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

Papers (1747)

Filter: Sort:
A Shannon-based scaling law shows why bigger or longer-trained LLMs sometimes get worse: if model size or data growth doesn't preserve signal-to-noise ratio, noise can dominate and induce U-shaped performance loss. The proposed law fits Pythia and OLMo2 loss basins better than classical power laws and correctly extrapolates to an unseen 12B model under multiple perturbations.
medium theoretical Full text arxiv
LLMs as Noisy Channels: A Shannon Perspective on Model Capacity and Scaling Laws · Xu Ouyang, Deyi Liu, Yuhang Cai, Jing Liu, Yuan Yang, Chen Zheng, Thomas Hartvig... · May 22, 2026
CHRONOS marries neural-ODE decay, changepoint-aware Shapley pricing and DP-bandit scheduling to boost retrieval recall to 0.937@10 while enforcing epsilon=4.25 (delta=1e-6), but at this privacy level released valuations are noise-dominated so most utility comes from public index routing and low-sensitivity scheduling.
low other Full text arxiv
CHRONOS: Temporally-Aware Multi-Agent Coordination for Evolving Data Marketplaces · Joydeep Chandra · May 22, 2026
AI-written explanations don’t make people more accurate: narrative explanations from large language models fail to raise classification accuracy over raw AI predictions, while making people more likely to follow the AI (even when it’s wrong) and slowing decisions.
medium rct Full text arxiv
Human Decision-Making with Persuasive and Narrative LLM Explanations · Laura R. Marusich, Mary Grace Kozuch Dhooghe, Jonathan Z. Bakdash, Murat Kantarc... · May 22, 2026
Detailed life-history prompts make LLM-based resident simulations noticeably more faithful, and curriculum-LoRA cuts personalization cost by about 10x while matching top-tier fidelity, bringing practical in-silico policy testing within reach for cash-strapped local governments.
medium descriptive Full text arxiv
Benchmarking LLMs for Community Governance Simulation with Life-history Narratives · Xu Chen, Yuanzi Li, Lei Wang, Nan Lu, Yang Wang, Anding Wang, Lei Shi, Xiaoxing ... · May 22, 2026
A three-arm ad experiment shows the ad-delivery algorithm—not the creative—drives most audience reallocation: in a Meta campaign the algorithm increased female impression share by 2.07 percentage points while the creative reduced it by 0.68 points, meaning roughly three-quarters of the shift was algorithmic and a standard two-arm A/B test underestimates that effect by about half.
high rct Full text arxiv
Algorithm or Creative? A Three-Arm Experimental Design for Decomposing Algorithmic Bias in Platform A/B Tests · Pallavi Pal, Anjana Susarla · May 22, 2026
A three‑phase distillation recipe compresses large retrieval models into compact encoders that nearly match teacher accuracy while slashing latency and raising ad revenue; deploying the student retriever in Bing Ads produced ~98% precision recovery, up to 27× lower latency, and a 1% revenue lift in production A/B tests.
high rct Full text arxiv
HARNESS-LM: A Three-Phase Training Recipe for Harnessing SLMs in Sponsored Search Retrieval · Vipul Gupta, Shikhar Mohan, Lakshya Kumar, Pranjal Chitale, Nikit Begwani, Amit ... · May 22, 2026
U.S. export controls on advanced chips and AI tools may be self-defeating: they can spur China’s push for technological self-reliance while raising legitimate WTO legal challenges, so policymakers should avoid overbroad security claims under GATT Article XXI(b).
low commentary Full text arxiv
Strategic Stalemates: The Paradox of Export Controls in the U.S.-China AI Race · Jingwen Liu, Jyh-An Lee · May 22, 2026
Regulation will determine whether AI spreads prosperity or concentrates power: the EU’s comprehensive digital and AI rules give it an institutional edge, while Ukraine’s reliance on older legal structures risks higher market concentration, inequality and slower innovation; policymakers should craft unified, nationally adapted AI regulations drawing on European models.
