasoning, and improving state-tracking could be a key step in equipping language models with better ToM capabilities.

1/ Researchers have developed a new approach called PRIME that helps AI models learn mathematics more efficiently, delivering better results while using only a fraction of the training data required by other methods. 2/ The PRIME-trained model, Eurus-2-7B-PRIME, outperformed larger models like GPT-4o and Llama-3.1-70B-Instruct across mathematical benchmarks, with a significant improvement of 16.7 percentage points compared to its predecessor, Qwen 2.5 Math 7B. 3/ PRIME provides continuous feedback throughout the problem-solving process using "implicit process rewards," requiring only 230,000 training examples and four solution attempts per problem to achieve better results than the Qwen2.5-Math-7B-Instruct model, which needed 2.5 million examples and 32 attempts. 👇 Read more #AItraining #GenerativeAI

1/ Researchers from Google Deepmind, Columbia University and UC San Diego have developed an AI system called CAT4D that can generate dynamic 3D scenes from ordinary videos. 2/ CAT4D uses a novel multi-view video diffusion model, trained on a mixture of real and synthetic data, to generate multiple views from different angles of a video and compute a changing 3D reconstruction. 3/ The technology could have applications in areas such as game development, film and augmented reality, although the system still struggles with temporal extrapolation beyond the input frames. 👇 Read more #Deepmind #GenerativeAI

1/ Anthropic and major music publishers have reached an agreement prohibiting the AI assistant Claude from generating copyrighted song lyrics. 2/ The deal requires Anthropic to put safeguards in place and promptly address any reports of system failures from publishers. However, the underlying issue of whether Anthropic has the right to use copyrighted data, such as song lyrics, to train AI remains unresolved. 3/ OpenAI has announced plans for a "media manager" to allow rights holders to exclude their content from AI training, but has not provided any further updates on this matter since May 2024.

1/ Google's Logan Kilpatrick predicts that AI vision capabilities will become mainstream by 2025, while AI agents may require additional development time until 2026. 2/ Microsoft's AI CEO, Mustafa Suleyman, believes that the current 80 percent accuracy of AI agents is insufficient for user confidence, and that a 99 percent accuracy rate is needed for widespread adoption. That could take two more generations of models. 3/ Anthropic suggests starting with simple prompts and moving to more complex agent systems as needed.

1/ Nvidia significantly increased its startup investments in 2024, investing $1 billion across 50 funding rounds and several corporate deals, up from $872 million and 39 rounds in 2023. 2/ The company focused on "core AI" companies that require substantial computing power, many of which are already Nvidia customers, and made acquisitions including Israeli AI workload management platform Run:ai. 3/ Regulators are closely examining whether Nvidia's dominant position and large investments might be pushing for exclusivity, but the company maintains there are no strings attached to its funding and that it aims to grow its ecosystem and support companies. 👇 Read more #Investments #Nvidia

1/ Researchers from Peking University, the Shanghai AI Laboratory, and Nanyang Technological University have developed DiffSensei, an AI system that can automatically turn written stories into manga-style comics while maintaining consistent character appearances and controlling page layouts. 2/ DiffSensei combines diffusion models with large language models to handle both the visual and narrative elements of manga creation. It generates manga in three steps: creating page layouts, drawing the characters, and adding dialogue, using a custom dataset called MangaZero containing over 43,000 annotated manga pages. 3/ Although DiffSensei struggles with unclear character references and generic artwork without specific style references, the researchers believe it could help streamline manga production by providing artists, publishers, and creators with a new tool for making personalized manga stories while maintaining control over characters and layouts. 👇 Read more #AIandart #GenerativeAI