The work introduces LocalFACE, a model-agnostic method for generating feasible and actionable counterfactual explanations using only locally acquired information. It improves privacy and model security by accessing only the data and model regions needed to construct a personalized recourse path.

The paper defines engagement in digital interventions as a state of simultaneous physical, affective, and cognitive energy investment in a task, distinct from adherence and motivation. It introduces the AIM-ACT framework, grounded in neuroscience, to explain how engagement unfolds moment to moment and to translate that into concrete design strategies.

The paper introduces PaperBanana, an agentic framework that automates the generation of publication-ready academic illustrations using vision-language and image generation models. Five specialized agents retrieve references, plan content and style, render figures, and refine them through self-critique, outperforming leading baselines in faithfulness, conciseness, readability, and aesthetics.

The paper introduces CGM-LSM, a Transformer decoder-based large sensor model pretrained on 1.6 million continuous glucose monitoring records from patients spanning different diabetes types, ages, and genders. By modeling patients as sequences of glucose time steps, it forecasts a two-hour horizon and reduces one-hour prediction error by 48.51% over prior methods.

The paper introduces TWOSOME, an online framework that deploys large language models as decision-making agents aligned with embodied environments through reinforcement learning, without prepared datasets or prior knowledge of the environment. Action probabilities queried from the model form the behavior policy, and a frozen LLM with low-rank adapters is trained using PPO.

This paper proposes a novel approach for detecting insulin pump faults (IPFs) by combining the ability of a long short term memory (LSTM) autoencoder to extract features with the strength of random forest to distinguish between anomalous and normal patterns.

The paper evaluates offline reinforcement learning as a way to develop effective insulin dosing policies without the potentially dangerous patient interaction required during online training. It benchmarks BCQ, CQL, and TD3-BC on the 30 virtual patients of the FDA-approved UVA/Padova simulator, including challenging scenarios such as incorrect bolus dosing and irregular meal timings.

The talk reviews two widely used deep reinforcement learning methods, Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC). It contrasts their on-policy and off-policy formulations, the role of the clipped surrogate objective and entropy regularization, and the trade-offs each makes between training stability and sample efficiency.

The paper introduces CaLoRAify, a vision-language framework that aligns ingredient recognition with calorie estimation by training on visual-text pairs. Built on CalData, a 330K image-text dataset, it applies low-rank adaptation and retrieval-augmented generation to estimate calories from a single food image while retaining conversational interaction.

The paper proposes a knowledge-infused conversational health agent for diabetes patients, built on the open-source openCHA framework and augmented with the American Diabetes Association dietary guidelines and the Nutritionix database. Analytical tools compute nutritional intake and compare it against those guidelines, outperforming GPT-4 across 100 questions on daily meal choices.

The paper introduces NLGraph, a benchmark of 29,370 problems spanning eight graph reasoning tasks described in natural language, from connectivity and shortest path to maximum flow. Evaluating GPT-3 and GPT-4, it finds preliminary graph reasoning ability that degrades on harder tasks, and proposes Build-a-Graph and Algorithmic prompting to narrow the gap.

The paper introduces LaFTer, which narrows the gap between zero-shot vision-language classifiers and supervised fine-tuning without any labels or paired vision-language data. A classifier is first trained on category descriptions auto-generated by an LLM, then adapted to an unlabeled image collection through pseudo-labeling, yielding up to 11.7% absolute improvement.

The paper proposes EmojiCrypt, a prompt encryption scheme that obfuscates private data with emojis and other non-linguistic symbols before it is sent to a cloud-based LLM. Across eight datasets it preserves, and sometimes improves, task performance while resisting simulated inference attacks, without requiring any change to the underlying model.

The paper introduces HuggingGPT, an LLM-powered agent that uses ChatGPT as a controller coordinating expert models hosted on Hugging Face. It follows a four-stage pipeline of task planning, model selection from function descriptions, task execution, and response generation, allowing complex multimodal requests to be decomposed and solved end to end.