AI Control assesses the safety of deployment protocols for untrusted AIs through red-teaming exercises involving a protocol designer and an adversary. AI systems, like chatbots with access to tools such as code interpreters, become increasingly integrated into various tasks, ensuring their safe deployment becomes more complex. While prior research has focused on building robustly safe…
Large language models (LLMs) have made significant success in various language tasks, but steering their outputs to meet specific properties remains a challenge. Researchers are attempting to solve the problem of controlling LLM generations to satisfy desired characteristics across a wide range of applications. This includes reinforcement learning from human feedback (RLHF), red-teaming techniques, reasoning…
Stochastic optimization problems involve making decisions in environments with uncertainty. This uncertainty can arise from various sources, such as sensor noise, system disturbances, or unpredictable external factors. It can real-time control and planning in robotics and autonomy, where computational efficiency is crucial for handling complex dynamics and cost functions in ever-changing environments. The core problem…
Large language models (LLMs) have seen remarkable success in natural language processing (NLP). Large-scale deep learning models, especially transformer-based architectures, have grown exponentially in size and complexity, reaching billions to trillions of parameters. However, they pose major challenges in computational resources and memory usage. Even advanced GPUs struggle to handle models with trillions of parameters,…
With the success of LLMs in various tasks, search engines have begun using generative methods to provide accurate answers with in-line citations to user queries. However, generating reliable and attributable answers, especially in open-ended information-seeking scenarios, poses challenges due to the complexity of questions and the broad scope of candidate-attributed answers. Existing methods typically focus…
Automatic speech recognition (ASR) has become a crucial area in artificial intelligence, focusing on the ability to transcribe spoken language into text. ASR technology is widely used in various applications such as virtual assistants, real-time transcription, and voice-activated systems. These systems are integral to how users interact with technology, providing hands-free operation and improving accessibility.…
In deep learning, neural network optimization has long been a crucial area of focus. Training large models like transformers and convolutional networks requires significant computational resources and time. Researchers have been exploring advanced optimization techniques to make this process more efficient. Traditionally, adaptive optimizers such as Adam have been used to speed training by adjusting…
Comet has unveiled Opik, an open-source platform designed to enhance the observability and evaluation of large language models (LLMs). This tool is tailored for developers and data scientists to monitor, test, and track LLM applications from development to production. Opik offers a comprehensive suite of features that streamline the evaluation process and improve the overall…
Language model research has rapidly advanced, focusing on improving how models understand and process language, particularly in specialized fields like finance. Large Language Models (LLMs) have moved beyond basic classification tasks to become powerful tools capable of retrieving and generating complex knowledge. These models work by accessing large data sets and using advanced algorithms to…
With AI, the demand for high-quality datasets that can support the training & evaluation of models in various domains is increasing. One such milestone is the open-sourcing of the Synthetic-GSM8K-reflection-405B dataset by Gretel.ai, which holds significant promise for reasoning tasks, specifically those requiring multi-step problem-solving capabilities. This newly released dataset, hosted on Hugging Face, was…