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

Papers (1747)

Filter: Sort:
Large language models' evaluations are pulled by the sentiment of preceding messages: histories laden with negative reviews make models systematically harsher, with the effect strongest for ambiguous items and negative context 1.62× more influential than positive. The bias appears across multiple providers and model sizes, and is largely mitigated by using a fresh context per item or balancing batched histories.
high quasi_experimental Full text arxiv
AMEL: Accumulated Message Effects on LLM Judgments · Sid-ali Temkit · May 21, 2026
AI tools at a large tech firm are blurring job boundaries and smoothing peer collaboration, but they are also fraying informal mentorship and feedback systems that sustain career growth.
low descriptive Full text arxiv
Beyond the Org Chart: AI and the Transformation of Invisible Work · Stephanie Rosenthal, Shamsi Iqbal · May 21, 2026
Users routinely lean on AI for trivial tasks that it does not meaningfully speed up, while underreporting how often they use it and overestimating time savings; prior exposure further entrenches reliance, risking an inefficient overreliance feedback loop.
medium quasi_experimental Full text arxiv
The efficiency-gain illusion: People underestimate the rate of AI use and overestimate its benefits on simple tasks · Sunny Yu, Myra Cheng, Ahmad Jabbar, Ilia Sucholutsky, Katherine M. Collins, Dan ... · May 21, 2026
Bigger LLMs can make riskier forecasts: larger and post-trained models inflate the upper tail on outbreaks, hyperinflation and housing booms, appearing better on common threshold tests but worse on tail-inclusive scoring, risking overly aggressive extrapolations in high-stakes forecasting.
medium descriptive Full text arxiv
Is Capability a Liability? More Capable Language Models Make Worse Forecasts When It Matters Most · Nick Merrill, Jaeho Lee, Ezra Karger · May 21, 2026
AI agents can assemble basic financial spreadsheets competently, but struggle to deliver professional-quality models as workflows grow complex; even top systems often err on multi-step calculations and fall short of enterprise standards.
medium descriptive Full text arxiv
WorkstreamBench: Evaluating LLM Agents on End-to-End Spreadsheet Tasks in Finance · Thomson Yen, Julian Poeltl, Harshith Srinivas Gear, Yilin Meng, Joshua Fan, Adam... · May 21, 2026
A clear, technical boundary—open pre-training artifacts versus proprietary post-training weights—could break the logjam in industry–academia ML deals; adopting the PBOS template would keep scientists at the negotiating table and streamline collaborations.
n/a commentary Full text arxiv
Position: The Pre/Post-Training Boundary Should Govern IP in Industry-Academia ML Collaborations · Dirk Bergemann, Soheil Ghili, Nitzan Mekel-Bobrov · May 21, 2026
Encoding procedures into small fine‑tuned agents matches orchestration for procedural tasks while cutting context costs and protecting proprietary logic; this approach resolves several practical barriers that have nevertheless kept developers favoring external orchestrators.
medium quasi_experimental Full text arxiv
Compiling Agentic Workflows into LLM Weights: Near-Frontier Quality at Two Orders of Magnitude Less Cost · Simon Dennis, Rivaan Patil, Kevin Shabahang, Hao Guo · May 21, 2026
A Nash-bargaining guided multi-agent RL system raises simulated peer-to-peer EV energy-trading welfare by roughly 62% and trading volume by 63% versus a double-auction baseline, while producing markedly fairer allocations; results are promising but confined to synthetic simulations without real-world validation.
medium descriptive Full text arxiv
Incentive-Aligned Vehicle-to-Vehicle Energy Trading via Nash-Integrated Multi-Agent Reinforcement Learning · Yujin Lin, Yue Yang, Hao Wang · May 21, 2026
Coordinated AI agents add value in multi-signal scientific tasks but not uniformly: cross-channel composites boost disease-emergence and exoplanet vetting performance, yet when one strong signal dominates gains fall to interpretation and provenance rather than top-line accuracy.
