Artificial intelligence (AI) introduces a new paradigm that challenges established information systems (IS) frameworks, prompting a reassessment of IS use in AI-driven systems. To address this, we conducted a three-iteration multi-approach review of 52 articles centered on AI facets and system-level properties as drivers of AI use. Our review identifies and classifies AI’s facets (anthropomorphism, autonomy, inscrutability, and learning) and system-level properties (explainability, transparency, and reliability). We propose working definitions to resolve conceptual issues. Building on these findings, we inductively develop a framework that positions AI use as a process in which AI facets disrupt or enable user interactions, while system-level properties mediate these effects, individually or in combination. Together, these elements shape established and emerging forms of AI use and outcomes. Leveraging our framework and iterations insights, we critically assess extant understanding and propose a research agenda with actionable paths to support future IS research addressing AI use complexities.
There is much debate as to whether family firms are socially responsible; We believe the question can be informed by changes that occur in managerial priorities across the length of founders’ careers. Drawing on behavioral theories of intertemporal choice, we argue that temporal discounting causes CSR to vary systematically across three succession stages: successor identification, power sharing, and founder exit. CSR related to founder legacy increases when founders identify successors because its socioemotional benefits then become more near-term and salient; subsequently, however, it declines as successors gain influence during power sharing with the founder, and further after founder exit because for successors any CSR legacy benefits are temporally remote and thus more discounted. Evidence from 2,072 listed Chinese family firms supports this multistage transformation, especially among older and first-born founders, and among younger heirs, heirs with shorter tenures, and heirs without children. By viewing legacy concerns through an intertemporal choice lens, we show them to be time-contingent—shaping CSR differently across succession stages and thereby helping reconcile prior findings.
Training-time privileged information (PI) can enable language models to succeed on tasks they would otherwise fail, making it a powerful tool for reinforcement learning in hard, long-horizon settings. However, transferring capabilities learned with PI to policies that must act without it at inference time remains a fundamental challenge. We study this problem in the context of distilling frontier models for multi-turn agentic environments, where closed-source systems typically hide their internal reasoning and expose only action trajectories. This breaks standard distillation pipelines, since successful behavior is observable but the reasoning process is not. We introduce π-Distill, a joint teacher–student framework that trains a PI-conditioned teacher and an unconditioned student simultaneously within a single shared-parameter model, enabling the teacher to learn how to use PI while mitigating distribution shift during transfer. We show that π-Distill effectively distills frontier agents using action-only privileged information, matching or outperforming industry-standard pipelines that assume access to full Chain-of-Thought supervision across multiple agentic benchmarks, models, and forms of PI. We complement our results with extensive analysis that characterize what factors enable effective learning with PI.
Bilevel linear programming (LP) is one of the simplest classes of bilevel optimization problems, yet it is known to be NP-hard in general. Specifically, determining whether the optimal objective value of a bilevel LP is at least as good as a given threshold, a standard decision version of the problem, is NP-complete. However, this decision problem becomes tractable when either the number of lower-level variables or the number of lower-level constraints is fixed, which prompts the question: What if restrictions are placed on the upper-level problem? In this paper, we address this gap by showing that the decision version of bilevel LP remains NP-complete even when there is only a single upper-level variable, no upper-level constraints (apart from the constraint enforcing optimality of the lower-level decision) and all variables are bounded between 0 and 1. This result implies that fixing the number of variables or constraints in the upper-level problem alone does not lead to tractability in general. On the positive side, we show that there is a polynomial-time algorithm that finds a local optimal solution of such a rational bilevel LP instance. We also demonstrate that many combinatorial optimization problems, such as the knapsack problem and the traveling salesman problem, can be written as such a bilevel LP instance.
Large Language Models (LLMs) excel at general-purpose tasks, but personalizing their responses to individual users remains challenging. Retrieval augmentation offers a lightweight alternative to fine-tuning by conditioning LLMs on user history records, yet existing strategies rely on heuristics (e.g., relevance to the query) that overlook the true contribution of records to personalization. Through a systematic motivation study, we show that (i) relevance does not reliably predict utility, and (ii) utility is non-monotonic across records: the best user profile is not simply the combination of the best individual records, and adding more records can even hurt performance. To address these limitations, we propose PURPLE, a contextual bandit framework that oPtimizes UseR Profiles for Llm pErsonalization. PURPLE operates as a re-ranking layer over candidate records, balancing efficiency with personalization quality. Across nine real-world personalization tasks spanning classification, regression, and short- and long-text generation, PURPLE consistently outperforms strong heuristic and retrieval-augmented baselines, establishing contextual bandit retrieval as a principled and scalable solution for personalized LLMs. Our anonymized code is available.
