Large language models in the past have typically relied on some form of reinforcement learning with human feedback (RLHF) to better align model responses with human preferences. However, because of oft-observed instabilities when implementing these RLHF pipelines, various reparameterization techniques have recently been introduced to sidestep the need for separately learning an RL reward model. Instead, directly fine-tuning for human preferences is achieved via the minimization of a single closed-form training objective, a process originally referred to as direct preference optimization (DPO) and followed by several notable descendants. Although effective in certain real-world settings, we introduce new evaluation criteria that serve to highlight unresolved shortcomings in the ability of existing DPO methods to interpolate between a pre-trained reference model and empirical measures of human preferences, as well as unavoidable trade-offs in how low- and high-quality responses are regularized and constraints are handled. Our insights then motivate an alternative DPO-like loss that provably mitigates these limitations. Empirical results serve to corroborate notable aspects of our analyses.

The widespread adoption of large language models (LLMs) has raised concerns about their safety and reliability, particularly regarding their vulnerability to adversarial attacks. In this paper, we propose a novel perspective that attributes this vulnerability to reward misspecification during the alignment process. We introduce a metric ReGap to quantify the extent of reward misspecification and demonstrate its effectiveness and robustness in detecting harmful backdoor prompts. Building upon these insights, we present ReMiss, a system for automated red teaming that generates adversarial prompts against various target aligned LLMs. ReMiss achieves state-of-the-art attack success rates on the AdvBench benchmark while preserving the human readability of the generated prompts. Detailed analysis highlights the unique advantages brought by the proposed reward misspecification objective compared to previous methods.

Continuous-Time Dynamic Graph (CTDG) precisely models evolving real-world relationships, drawing heightened interest in dynamic graph learning across academia and industry. However, existing CTDG models encounter challenges stemming from noise and limited historical data. Graph Data Augmentation (GDA) emerges as a critical solution, yet current approaches primarily focus on static graphs and struggle to effectively address the dynamics inherent in CTDGs. Moreover, these methods often demand substantial domain expertise for parameter tuning and lack theoretical guarantees for augmentation efficacy. To address these issues, we propose Conda, a novel latent diffusion-based GDA method tailored for CTDGs. Conda features a sandwich-like architecture, incorporating a Variational Auto-Encoder (VAE) and a conditional diffusion model, aimed at generating enhanced historical neighbor embeddings for target nodes. Unlike conventional diffusion models trained on entire graphs via pre-training, Conda requires historical neighbor sequence embeddings of target nodes for training, thus facilitating more targeted augmentation. We integrate Conda into the CTDG model and adopt an alternating training strategy to optimize performance. Extensive experimentation across six widely used real-world datasets showcases the consistent performance improvement of our approach, particularly in scenarios with limited historical data.

Programmers frequently engage with machine learning tutorials in computational notebooks and have been adopting code generation technologies based on large language models (LLMs). However, they encounter difficulties in understanding and working with code produced by LLMs. To mitigate these challenges, we introduce a novel workflow into computational notebooks that augments LLM-based code generation with an additional ephemeral UI step, offering users UI scaffolds as an intermediate stage between user prompts and code generation. We present this workflow in BISCUIT, an extension for JupyterLab that provides users with ephemeral UIs generated by LLMs based on the context of their code and intentions, scaffolding users to understand, guide, and explore with LLM-generated code. Through a user study where 10 novices used BISCUIT for machine learning tutorials, we found that BISCUIT offers users representations of code to aid their understanding, reduces the complexity of prompt engineering, and creates a playground for users to explore different variables and iterate on their ideas.

Scheduling university courses is extra challenging when classroom capacities are reduced because of social distancing requirements that are implemented in response to a pandemic such as COVID-19. In this work, we propose an expanded taxonomy of course delivery modes, present an integer program, and develop a course scheduling algorithm to enable all course sections -- even the largest -- to have a significant classroom learning component during a pandemic. Our approach is fair by ensuring that a certain fraction of the instruction in every course section occurs in the classroom. Unlike previous studies, we do not allow rotating attendance and instead require simultaneous attendance in which all students in a section meet in 1-5 rooms at the same time but less often than in a normal semester. These mass meetings, which create opportunities for in-person midterm exams and group activities, are scheduled at high precision across all days of the semester rather than a single, repeating week. A fast heuristic algorithm makes the schedule in an hour. Results: We consider the 1834 in-person course sections, 172 classrooms, and 96 days in the fall 2022 semester at [UniversityXYZ]. If average classroom capacity is reduced by 75% due to a pandemic, our approach still allows at least 25% of the instruction in every section, and more than 49% of all instruction across the entire campus, to be in the classroom. Our method also produces excellent results for regular classroom assignment. Managerial implications: An algorithm based on the principles of fairness and simultaneous attendance can significantly improve university course schedules during a pandemic and in normal times. High-precision schedules that prepare a campus for various pandemic possibilities can be created with minimal administrative effort and activated at a moment's notice before or during a semester if an outbreak occurs.

