The Commonplace
Home Papers Evidence Explore Trends Syntheses Digests About 🎲 Workforce Futures

Papers (1747)

Filter: Sort:
A student-built benchmark exposes large gaps in deep research AIs: across 256 humanities and social-science questions the average pass rate is under 17%, with the leading system (GPT-5.5) clearing 57.6%; building such tests in class helps students learn to judge machine-produced knowledge.
medium descriptive Full text arxiv
Teaching AI Through Benchmark Construction: QuestBench as a Course-Based Practice for Accountable Knowledge Work · Haiyang Shen, Jiuzheng Wang, Taian Guo, Mugeng Liu, Wenchun Jing, Chongyang Pan,... · May 20, 2026
Individually rational delegation to AI can produce a collective 'sociotechnical lock‑in' that erodes shared standards; stronger communication and institutional norms can avert this failure by imposing social commitments.
n/a theoretical Full text arxiv
The Human-AI Delegation Dilemma: Individual Strategies, Collective Equilibria and Sociotechnical Lock-in · Angjelin Hila · May 20, 2026
Three-quarters of AI design mismatches trace back to developers’ focus on speed: 83% of workplace AI incidents arise when systems don’t match worker needs, and developers’ emphasis on efficiency explains 74% of task-level misalignments, especially in people-facing roles.
low correlational Full text arxiv
The Quiet Path from Seemingly Minor Design Errors to Workplace AI Incidents · Julia De Miguel Velázquez, Sanja Šćepanović, Andrés Gvirtz, Daniele Quercia · May 20, 2026
Software and software R&D—linked to AI—explain half of U.S. nonfarm business labor productivity growth between 2017 and 2024, and half of the 1.2 percentage-point acceleration versus the prior five years, implying AI is already materially affecting official productivity measures.
low descriptive Summary semantic_scholar
AI as an Innovation in the Method of Innovation: Implications for Productivity Growth · Filippo Bontadini, Carol Corrado, Jonathan Haskel, C. Jona-Lasinio · Fetched May 20, 2026
Analysis of AI patents (2002–2021) reveals four distinct classes of AI inventions and the topical pathways through which scientific research is incorporated into technology; the mapped science-to-technology links offer targeted insights for corporate R&D strategy and innovation policy.
n/a descriptive Summary semantic_scholar
Knowledge flows from science to AI technology: Identifying core and brokerage technological roles · Seokhui Lee, Jisoo Hur, Junseok Hwang, D. Kogler, Keungoui Kim · Fetched May 20, 2026
AI-driven learning cultures in Pakistani IT firms are linked to stronger innovation performance, but improvements appear to come less from standalone tools than from better coordination of knowledge and firm-level intelligence.
low correlational Summary openalex
Enhancing innovation in Pakistan’s IT sector · · May 20, 2026
Digital talent and industrial digitalization spur regional growth only up to a point: across Chinese provinces their combined impact follows an inverted-U, implying diminishing (and eventually negative) returns to excessive clustering—except in the Yangtze River Delta, where coordinated digitalization and talent agglomeration still drive growth.
low correlational Full text openalex
Emerging Technology-Driven Development: The Interactive Relation Among Digital Talent Agglomeration, Industrial Digitalization, and China's Economic Growth · Xiaoyue Liu, Jing Zuo · May 20, 2026
Organizations should evaluate humans and AI on shared capability 'profiles' rather than disparate metrics; doing so makes task allocation more predictive, explainable and auditable, but requires new measurement standards and infrastructure.
n/a theoretical Full text openalex
Reverse Turing Tests for Human-Machine Task Suitability Assessments Should be Profile-Driven · Jonathan Prunty, Marko Tešić, Ben Slater, Zachary Tidler, Paul Clothier, Luning ... · May 20, 2026
Open-world tests find real-world AI reach: in a CRUX pilot, an AI agent built and published a simple iOS app with only one avoidable human intervention, indicating that messy, long-horizon evaluations can reveal capabilities that standard benchmarks understate.