medium descriptive Full text arxiv
Cross-domain benchmarks reveal when coordinated AI agents improve scientific inference from partial evidence · Fiona Y. Wong, Markus J. Buehler · May 21, 2026
In a timed multi-stage Risk tournament Gemini-3.1-pro-preview won 20 of 32 matches against three rivals, but when all planners' outputs were executed on a common scaffold planner performance was statistically indistinguishable. The result implies that system-level execution, objective-tracking, and runtime reliability — not planning ability alone — determine live-agent effectiveness.
medium quasi_experimental Full text arxiv
Evaluating Large Language Models as Live Strategic Agents: Provider Performance, Hybrid Decomposition, and Operational Gaps in Timed Risk Play · H. C. Ekne · May 21, 2026
A new 10,071‑pair agricultural visual‑grounding benchmark exposes large capability gaps in current models: the best zero‑shot multi‑target F1 is only 0.35 and high‑IoU mask success rates remain under 17%, underscoring limits to reliable field automation.
n/a descriptive Full text arxiv
AgroVG: A Large-Scale Multi-Source Benchmark for Agricultural Visual Grounding · Haocheng Li, Juepeng Zheng, Zenghao Yang, Kaiqi Du, Guilong Xiao, Gengmeng Pu, H... · May 21, 2026
A system called Echo recycles users' corrections into training data and lifts code-completion acceptance rates from 25.7% to 35.7% in production, showing that continuous learning from real interaction logs can materially improve AI performance. The finding suggests a scalable path to break static performance ceilings where users routinely refine agent outputs.
medium quasi_experimental Full text arxiv
Echo: Learning from Experience Data via User-Driven Refinement · Hande Dong, Xiaoyun Liang, Jiarui Yu, Jiayi Lin, Changqing Ai, Feng Liu, Wenjun ... · May 21, 2026
Fine-tuned LLMs make ad delivery more stable and predictive: a semantic, graph-augmented retrieval layer reduced sensitivity to small creative changes and improved engagement in production A/B tests, demonstrating practical gains from LLMs in large-scale ad recommendation.
medium rct Full text arxiv
LLM Retrieval for Stable and Predictable Ad Recommendations · Vinodh Kumar Sunkara, Satheeshkumar Karuppusamy, Hangjun Xu, Sai Deepika Regani,... · May 21, 2026
Making an AI transparent doesn't automatically win more data: transparency raises stated willingness to share only for users who already trust AI, while immediate sharing decisions remain high and unchanged across human, white-box, and black-box scenarios.
medium rct Summary openalex
Understanding Data-Sharing with AI Systems: The Roles of Transparency, Trust, and the Processing Entity · Julian Rosenberger, Sophie Kuhlemann, Verena Tiefenbeck, Mathias Kraus, Patrick ... · May 21, 2026
Intelligent technologies are reshaping international marketing: reported ROI gains of about 12–25% accompany a shift from linear to nonlinear consumer journeys, prompting new data-driven frameworks while cross-border data rules, algorithmic opacity and cultural frictions complicate implementation.
n/a review_meta Full text openalex
Research on International Marketing in the Context of Intelligence · Xiaoqi Shao, Juntao Fang · May 21, 2026
AI-driven performance management is making healthcare workplaces more responsive and efficient—speeding feedback, surfacing leaders and saving managerial time—while predictive analytics help manage staffing and burnout; however, benefits are fragile without strong safeguards against bias, privacy breaches and governance failures.
medium review_meta Full text openalex
The influence of AI-Driven Employee Performance Management (EPM) Systems in the sustained enhancement of Employee Mobilization, Organizational Performance, and the Change Culture in the healthcare sector: A scoping review analysis · Dimitris Karaferis, Dimitra Balaska, Demetra Konstantakopoulou, Yannis Pollalis · May 21, 2026
China’s digital economy has shifted jobs toward services and away from industry across provinces between 2013 and 2024, led more by firms digitally integrating production than by pure digital industries; the 2017 national digital strategy marked a turning point after which digital industrialization began to support manufacturing while industrial digitalization continued to propel servicization.
