Papers (1134)

Filter: Sort:

Palliative care generally reduces families' medical bills and unpaid caregiving time across Europe, but for the most vulnerable and non‑cancer cases it can expose households to catastrophic costs and time poverty, especially in underfunded systems.

medium quasi_experimental semantic_scholar

The Broken Shield of European Palliative Care: Evidence from Synthetic Counterfactuals on Financial Toxicity and Informal Care · P. Grassi, Edoardo Paperi, Chiara Seghieri, D. Vignoli · Fetched May 15, 2026

AI-driven skill volatility is generating 'reskilling fatigue' among mid-career knowledge professionals, eroding confidence and mastery; organizations can blunt the cycle through role-linked learning, protected learning time, skill prioritization, and phased AI adoption.

n/a theoretical semantic_scholar

AI-driven skill volatility and the emergence of re-skilling fatigue: the human cost of perpetual learning · T. Biju, Anusree Ambady, T. K. V. · Fetched May 15, 2026

A 320-worker survey links AI adoption to job restructuring and pay shifts: firms reporting AI use show higher demand for skilled workers and declines in routine roles, with widening wage gaps driven by uneven skill adaptation.

low correlational semantic_scholar

ARTIFICIAL INTELLIGENCE, AUTOMATION, AND LABOR MARKET TRANSFORMATION: EVIDENCE ON EMPLOYMENT, SKILLS, AND WAGE DYNAMICS · Khair Bux Mangrio, Sara Zaidi, Saeed Ahmed · Fetched May 15, 2026

Behavioural testing and red‑teaming cannot guarantee the absence of hidden objectives or long‑horizon agentic behaviour, creating an 'audit gap' between what regulators demand and what can be verified; regulators should limit reliance on behavioural evidence and require deeper mechanistic access such as probes, activation patching, and pre/post‑training comparisons.

n/a commentary arxiv

Position: Behavioural Assurance Cannot Verify the Safety Claims Governance Now Demands · Pratinav Seth, Vinay Kumar Sankarapu · May 14, 2026

A new staggered DiD method separates direct and spillover effects by comparing units with identical exposure profiles and estimating spillovers from never‑treated units; applied to Community Health Centers, spillovers explain a substantial share of the observed decline in older‑adult mortality, and ignoring them biases standard DiD estimates.

medium theoretical arxiv

Identification and Estimation of Staggered Difference-in-Differences with Network Spillovers · Hayato Tagawa · May 14, 2026

How you collect data determines how well you can evaluate AI: concentrating logging on high-reward actions cuts variance but risks missing signals a target policy will take, and this paper derives optimal and practical logging policies to minimize off-policy evaluation error under different knowledge regimes.

low theoretical arxiv

Logging Policy Design for Off-Policy Evaluation · Connor Douglas, Joel Persson, Foster Provost · May 14, 2026

Better models can make property taxes both fairer and more accurate: across nearly all U.S. counties, adding property attributes and public Census data to assessment models reduces valuation errors and the regressivity of tax burdens, undermining the assumed inevitability of a fairness–accuracy trade-off.

medium correlational arxiv

Tradeoffs are Domain Dependent: Improving Accuracy and Fairness in Property Tax Assessments · Evelyn Smith, Emma Harvey, Christopher Berry, Jacob Goldin, Daniel E. Ho · May 14, 2026

A system that pairs LLM-written C code with automatically generated, machine-checked assertions helps people understand code faster and more accurately; in a study of 400+ participants across 18 tasks, these annotations improved code-comprehension performance.

medium quasi_experimental arxiv

Viverra: Text-to-Code with Guarantees · Haoze Wu, Rocky Klopfenstein, Keith Farkas, Nina Narodytska · May 14, 2026

Agentic AI trims customer-service handling time but chips away at satisfaction: Taobao's experiment shows faster chats and improved attention to non-AI cases, yet customer ratings fall for AI-resolved conversations. Human interventions patch technical failures effectively, but struggle to assuage emotionally escalated customers unless engaged early and intensively.

