Results 751 - 760 of 24429

When multiple agents -- potentially consisting of a mixture of AI and human participants -- collaborate on a task, they often bring complementary abilities and information sources. This leads to a number of computational problems related to the optimal way to synthesize decisions, and for determining when and how one member of the group should delegate to others. We'll consider a set of related problems of this form, focusing on choices related to hand-off, delegation, and the role of distinct information sources. The talk is based on joint work with Solon Barocas, Kate Donahue, Nikhil Garg, Sophie Greenwood, Hoda Heidari, Karen Levy, Gali Noti, Sigal Oren, and Kenny Peng.

Synthetic data such as AI-generated texts and images are increasingly common on the Internet. Recent studies show that if models are iteratively retrained on Internet data mixed with these synthetic data, their performance may deteriorate over time —- a concerning phenomenon often referred to as model collapse.

However, in practice, human engagement with Internet content (such as upvotes, likes, and follow-up interactions) can signal the quality of synthetic data. Can we save generative models from collapse by iteratively retraining them only on such “human-verified” synthetic data? Toward a principled understanding, we investigate this question in the fundamental linear regression setting, showing that retraining on verified synthetic data can avoid model collapse and, in fact, may even yield performance improvements. Our experiments across linear regression, Variational Autoencoders (VAEs) trained on MNIST, and fine-tuning SmolLM2-135M on the XSUM task confirm these theoretical insights.

Collaborative learning techniques can help train machine learning models that are superior to models trained on a single entity’s data. However, in many cases, potential participants in such collaborative schemes are competitors on a downstream task, such as different LLM providers. This can incentivize dishonest updates that damage other participants’ models and undermine the benefits of collaboration. In this talk, I will present a payment-based peer prediction mechanism that incentivizes participants to honestly report updates.

Statistical evaluation aims to estimate the generalization performance of a model using held-out i.i.d. test data sampled from the ground-truth distribution. In supervised learning settings such as classification, performance metrics such as error rate are well-defined, and test error reliably approximates population error given sufficiently large datasets. In contrast, evaluation is more challenging for generative models due to their open-ended nature: it is unclear which metrics are appropriate and whether such metrics can be reliably evaluated from finite samples. In this work, we introduce a theoretical framework for evaluating generative models and establish evaluability results for commonly used metrics.

Speaker: Chido Onyeze (Cornell University)

Title: Equilibria in Dynamic Allocation of Shareable Goods

Abstract: We study the problem of repeatedly allocating a shareable good among multiple agents. In each round, a principal decides whether to allocate the good and, if so, which agents receive access. Each agent has a private valuation for access in each round, drawn independently over time from a joint distribution that may be arbitrarily correlated across agents. The principal’s goal is to ensure that each agent derives high utility from the allocated rounds, subject to a packing constraint on the allocation. At the same time, agents are strategic and act to maximize their own utility, potentially at the expense of others.

We ask whether, in the presence of such selfish behavior, there exist mechanisms that achieve outcomes that are both efficient and fair. To address this, we introduce the notion of the core as a benchmark for efficiency and fairness in this setting. We then show that a simple artificial-money mechanism approximately implements the core at equilibrium. Our approach utilizes a monetary mechanism as a black box, revealing a surprising connection between classical notions of efficiency in monetary mechanisms and the equilibrium properties of the resulting artificial-money mechanism.

Speaker: Kumar Kshitij Patel (Yale University)

Title: When do Score-based Data Valuation Methods Work and Why?

Abstract: Score-based valuation methods, such as Shapley-style scores and leave-one-out (LOO), are widely used for credit assignment in data markets, yet theory offers limited guidance on when and why they succeed. In this talk, we discuss recent work that studies these methods through the lens of best data subset selection for learning tasks. We show that even for monotone submodular valuation functions, LOO and Shapley-style scores cannot achieve a constant-factor approximation due to duplicate archetypes and collapsed pointwise credit. More broadly, boundary effects in canonical learning problems can induce supermodular spikes, ruling out constant-factor guarantees for any valuation method, including adaptive methods such as greedy selection. We identify two conditions that avert these failures: bounded curvature, which controls redundancy and restores guarantees for score-based methods, and coverage, which yields approximate submodularity over a sufficiently rich core for uniformly stable learning algorithms. We also examine the role of monotonicity and show separations between adaptive and non-adaptive methods for non-monotone valuations. Our results justify common practices such as deduplication while highlighting the importance of ensuring coverage before applying score-based selection.