n/a theoretical Full text semantic_scholar
ECONOMIC SYSTEMS IN THE CONTEXT OF DIGITALISATION AND AI: THEORETICAL AND LEGAL MODELLING · Yana Tytska, V. Anisimov, O. Matvieiev · Fetched May 22, 2026
Siting modular AI compute at wind farms could unlock vast, underutilized renewable capacity—890+ GW lie within 50 ms of Azure data centers—and a new router, XWind, cuts P99 inference latency by up to 52% in a 64‑GPU emulation, making behind‑the‑meter AI deployments materially more performant. But the results are from a controlled testbed and feasibility mapping, leaving open questions about costs, grid impacts, and real‑world scalability.
medium descriptive Full text arxiv
XWind: A Cross-site Router for Large Language Model Inference Serving at Renewable Energy Farms · Tella Rajashekhar Reddy, Atharva Deshmukh, Liangcheng Yu, Chaojie Zhang, Mike Sh... · May 22, 2026
Benchmarks that mimic NLP tasks overstate LLMs' readiness for real-world knowledge work; a three-step framework — name the work activity, specify the tested setting, and score the work product — plus an O*NET-derived activity inventory clarifies the strongest work claims a benchmark score can support and exposes gaps in popular datasets.
n/a theoretical Full text arxiv
Design and Report Benchmarks for Knowledge Work · Yining Hua, Hongbin Na, Cyrus Ayubcha, Levi Lian · May 22, 2026
A procedurally generated strategic benchmark reveals that leading LLMs, while similar in overall strength, differ sharply in the kinds of strategic competence they possess — and some top models are markedly more locally volatile. GENSTRAT's capability profiles and a new jaggedness measure expose deployment risks that simple leaderboard scores miss.
medium descriptive Full text arxiv
GENSTRAT: Toward a Science of Strategic Reasoning in Large Language Models · Vartan Shadarevian, Kia Ghods, Alex Kenich, Anany Kotawala · May 22, 2026
The Foundation Protocol: a proposed graph-based coordination layer aims to make autonomous agents interoperable, economically coherent and auditable by treating coordination as shared infrastructure; it is a detailed conceptual blueprint but untested in real-world deployments.
n/a theoretical Full text arxiv
Foundation Protocol: A Coordination Layer for Agentic Society · Bang Liu, Yongfeng Gu, Jiayi Zhang, Zhaoyang Yu, Sirui Hong, Maojia Song, Xiaoqi... · May 22, 2026
A production-aligned benchmark finds leading LLM kernel agents fail to improve real-world inference speed: the best agent is 6% slower than hardened systems while others perform markedly worse. FastKernels—covering 46 representative architectures and matching production interfaces—highlights benchmark-production misalignment as a key bottleneck to deploying agent-generated kernels.
medium descriptive Full text arxiv
FastKernels: Benchmarking GPU Kernel Generation in Production · Gabriele Oliaro, Yichao Fu, May Jiang, Owen Lu, Junli Wang, Zhihao Jia, Hao Zhan... · May 22, 2026
AI is shifting from isolated assistance to coordinating entire research workflows, promising faster and more hybridized science; but autonomy currently only looks credible in structured, rapidly verifiable domains, with reproducibility, provenance, validation, and accountability still anchoring humans to the loop.
n/a review_meta Full text arxiv
AutoResearch AI: Towards AI-Powered Research Automation for Scientific Discovery · Guiyao Tie, Jiawen Shi, Dingjie Song, Yixiao Huang, Ziji Sheng, Xueyang Zhou, Da... · May 22, 2026
Orchestrating AI makes it cheaper to modularize systems, but tasks requiring verifiable, auditable, or assignable outputs resist unbundling; firms will either keep capabilities integrated, run parallel tracks, or accept mounting governance 'rule debt'.
n/a theoretical Full text arxiv
Redrawing the AI Map: A Theory of Accountability Boundaries in Agentic Ecosystems · Muhammad Zia Hydari, Farooq Muzaffar · May 22, 2026
Large experiment finds a 'speedup illusion': users expect LLMs to be faster but actual completion times on simple tasks are unchanged, even as subjective effort falls; the bias is specific to AI and not seen when imagining human help.
medium rct Full text arxiv
Cognitive offloading and the speedup illusion in human-AI interaction · Sunny Yu, Myra Cheng, Ahmad Jabbar, Ilia Sucholutsky, Katherine M. Collins, Dan ... · May 22, 2026
Firms reshape hiring and job tasks as generative AI diffuses: hiring reallocation explains just over half of the decline in AI exposure while within-job redesign accounts for nearly 40%, and senior roles adjust earlier mainly through reallocation whereas junior roles show a broader mix of redesign and reallocation.