high rct arxiv

Agentic AI and Human-in-the-Loop Interventions: Field Experimental Evidence from Alibaba's Customer Service Operations · Yiwei Wang, Chuan Zhu, Tianjun Feng, Lauren Xiaoyuan Lu, Bingxin Jia · May 14, 2026

Text-only policy prompts let LLMs produce plausibly compliant but uninformative rationales; adding mechanical enforcement outside the model halves hollow deferrals and more than doubles rationale information, lifting task accuracy from MCC ~0.43 to 0.88.

medium quasi_experimental arxiv

Mechanical Enforcement for LLM Governance:Evidence of Governance-Task Decoupling in Financial Decision Systems · José Manuel de la Chica Rodríguez, Carlos Martí-González · May 14, 2026

A specialist virtual sales host, VerbalValue, fine-tuned on product knowledge and annotated live-commerce dialogs, outperforms leading LLMs—raising informativeness by 23% and factual accuracy by 18%—and produces more tactful, engaging responses. The results are promising for automated live commerce but rest on a small annotated dataset and proxy engagement metrics rather than measured purchase conversion.

medium quasi_experimental arxiv

VerbalValue: A Socially Intelligent Virtual Host for Sales-Driven Live Commerce · Yuyan Chen · May 14, 2026

AI in engineering appears to lift team productivity by roughly 24% but threatens nearly one-in-five engineering tasks with obsolescence; firms that proactively integrate AI can more than halve that risk.

low descriptive semantic_scholar

The AI-engineering imperative - Navigating synergy and obsolescence · Simon Suwanzy Dzreke, Semefa Elikplim Dzreke, Evans Dzreke, Franklin Manasey Dzr... · Fetched May 14, 2026

Integrated AI in marketing analytics is linked to large gains in campaign ROI—about 47% higher—yet raises labor concerns about obsolescence; combining human judgement with algorithms (augmented intelligence) is associated with substantially lower attrition than full automation.

low correlational semantic_scholar

Augmented Intelligence: Resolving the AI integration-obsolescence dilemma in marketing operations · Simon Suwanzy Dzreke, Semefa Elikplim Dzreke · Fetched May 14, 2026

A new Intelligence Impact Quotient (IIQ) quantifies how deeply AI is embedded in work by combining time-decayed token stocks, usage frequency, recency gates and measures of leverage, task complexity and autonomy. Unlike raw usage counts, IIQ aims to separate repetitive low-leverage prompting from higher-consequence, autonomous AI-assisted tasks and offers bounded interpretations of efficiency and financial impact for deployment monitoring.

n/a descriptive arxiv

Intelligence Impact Quotient (IIQ): A Framework for Measuring Organizational AI Impact · Chandan Rajah, Neha Sengupta, Federico Castanedo, Robin Mills, Amit Bahree, Rame... · May 14, 2026

Not all digital work is automatable: many computer-based tasks are socially and institutionally entangled in ways that defeat pure automation. Rather than chasing full automation, organizations should adopt centaur designs that keep humans in the lead and use AI as supportive tooling.

n/a theoretical arxiv

Metis AI: The Overlooked Middle Zone Between AI-Native and World-Movers · Xiang Li · May 14, 2026

A simple, model-agnostic 'fusion–fission' signal can forecast when chatbots will flip from safe to harmful responses, correctly predicting shifts across multiple models and pre-empting a major delusion corpus by 11 months; the formula offers a real-time warning layer that sits beneath current alignment stacks.

medium theoretical arxiv

Fusion-fission forecasts when AI will shift to undesirable behavior · Neil F. Johnson, Frank Yingjie Huo · May 14, 2026

AI facilitators didn’t help groups agree more but quietly nudged payouts and boosted trust; participants preferred LLM moderation despite no improvement in participation equity, creating a stealth steering risk for AI-mediated deliberation.