As more clinical workflows continue to be augmented by artificial intelligence (AI), AI literacy among physicians will become a critical requirement for ensuring safe and ethical AI-enabled patient care. Despite the evolving importance of AI in healthcare, the extent to which it has been adopted into traditional and often-overloaded medical curricula is currently unknown. In a scoping review of 1,699 articles published between January 2016 and June 2024, we identified 18 studies which propose guiding frameworks, and 11 studies documenting real-world instruction, centered around the integration of AI into medical education. We found that comprehensive guidelines will require greater clinical relevance and personalization to suit medical student interests and career trajectories. Current efforts highlight discrepancies in the teaching guidelines, emphasizing AI evaluation and ethics over technical topics such as data science and coding. Additionally, we identified several challenges associated with integrating AI training into the medical education program, including a lack of guidelines to define medical students AI literacy, a perceived lack of proven clinical value, and a scarcity of qualified instructors. With this knowledge, we propose an AI literacy framework to define competencies for medical students. To prioritize relevant and personalized AI education, we categorize literacy into four dimensions: Foundational, Practical, Experimental, and Ethical, with tailored learning objectives to the pre-clinical, clinical, and clinical research stages of medical education. This review provides a road map for developing practical and relevant education strategies for building an AI-competent healthcare workforce.

Rapid progress in general-purpose AI has sparked significant interest in "red teaming," a practice of adversarial testing originating in military and cybersecurity applications. AI red teaming raises many questions about the human factor, such as how red teamers are selected, biases and blindspots in how tests are conducted, and harmful content's psychological effects on red teamers. A growing body of HCI and CSCW literature examines related practices-including data labeling, content moderation, and algorithmic auditing. However, few, if any, have investigated red teaming itself. This workshop seeks to consider the conceptual and empirical challenges associated with this practice, often rendered opaque by non-disclosure agreements. Future studies may explore topics ranging from fairness to mental health and other areas of potential harm. We aim to facilitate a community of researchers and practitioners who can begin to meet these challenges with creativity, innovation, and thoughtful reflection.

AI systems make decisions in physical environments through primitive actions or affordances that are accessed via API calls. While deploying AI agents in the real world involves numerous high-level actions, existing embodied simulators offer a limited set of domain-salient APIs. This naturally brings up the questions: how many primitive actions (APIs) are needed for a versatile embodied agent, and what should they look like? We explore this via a thought experiment: assuming that wikiHow tutorials cover a wide variety of human-written tasks, what is the space of APIs needed to cover these instructions? We propose a framework to iteratively induce new APIs by grounding wikiHow instruction to situated agent policies. Inspired by recent successes in large language models (LLMs) for embodied planning, we propose a few-shot prompting to steer GPT-4 to generate Pythonic programs as agent policies and bootstrap a universe of APIs by 1) reusing a seed set of APIs; and then 2) fabricate new API calls when necessary. The focus of this thought experiment is on defining these APIs rather than their executability. We apply the proposed pipeline on instructions from wikiHow tutorials. On a small fraction (0.5%) of tutorials, we induce an action space of 300+ APIs necessary for capturing the rich variety of tasks in the physical world. A detailed automatic and human analysis of the induction output reveals that the proposed pipeline enables effective reuse and creation of APIs. Moreover, a manual review revealed that existing simulators support only a small subset of the induced APIs (9 of the top 50 frequent APIs), motivating the development of action-rich embodied environments.

Integrating AI-driven tools in higher education is an emerging area with transformative potential. This paper introduces Iris, a chat-based virtual tutor integrated into the interactive learning platform Artemis that offers personalized, context-aware assistance in large-scale educational settings. Iris supports computer science students by guiding them through programming exercises and is designed to act as a tutor in a didactically meaningful way. Its calibrated assistance avoids revealing complete solutions, offering subtle hints or counter-questions to foster independent problem-solving skills. For each question, it issues multiple prompts in a Chain-of-Thought to GPT-3.5-Turbo. The prompts include a tutor role description and examples of meaningful answers through few-shot learning. Iris employs contextual awareness by accessing the problem statement, student code, and automated feedback to provide tailored advice. An empirical evaluation shows that students perceive Iris as effective because it understands their questions, provides relevant support, and contributes to the learning process. While students consider Iris a valuable tool for programming exercises and homework, they also feel confident solving programming tasks in computer-based exams without Iris. The findings underscore students' appreciation for Iris' immediate and personalized support, though students predominantly view it as a complement to, rather than a replacement for, human tutors. Nevertheless, Iris creates a space for students to ask questions without being judged by others.

Reinforcement learning (RL) and causal modelling naturally complement each other. The goal of causal modelling is to predict the effects of interventions in an environment, while the goal of reinforcement learning is to select interventions that maximize the rewards the agent receives from the environment. Reinforcement learning includes the two most powerful sources of information for estimating causal relationships: temporal ordering and the ability to act on an environment. This paper examines which reinforcement learning settings we can expect to benefit from causal modelling, and how. In online learning, the agent has the ability to interact directly with their environment, and learn from exploring it. Our main argument is that in online learning, conditional probabilities are causal, and therefore offline RL is the setting where causal learning has the most potential to make a difference. Essentially, the reason is that when an agent learns from their {\em own} experience, there are no unobserved confounders that influence both the agent's own exploratory actions and the rewards they receive. Our paper formalizes this argument. For offline RL, where an agent may and typically does learn from the experience of {\em others}, we describe previous and new methods for leveraging a causal model, including support for counterfactual queries.