medium descriptive Full text arxiv
Generative AI and the Reorganization of Labor Demand · Fangyan Wang, Zaiyan Wei, Yang Wang · May 22, 2026
An LLM-driven synthesis system shrinks months of expert formal-verification work to hours: IDS fully implements and verifies seven distributed key-value-store specs in ~6.8 hours at $106 each, outperforming prior coding agents and producing implementations up to three times faster than published verified systems.
medium other Full text arxiv
Inductive Deductive Synthesis: Enabling AI to Generate Formally Verified Systems · Shubham Agarwal, Alexander Krentsel, Shu Liu, Mert Cemri, Audrey Cheng, Rui Meng... · May 22, 2026
Import-driven adoption of automation boosts wages at adopters—about a 4% average gain five years after an automation spike—largely concentrated in small firms and among older, managerial and white-collar staff, while small firms see a notable rise in within-firm wage inequality (≈7.5%).
medium quasi_experimental Full text openalex
Firm size and the automation wage premium · Laura Bisio, Angelo Cuzzola, Marco Grazzi, Daniele Moschella · May 22, 2026
Digital transformation in China's listed manufacturers expands, not contracts, labor demand—raising demand especially for highly educated, high-skilled workers. Productivity gains and improved worker digital literacy help explain the rise, though effects differ by ownership, factor intensity and firm size.
medium correlational Summary openalex
How Does Digital Transformation Reshape Manufacturing Firms' Labor Demand? · Yongming Wang, Xin Liu, Yujie Zhu, Huifen Cai, Xuefeng Shao · May 22, 2026
Control over data and platforms, not mere connectivity, now drives inequality in industrial economies; workers, small firms and poorer countries participate in digital systems but capture little value. Stronger governance and redistribution of digital value are needed to make industrial digitalisation compatible with inclusive growth and SDG goals.
n/a review_meta Full text openalex
Beyond Access: Rethinking Digital Power in Data-Driven Industrial Economies for Sustainable Development · David Mhlanga · May 22, 2026
Uneven AI adoption risks widening capability inequality: low-income groups and segmented labour markets—notably in India’s IT services—face accelerated skill-change and transition risks, calling for coordinated policies on learning equity, transition protections and accountable algorithmic management.
medium review_meta Full text openalex
ARTIFICIAL INTELLIGENCE, INEQUALITIES OF KNOWLEDGE AND RESOURCES, AND SOCIO-ECONOMIC SECURITY: GLOBAL EVIDENCE WITH IMPLICATIONS FOR INDIA'S LABOR AND IT WORKFORCE · Shrikant Dattatraya Wagh, Dr. Shyam Jivan Salunkhe · May 22, 2026
AI's labour disruption will be shaped as much by corporate governance as by automation: the authors propose a Workforce Resilience Governance Framework and two indices to measure organizational readiness and employee trust, arguing firms that manage transitions responsibly can reduce displacement and foster augmentation.
low theoretical Full text openalex
From Automation Panic to Workforce Resilience: A Governance Framework for Enterprise AI Transformation · PannaLal Jaiswal, Rajeew Vishvakarma, Sooraj Jacob · May 22, 2026
Internal HR quality—not external pressures—drives perceived HR effectiveness in Kazakhstan's civil service, and managers are about 61% more likely to be active AI users while each higher tenure category reduces AI use (odds ratio 0.846); access to modern digital tools further boosts uptake.
medium correlational Full text openalex
Determinants of Artificial Intelligence Adoption in Public Sector Human Resource Management: Empirical Evidence from Kazakhstan · Aliya Daueshova, Azamat Zhanseitov, Aigerim Amirova, Saule ISKENDIROVA, Zhansaya... · May 22, 2026
Large language models can map who-knows-what from workplace Slack logs, but accuracy is limited: Gemini 2.5 Flash estimated individual expertise with a mean absolute error of 21.1 percentage points while GPT models performed considerably worse, and more messages did not substantially improve inference—highlighting feasibility, limits, and privacy concerns for automated expertise mapping.
medium correlational Full text arxiv
Can AI Guess What You Know? Performance Comparison of Large Language Models for Human Domain Knowledge Estimation From Communication Logs · Ko Watanabe, Shoya Ishimaru · May 21, 2026