Our research investigates the capability of modern multimodal reasoning models, powered by Large Language Models (LLMs), to facilitate vision-powered assistants for multi-step daily activities. Such assistants must be able to 1) encode relevant visual history from the assistant's sensors, e.g., camera, 2) forecast future actions for accomplishing the activity, and 3) replan based on the user in the loop. To evaluate the first two capabilities, grounding visual history and forecasting in short and long horizons, we conduct benchmarking of two prominent classes of multimodal LLM approaches -- Socratic Models and Vision Conditioned Language Models (VCLMs) on video-based action anticipation tasks using offline datasets. These offline benchmarks, however, do not allow us to close the loop with the user, which is essential to evaluate the replanning capabilities and measure successful activity completion in assistive scenarios. To that end, we conduct a first-of-its-kind user study, with 18 participants performing 3 different multi-step cooking activities while wearing an egocentric observation device called Aria and following assistance from multimodal LLMs. We find that the Socratic approach outperforms VCLMs in both offline and online settings. We further highlight how grounding long visual history, common in activity assistance, remains challenging in current models, especially for VCLMs, and demonstrate that offline metrics do not indicate online performance.

These first two papers are similar in that they both probe the internals of a neural network by using gradient ascent. In other words, they consider what small changes to the input or to the activations will increase the probability of a predicted class. The first paper applies this to the activations, and the authors report that "it is [possible] to find good qualitative interpretations of high level features. We show that, perhaps counter-intuitively, such interpretation is possible at the unit level, that it is simple to accomplish and that the results are consistent across various techniques."

Deep learning (DL) algorithms have, over the past decade, emerged as the most competitive image recognition algorithms; however, they are by default "black box" algorithms: it is difficult to explain why they make a specific prediction. Why is that an issue? Users of ML models often want the ability to interpret which parts of the image led to the algorithm's prediction for many reasons: Motivated by these use cases, during the last decade, researchers developed many different methods to open the "black box" of deep learning, aiming to make underlying models more explainable. Some methods are specific for certain kinds of algorithms, while some are general. Some are fast, and some are slow.