low descriptive Full text arxiv
Open-World Evaluations for Measuring Frontier AI Capabilities · Sayash Kapoor, Peter Kirgis, Andrew Schwartz, Stephan Rabanser, J. J. Allaire, R... · May 19, 2026
A privacy-first AI coaching system doubled daily check-in adherence and produced larger average weight loss in a real-world deployment compared with static grouping, using four-view data separation and a privacy-constrained contextual bandit to personalize peer support while preventing PII leakage.
medium quasi_experimental Full text arxiv
Privacy-by-Design Adaptive Group Assignment for Digital Lifestyle Coaching at Scale · Nariman Mani, Salma Attaranasl · May 19, 2026
Autonomous coding agents speed some development tasks but do not replace engineering discipline; firms need stricter process controls, verification, and human gates. The paper proposes Agentic Agile-V and a SCOPE-V task loop to convert conversational prompts into verifiable engineering artifacts.
medium review_meta Full text arxiv
Agentic Agile-V: From Vibe Coding to Verified Engineering in Software and Hardware Development · Christopher Koch · May 19, 2026
A short checklist sharpens LLM answers and trims interaction: checklist prompts raised mean rubric scores to 7.50/8 and used fewer tokens than both raw and clarifying-question prompts across ChatGPT, Claude and Grok.
medium quasi_experimental Full text arxiv
Less Back-and-Forth: A Comparative Study of Structured Prompting · Saurav Ghosh, Gabriella Polach, Abdou Sow · May 19, 2026
Cleaner code does not make autonomous coding agents more likely to pass tests, but it makes them notably more efficient: repositories with fewer static-analysis violations reduce an agent's token use by about 7–8% and slash file revisitations by roughly one-third, implying cleaner code lowers compute and navigational overhead even when correctness is unchanged.
medium quasi_experimental Full text arxiv
Does Code Cleanliness Affect Coding Agents? A Controlled Minimal-Pair Study · Priyansh Trivedi, Olivier Schmitt · May 19, 2026
Structured 'Skills' provide little marginal gain for an autonomous cybersecurity agent: across 180 CTF runs the full-Skills condition outperformed no-Skills by only 8.9 percentage points (statistically insignificant), and in environments that return precise, low-latency feedback curated Skills often cease to help or can degrade performance.
medium quasi_experimental Full text arxiv
When Skills Don't Help: A Negative Result on Procedural Knowledge for Tool-Grounded Agents in Offensive Cybersecurity · Samuel Jacob Chacko, James Hugglestone, Chashi Mahiul Islam, Xiuwen Liu · May 19, 2026
Personal 'twin agents' that speak and act for absent workers blur the line between person and machine, creating new attribution and trust problems; existing safeguards against automation overreliance fail because there is no clear boundary for accountability.
n/a commentary Full text arxiv
From Role to Person: Trust Calibration Challenges in Twin Agents · Hugo Andersson, Niklas Elmqvist · May 19, 2026
OpenComputer creates 1,000 verifiable desktop tasks across 33 applications and finds verifier-based scoring matches human judgments better than LLM-as-judge; despite partial successes, even frontier agents struggle to complete tasks end-to-end, revealing a gap in robust desktop automation.
medium descriptive Full text arxiv
OpenComputer: Verifiable Software Worlds for Computer-Use Agents · Jinbiao Wei, Qianran Ma, Yilun Zhao, Xiao Zhou, Kangqi Ni, Guo Gan, Arman Cohan · May 19, 2026
Pretrained tabular models are systematically biased when people game features, but a lightweight inference-time fix (SPN) that simulates strategic responses restores accuracy; experiments on synthetic and real tabular data show consistent robustness gains without retraining.
