This paper describes an effort to improve the accuracy of historical U.S. Census transcriptions using a machine learning model. The authors focused on correcting errors in name transcriptions from the 1940 census for Rhode Island, specifically targeting records where independent human transcriptions from Ancestry.com and FamilySearch.org disagreed. The improved transcriptions significantly increased the rate of linking individuals across census records, particularly benefiting records with low original legibility where human transcribers typically struggle. This approach promises to enhance the utility of historical census data for economic and social research by creating a higher quality, linked dataset across multiple periods.

This is from a working paper that analyzes the regulatory distortion in investment incentives within the United States natural gas pipeline network. The authors develop and estimate a structural model to compare the marginal social value of pipeline capacity—tied to regional gas price differences—against the financial incentives of firms operating under fixed rate-of-return regulation. The paper finds that firms' incentives to invest frequently exceed the social value of capital, emphasizing the critical role of the costly regulatory approval process as a secondary tool to control investment and prevent overcapitalization. Ultimately, the authors suggest a welfare-improving reallocation of regulatory costs, recommending streamlining approval in the Northeast while tightening scrutiny in the Southeast and Mountain West regions to address a persistent spatial misallocation of capital.

The document introduces STEER-ME, a new benchmark designed to assess the microeconomic reasoning abilities of Large Language Models (LLMs), specifically focusing on non-strategic settings like supply and demand analysis. To address the limitations of existing benchmarks, the researchers taxonomize microeconomic reasoning into 58 distinct elements, covering areas like consumption decisions, production decisions, and market equilibrium. The benchmark utilizes a novel, automated data generation protocol called auto-STEER to create a large, varied set of multiple-choice questions, mitigating the risk of LLMs overfitting to evaluation data. A case study involving 27 LLMs demonstrated significant performance variation, highlighting that even sophisticated models often rely on shortcuts or produce "near-miss" solutions when faced with complex computational or conceptual tasks.

This paper examines the challenges of conditional ignorability, a key assumption in causal inference used to identify causal effects from observational data. It argues that assessing this assumption is more complex than often perceived, as it implicitly requires evaluating numerous structural configurations within covariate sets. To address this, the authors propose a new framework using Cluster Causal Diagrams (CG(3)), which abstracts the internal structure of covariates into three blocks: treatment (X), outcome (Y), and adjustment covariates (Z). This approach introduces structural ignorability, a concept evaluated using a modified back-door criterion on CG(3) diagrams, offering a more transparent and practical method for assessing causal assumptions. The paper highlights that while conditional ignorability cannot be reliably assessed at this level of abstraction, structural ignorability provides a principled middle ground between the traditional potential outcomes (PO) framework and comprehensive structural causal models (SCMs).

We explore the imminent shift in the digital economy from a defensive model, where websites block automated agents, to an open, transactional "agent economy." This transformation is driven by the realization that website content is a valuable capital good for AI models, leading to a move towards monetized access via APIs. We detail the unsustainable "arms race" between scrapers and blocking technologies, advocating for APIs as an economically rational solution to monetize information goods. It then introduces the rise of autonomous AI agents as new consumers, necessitating machine-centric API design, new platforms (marketplaces, A2A commerce), and evolving business models like Agent-as-a-Service and outcome-based pricing. Finally, We highlight critical challenges, including the need for a micropayment infrastructure, widespread standardization, and robust regulatory frameworks to address data privacy, intellectual property, and potential algorithmic collusion.

This discussion offers a comprehensive examination of the strategic importance of advanced analytics for driving business expansion. It highlights how organizations often struggle to leverage data effectively due to poor data quality, cultural resistance, and a lack of strategic alignment, despite the immense potential for growth. It differentiates between traditional Business Intelligence (BI) and forward-looking advanced analytics, emphasizing techniques like predictive and prescriptive modeling, Machine Learning (ML), and Artificial Intelligence (AI) for uncovering insights and recommending actions. It further illustrates the tangible value of analytics through case studies in customer personalization, marketing ROI optimization, and cross-functional operational improvements, while also addressing the significant challenges and ethical considerations such as data privacy and algorithmic bias. Ultimately, it proposes a phased strategic blueprint for implementing a sustainable analytics capability, focusing on executive alignment, data governance, and fostering a data-informed culture.

This academic paper from Meta Platforms introduces **AdLlama**, a novel large language model (LLM) designed to enhance generative advertising text on Facebook. The core innovation is **Reinforcement Learning with Performance Feedback (RLPF)**, a post-training method that utilizes historical ad performance data, specifically click-through rates (CTR), as a reward signal to fine-tune the LLM. Unlike traditional methods relying on human preferences, RLPF optimizes for measurable real-world outcomes. A large-scale A/B test involving nearly 35,000 advertisers demonstrated that AdLlama significantly improved advertiser-level CTR by 6.7% and increased the number of ad variations created, showcasing the tangible economic impact of this new reinforcement learning approach.