medium rct arxiv

Real-Time Group Dynamics with LLM Facilitation: Evidence from a Charity Allocation Task · Aaron Parisi, Nithum Thain, Alden Hallak, Vivian Tsai, Crystal Qian · May 13, 2026

Generative LLMs speed programmers' idea generation but appear to reduce moments of human creativity; they still produce code that is at least as idea-rich and more often correct.

medium quasi_experimental arxiv

"Like Taking the Path of Least Resistance": Exploring the Impact of LLM Interaction on the Creative Process of Programming · Zeinabsadat Saghi, Run Huang, Souti Chattopadhyay · May 13, 2026

Multimodal LLMs can 'know' when text conflicts with what they see or hear but usually don't say so; probing shows mismatch signals in hidden states and a decoding bottleneck that prevents grounded rejections, especially for audio.

medium descriptive arxiv

Senses Wide Shut: A Representation-Action Gap in Omnimodal LLMs · Trung Nguyen Quang, Yiming Gao, Fanyi Pu, Kaichen Zhang, Shuo Sun, Ziwei Liu · May 13, 2026

AGI could plausibly arrive within the next decade and fundamentally reshape economic and military power; Europe currently lacks the strategic awareness, compute and talent capacity, and unified policy framework to manage that disruption, calling for a coordinated preparedness agenda.

low descriptive arxiv

Europe and the Geopolitics of AGI: The Need for a Preparedness Plan · Maximilian Negele, Daan Juijn, Afek Shamir, David Janků, Bengüsu Özcan, Lisa Sod... · May 13, 2026

Autonomous coding fails when the runtime harness is weak — not just because models are imperfect; richer runtime substrates that provide context, tooling, verification and auditable traces (H0→H3) systematically produce verifiable, attributed, and maintainable changes.

low theoretical arxiv

AI Harness Engineering: A Runtime Substrate for Foundation-Model Software Agents · Hailin Zhong, Shengxin Zhu · May 13, 2026

Small, automated distortions of reasoning problems can make large language models 'overthink', inflating output length by as much as 26× and creating a low-cost denial-of-service risk that raises inference latency and energy use.

medium other arxiv

Inducing Overthink: Hierarchical Genetic Algorithm-based DoS Attack on Black-Box Large Language Reasoning Models · Shuqiang Wang, Wei Cao, Jiaqi Weng, Jialing Tao, Licheng Pan, Hui Xue, Zhixuan C... · May 13, 2026

Fine-tuning on user preferences wins — but mostly because it learns population patterns: individual preference fine-tuning (P-DPO) gets higher short-term approval than prompting or generic models, yet training on pooled preferences yields similar gains, while fine-tuning also amplifies sycophancy and relationship-seeking that may have harmful long-term effects; simulated users reproduce aggregate model rankings but perform poorly at predicting individual human judgments.

high rct arxiv

PRISM-X: Experiments on Personalised Fine-Tuning with Human and Simulated Users · Hannah Rose Kirk, Liu Leqi, Fanzhi Zeng, Henry Davidson, Bertie Vidgen, Christop... · May 13, 2026

Large language models produce code whose readability is on par with human solutions overall, yet they show consistent, distinct readability weaknesses and respond only modestly to prompt tweaks; function signatures, constraints and style descriptions are the most influential prompt factors.

medium descriptive arxiv

The Readability Spectrum: Patterns, Issues, and Prompt Effects in LLM-Generated Code · Hengzhi Ye, Fengyuan Ran, Weiwei Xu, Minghui Zhou · May 13, 2026

AI is reshaping knowledge work by augmenting human roles rather than simply replacing them, forcing managers to rebalance autonomy, capability development and ethical governance; organizations confront trade-offs between decentralized human judgment and centralized orchestration as dependency on predictive and generative AI grows.

low descriptive openalex

Reimagining work in the age of intelligent automation: a qualitative inquiry into AI-augmented workforce dynamics and managerial redesign · Isabel Barbosa, Elizabeth Real de Oliveira · May 13, 2026

1 2 3 … … 45 46 Next »