medium quasi_experimental Full text openalex
The impact of China's digital economy development on changes in the labor structure · Xinyan Luo · May 21, 2026
Georgia risks missing AI-fuelled economic gains because it has no national strategy or legal framework; adopting an integrated, adaptive policy and infrastructure program modeled on successful small-country approaches would better position it as a regional technology leader.
low descriptive Full text openalex
Economic Impact of Artificial Intelligence and Policy Framework: Analysis of International Practice and Strategic Recommendations for Georgia · Nino Bitsadze, Nino Taganashvili · May 21, 2026
Algorithmic authority alone does not secure legitimacy: AI can improve efficiency but decision legitimacy in organizations hinges on transparency, explainability and maintained human judgment; without these, algorithms may both bolster and erode stakeholders' trust.
n/a theoretical Full text openalex
Decision Legitimacy in AI-Enabled Organizations: A Multilevel Framework of Algorithmic Authority and Human Judgment · Aslı Tenderis · May 21, 2026
Vietnam should adopt a six-pronged policy framework to harness the productivity and innovation potential of autonomous AI agents while containing new privacy, security and displacement risks; the paper distils international practices into actionable regulatory and institutional steps for the Vietnamese context.
low commentary Full text openalex
Regulatory Policy for the Agent Economy in the Digital Age: Lessons for Vietnam · Le Hai Ha · May 21, 2026
Replacing fragile item IDs with multimodal, hierarchical codes improved livestream recommendation: a production A/B test showed +0.55% quality watch duration, +2.05% cold-start room views and +0.05% active hours across a billion-user platform.
high rct Full text arxiv
FLUID: From Ephemeral IDs to Multimodal Semantic Codes for Industrial-Scale Livestreaming Recommendation · Xinhang Yuan, Zexi Huang, Anjia Cao, Xudong Lu, Zikai Wang, Penghao Zhou, Chang ... · May 20, 2026
Occupation-exposure scores built from AI chat logs reflect who uses the platforms, not the workforce, and change estimated employment effects nearly twofold across vendors and channels; correcting for workforce composition cuts estimated impacts by up to 93% and theoretical bounds show the bias typically masks substitution rather than exaggerating it.
medium quasi_experimental Full text arxiv
Who Uses AI? Platforms, Workforce, and AI Exposure · Michelle Yin, Burhan Ogut · May 20, 2026
Heavy use of on-demand AI can undermine learning in a controlled reasoning task, while informative AI boosts short-term performance and—on average—does not reduce later independent performance; outcomes, however, vary substantially across users.
medium quasi_experimental Full text arxiv
The Impact of AI Usage and Informativeness on Skill Development in Logical Reasoning · Shang Wu, Hongyu Yao, Catarina Belem, Shuyuan Fu, Mark Steyvers, Padhraic Smyth · May 20, 2026
Human-AI teams rarely achieve true synergy; closing the 'synergy gap' requires redesigning more than interfaces. A six-part sociotechnical framework shows why context, decision frameworks, participant roles, AI capabilities, interaction design and holistic evaluation must all be aligned to unlock combined performance gains.
n/a theoretical Full text arxiv
Addressing the Synergy Gap: The Six Elements of the Design Space · Tommaso Turchi, Ben Wilson, Matt Roach, Alan Dix, Alessio Malizia · May 20, 2026
AI agents produce modest refactoring gains but also introduce lint and security issues: 22.5% of agent commits raise measured code quality (usability up most), yet 24% of files add Pylint issues and 4.7% add Bandit findings — nonetheless, 73.5% of such PRs are merged.
medium descriptive Full text arxiv
Quality and Security Signals in AI-Generated Python Refactoring Pull Requests · Mohamed Almukhtar, Anwar Ghammam, Hua Ming · May 20, 2026