We explore the **evolution of marketing** from AI-assisted to **AI-autonomous functions**, highlighting the profound implications for Chief Marketing Officers (CMOs). We argue that while AI currently boosts efficiency in tactical tasks, the future involves **specialized AI agents** operating with a high degree of autonomy, necessitating a shift in the CMO's role to that of a **systems architect and ethical steward**. We argue the crucial role of **multi-objective reward models** in guiding AI behavior and managing **inherent systemic risks** like algorithmic bias and "reward hacking." Ultimately, we outline a **phased roadmap** for organizations to transition to this advanced, AI-driven marketing ecosystem, stressing the importance of **human oversight and ethical governance**.

We explore **transformative shift in marketing** due to AI, moving from AI-assisted to **AI-autonomous functions**. It highlights that while AI excels at **tactical tasks** like content optimization and performance marketing, **human oversight** remains crucial for strategic areas such as brand management and crisis communication. The text emphasizes the evolving role of the **Chief Marketing Officer (CMO)**, who will become the **architect and ethical steward** of these autonomous systems. This new paradigm necessitates the CMO's understanding of **Reinforcement Learning (RL)** and the critical importance of **designing reward models** to align AI actions with complex, **multi-objective business goals**. Finally, the source addresses significant **risks associated with autonomous AI**, including the "black box" problem, algorithmic bias, and the potential loss of human intuition, while outlining a **strategic roadmap for organizations** to navigate this transition effectively.

We analyze the urgent need for a radical overhaul in marketing education due to the rise of Artificial Intelligence (AI) and Large Language Models (LLMs)**. It argues that traditional MBA programs are facing a **"crisis of relevance," highlighted by student dissatisfaction at institutions like Stanford GSB**, because their curricula fail to address the AI-driven transformation of the marketing industry. The report **details how AI is reshaping every aspect of marketing**, from hyper-personalization to content creation, and **outlines the new "AI-augmented marketer" profile**—a strategic orchestrator requiring skills beyond mere technical proficiency. Finally, it **proposes a comprehensive three-pillar framework for curriculum redesign**, emphasizing a redefined core, advanced specializations, and applied pedagogy to cultivate leaders capable of strategically leveraging AI.

We outline **transformation of Artificial Intelligence (AI) in marketing**, moving from its current role as an **efficiency-boosting tool** to its future as an **autonomous, decision-making agent**. It categorizes this evolution into distinct phases: the present, where AI augments human tasks like content creation, and the impending **"agentic inflection point,"** where AI systems will independently execute complex, multi-step marketing goals. The text also emphasizes the **critical need for educational reform** to prepare marketers for this shift, highlighting new competencies like data fluency and ethical AI application, and it concludes with **strategic recommendations for both businesses and academic institutions** to navigate this profound change.

This research introduces a novel framework for Visual In-Context Learning (VICL), a method where artificial intelligence models learn from provided visual examples. The primary focus is on optimizing the selection of these "in-context examples," which significantly impacts the model's performance on tasks like image segmentation, object detection, and colorization. The authors propose a transformer-based list-wise ranker to identify the most relevant examples, overcoming limitations of previous pair-wise ranking methods that often rely on visual similarity. Furthermore, a consistency-aware ranking aggregator is introduced to synthesize more reliable global rankings from the partial predictions of the ranker. Extensive experiments demonstrate that this new approach consistently outperforms existing methods, leading to state-of-the-art results across various visual tasks.

This paper introduces GEPA (Genetic-Pareto), a novel prompt optimizer designed for large language models (LLMs) and compound AI systems. Unlike traditional reinforcement learning (RL) methods that rely on numerical rewards and extensive "rollouts" (tens of thousands), GEPA leverages natural language reflection to learn high-level rules from trial and error, significantly reducing the required number of rollouts. It achieves this by analyzing system trajectories, diagnosing problems, proposing prompt updates, and combining effective lessons through a Pareto frontier search. This paper presents evidence that GEPA outperforms existing RL and prompt optimization techniques in sample efficiency and generalization across various benchmarks, while also producing shorter, more efficient prompts compared to other methods like MIPROv2.

We introduce and defend the "prediction policy problems" (PPP) framework, which posits that many public policy and economic challenges have an often-overlooked predictive element that machine learning (ML) can significantly enhance. The document addresses key criticisms, arguing that the framework doesn't seek to replace causal inference but rather to improve the predictive "bricks" within complex policy decisions, which inherently include prediction, causal inference, and normative judgment. It emphasizes that accurate prediction is crucial for efficient and equitable resource allocation and that the framework has spurred the development of causal ML methods that integrate prediction with causal analysis. Furthermore, the text contends that challenges like target-construct mismatch and dynamic systems are inherent to quantitative policy analysis and that the PPP framework offers a more transparent and adaptable approach than traditional methods. Finally, it stresses that responsible implementation requires a robust "institutional wrapper" encompassing transparency, human oversight, and contestability, asserting that the proper comparison for algorithmic systems is not perfection, but the often-flawed human-centric status quo.