medium descriptive Full text arxiv
When Tabular Foundation Models Meet Strategic Tabular Data: A Prior Alignment Approach · Xinpeng Lv, Yunxin Mao, Renzhe Xu, Chunyuan Zheng, Yikai Chen, Haoxuan Li, Jinxu... · May 19, 2026
A general-purpose LLM-driven optimizer outperforms specialized tools across six domains—tripling ARC-AGI accuracy (32.5% to 89.5%), cutting cloud scheduling costs by ~40%, and generating CUDA kernels that match or beat PyTorch in 87% of cases; multi-task search and actionable side information materially accelerate and improve outcomes.
medium other Full text arxiv
optimize_anything: A Universal API for Optimizing any Text Parameter · Lakshya A Agrawal, Donghyun Lee, Shangyin Tan, Wenjie Ma, Karim Elmaaroufi, Rohi... · May 19, 2026
A new benchmark, Cutverse, tests autonomous GUI agents on 186 realistic media-editing tasks across seven professional applications and finds just 36% task success, highlighting persistent failures in long-horizon planning and domain-specific workflows.
medium descriptive Full text arxiv
CutVerse: A Compositional GUI Agents Benchmark for Media Post-Production Editing · Haobo Hu, Xiangwu Guo, Zhiheng Chen, Difei Gao, Haotian Liu, Libiao Jin, Qi Mao · May 19, 2026
A new auto-bidding system that pairs directed exploration with a safety fallback lifts advertising performance: live Taobao tests show a 4.10% increase in ad GMV and a 3.52% rise in ad ROI while also raising clicks and spend.
medium other Full text arxiv
Generative Auto-Bidding with Unified Modeling and Exploration · Mingming Zhang, Feiqing Zhuang, Na Li, Shengjie Sun, Xiaowei Chen, Junxiong Zhu,... · May 19, 2026
Companies that orient toward AI report lower carbon intensity, and their peers and suppliers show downstream emissions reductions too; the link appears routed through higher-quality green innovation though causality is not fully established.
medium correlational Full text openalex
Artificial intelligence orientation and decarbonization spillovers: evidence from Chinese listed firms · Lin Li, Meng Li · May 19, 2026
AI is rewriting the fundamentals of corporate strategy: algorithms act as cognitive partners, foundation models reorganize ecosystems, and control over critical resources shifts toward stakeholder networks, forcing firms to rethink boundaries, governance and everyday strategic practice.
n/a theoretical Full text openalex
Infusing Artificial Intelligence into Strategy Theory: Synthesizing Five Classic Debates · Rajiv Kashyap, Jim Samuel, Soohyun Lee, Raza Mir · May 19, 2026
Autonomous AI agents will transform payments but outpace existing rules; agent-led transactions promise efficiency and new services in retail, e-commerce and DeFi while creating cybersecurity, liability and compliance gaps that current regulation cannot easily close, prompting a proposed multi-layered governance approach to restore accountability and resilience.
n/a theoretical Full text openalex
AI Agents in Payments: Applications, Risks and Regulations · David Restrepo Amariles, Damien Charlotin, Liyun He-Guelton · May 19, 2026
AI is reshaping sovereignty: control over data, computation and platforms is displacing territorial authority, creating infrastructural chokepoints that reproduce colonial patterns of extraction. Narrow regulations targeting applications will leave upstream power intact unless alternative, stewardship-based models of techno-sovereignty are supported.
n/a theoretical Full text openalex
Digital colonialism, techno-sovereignty, and infrastructural power · Artur Ishkhanyan · May 19, 2026
Generative AI turns ads from visible slots into hidden interventions embedded in model generation, enabling subtle commercial influence across product mentions, framing and long-term preference shaping; existing systems and governance focus on the most observable forms while leaving deeper, harder-to-detect influence largely unmeasured and unregulated.
n/a theoretical Full text arxiv
Generative AI Advertising as a Problem of Trustworthy Commercial Intervention · Jingyi Qiu, Qiaozhu Mei · May 18, 2026