This academic article critiques the widespread deployment of "predictive optimization" algorithms, which use machine learning to make decisions about individuals based on future predictions. The authors argue that despite claims of accuracy, efficiency, and fairness, these systems inherently fail on their own terms due to seven recurring shortcomings. These issues include inability to translate predictions into optimal interventions, mismatches between intended outcomes and measurable data, biased training data, limitations in predicting social outcomes, unavoidable disparate performance across groups, lack of effective contestability, and vulnerability to strategic manipulation (Goodhart's Law). The research analyzes eight real-world case studies across diverse domains like criminal justice and healthcare, demonstrating that these critiques apply broadly and are not easily resolved through minor design changes, ultimately challenging the legitimacy of such deployments. The paper concludes by providing a rubric of critical questions for assessing these systems and advocating for alternative decision-making approaches.

This research introduces a framework for integrating human expertise into algorithmic predictions, specifically focusing on instances where algorithms deem inputs "indistinguishable." The authors propose a method for selectively incorporating human judgment in these cases, demonstrating its proven ability to enhance the performance of any feasible algorithmic predictor. Empirical studies, including X-ray classification and visual prediction tasks, reveal that even when algorithms generally outperform humans, human input significantly improves predictions on specific, identifiable instances, which can constitute a substantial portion of the data. Furthermore, the paper explores how this framework can lead to algorithms that are robust to varying levels of user compliance, providing near-optimal predictions even when users selectively defer to the algorithm. Ultimately, the work advocates for human-AI collaboration to mitigate algorithmic monoculture by leveraging diverse human perspectives in prediction tasks.

This academic paper investigates the critical issue of data leakage in applying machine learning (ML) to panel data, which combines cross-sectional and time-series observations. The authors explain that standard ML practices, when unsuited for panel data's inherent structure, can lead to temporal leakage (future information affecting past predictions) and cross-sectional leakage (information sharing across training and testing units). This leakage results in inflated model performance and misleading policy recommendations, as empirical applications, particularly for income prediction in U.S. counties, vividly demonstrate. To counter this, the paper offers practical guidelines for practitioners, emphasizing the importance of clearly defining research goals—whether for cross-sectional prediction or sequential forecasting—and implementing appropriate data splitting and cross-validation strategies to ensure robust and realistic ML model evaluation.

This paper introduces the concept of "prediction policy problems," arguing that not all policy decisions require causal inference; many benefit significantly from accurate predictions. The authors distinguish these from traditional "causal inference" problems through examples, such as deciding whether to take an umbrella (prediction) versus whether a rain dance causes rain (causal). They explain how machine learning (ML) excels in prediction by effectively managing the bias-variance trade-off and allowing for flexible models, unlike conventional methods like Ordinary Least Squares (OLS) that prioritize unbiasedness. An illustrative application in healthcare demonstrates how ML can identify and reduce "futile surgeries" by predicting patient mortality, leading to substantial savings and improved patient outcomes. The text concludes by highlighting the widespread applicability and importance of prediction problems across various policy domains, suggesting they warrant greater attention and reorientation in economic research.

We explore the evolving landscape of immersive marketing across Meta's virtual and augmented reality platforms, specifically focusing on Quest headsets and Ray-Ban smart glasses. They detail how advancements in hardware, AI, and sensor technology will enable deeply personalized, context-aware, and interactive advertising experiences that go beyond traditional 2D formats. The texts also highlight the critical ethical and privacy considerations associated with collecting sensitive biometric and first-person data, emphasizing the need for transparency and user trust. Finally, the sources discuss the strategic shifts required for brands and marketers to adapt to these new "dimensional" realities, including developing 3D assets, cultivating new skillsets, and redefining how marketing ROI is measured.

The paper "The Limitations of Large Language Models for Understanding Human Language and Cognition" from "Open Mind: Discoveries in Cognitive Science" argues that Large Language Models (LLMs) offer limited insights into human language and cognition, particularly concerning acquisition and evolution. The authors, Christine Cuskley, Rebecca Woods, and Molly Flaherty, contend that while LLMs can functionally imitate human writing, their underlying mechanisms and developmental processes are fundamentally different from how humans acquire and use language. They employ an ethological "four questions" framework to highlight these distinctions, emphasizing that LLMs lack true meaning, multimodality, and the diverse, interactive aspects characteristic of human language. Ultimately, the report concludes that LLMs should be viewed as tools for specific, carefully constructed research questions rather than comprehensive models for understanding the full scope of human linguistic behavior.