The Commonplace
Home Dashboard Papers Evidence Syntheses Digests 🎲

Papers (1134)

Filter: Sort:
A large study of 6,000 live coding-agent sessions finds agents either write almost all or none of committed code — 41% 'vibe coding' versus 23% human-only — yet only 44% of agent-produced code survives into commits and agent contributions carry more security flaws, with users pushing back in 44% of interactions.
medium descriptive arxiv
SWE-chat: Coding Agent Interactions From Real Users in the Wild · Joachim Baumann, Vishakh Padmakumar, Xiang Li, John Yang, Diyi Yang, Sanmi Koyej... · Apr 22, 2026
AI-assisted government summaries underrepresent dissent: Canada's 2025–26 AI Strategy consultation summaries excluded 15–17% of participants and performed 8–9 percentage points worse than a random-participant baseline, with critics of AI facing the highest exclusion rates.
medium correlational arxiv
Participatory provenance as representational auditing for AI-mediated public consultation · Sachit Mahajan · Apr 22, 2026
Large language models reliably flag fraudulent investment opportunities and resist motivated persuasion more than lay humans; in controlled tests humans endorsed fraud ~13–14% of the time while tested LLMs never did, and motivated framing did not suppress — and slightly increased — AI warnings.
high rct arxiv
Large Language Models Outperform Humans in Fraud Detection and Resistance to Motivated Investor Pressure · Nattavudh Powdthavee · Apr 22, 2026
Structuring GenAI-driven development with executable requirements and architectural artifacts curbs agent drift and moves engineers from line-by-line coding to higher-level design work; early lab tests on a web app show improved traceability and maintainability compared with unstructured vibe coding.
low descriptive arxiv
Shift-Up: A Framework for Software Engineering Guardrails in AI-native Software Development -- Initial Findings · Petrus Lipsanen, Liisa Rannikko, François Christophe, Konsta Kalliokoski, Vlad S... · Apr 22, 2026
Firms that disclose using AI in financial reporting show better audit outcomes in Nigeria — fewer restatements and audit-fee patterns consistent with higher audit quality — partly because AI disclosure correlates with greater transparency and stronger internal controls.
medium correlational semantic_scholar
Artificial Intelligence Adoption in Financial Reporting and Audit Quality: Evidence from Nigerian Listed Firms · Akomolehin Fo, Oluwaremi Jb, Aluko Or, Famoroti Jo · Fetched Apr 22, 2026
AI deployments depress operating margins initially for KOSDAQ firms — consistent with short‑term transition costs — while investors value AI as a growth option in ICT firms, which show significant increases in market valuation.
medium quasi_experimental semantic_scholar
The Dynamic Causal Effects of Corporate AI Adoption on Profitability and Market Value : An Empirical Analysis of KOSDAQ Panel Data · Jungsoo Kim, Bong-Gi Baek · Fetched Apr 22, 2026
Industry-level national-account analysis suggests AI adoption raises measured productivity and reduces input and labor intensity while tilting employment toward younger, less-educated workers; however, estimates depend importantly on how and when AI pervasiveness is dated, highlighting substantial measurement uncertainty.
medium quasi_experimental semantic_scholar
Early Estimates of the Impact of AI Within BEA’s Industry Economic Accounts · Tina Highfill, Jon D. Samuels · Fetched Apr 22, 2026
Selective forgetting makes LLM agents leaner, cleaner and safer: in controlled tests, memory-pruning mechanisms raised access efficiency by ~8.5% and improved content signal-to-noise by ~29%, while eliminating tested security risks; the paper argues forgetting is as important as remembering for deployed agents.
medium descriptive arxiv
FSFM: A Biologically-Inspired Framework for Selective Forgetting of Agent Memory · Yingjie Gu, Wenjian Xiong, Liqiang Wang, Pengcheng Ren, Chao Li, Xiaojing Zhang,... · Apr 22, 2026
A stateless memory design gives enterprises the auditability and determinism they need without sacrificing accuracy: Deterministic Projection Memory cuts decision-time LLM calls to one, runs 7–15× faster under tight budgets and improves factual precision and reasoning coherence markedly versus summarization pipelines at 20× compression.
medium other arxiv
Stateless Decision Memory for Enterprise AI Agents · Vasundra Srinivasan · Apr 22, 2026
A spreadsheet assistant that reveals its stepwise decisions lets users catch mistakes and feel ownership: exposing and controlling each action led participants to detect errors, understand the task better, and view themselves as collaborators rather than passive reviewers.
medium quasi_experimental arxiv
Auditing and Controlling AI Agent Actions in Spreadsheets · Sadra Sabouri, Zeinabsadat Saghi, Run Huang, Sujay Maladi, Esmeralda Eufracio, S... · Apr 22, 2026
A rights‑based 'white‑box' protocol would require Mexican fintechs to make credit and financial algorithms transparent, auditable and non‑discriminatory to protect citizens' digital sovereignty and financial inclusion. While these rules aim to curb bias and strengthen oversight, they risk raising compliance costs and favoring incumbents unless paired with sandboxes, shared datasets and capacity support.
low review_meta openalex
Diego Saucedo Portillo Sauceport Research · Diego Saucedo Portillo · Apr 22, 2026
A Mexican legal blueprint would make 'Digital Sovereignty' a constitutional right and require 'white‑box' audits for credit algorithms, enforcing transparency and anti‑discrimination in fintech. The measures are likely to favour well‑capitalised incumbents and spawn a certification/audit industry, while risking higher short‑run lending costs and barriers for small providers unless offset by proportional rules or sandboxes.
low theoretical openalex
Diego Saucedo Portillo Sauceport Research · Diego Saucedo Portillo · Apr 22, 2026
Generative-AI algorithms are quietly binding gig couriers to platforms, eroding the employee/contractor divide; courts should test for substantive algorithmic control, legislatures should create a 'quasi-employee' status with tiered protections, and platforms must bear greater algorithmic accountability.
n/a commentary openalex
AIGC+ Determination of Labor Relations in the Context of the “Gig Economy” · Manyu Wei · Apr 22, 2026
U.S. chip export controls rewire, not sever, Sino‑technology links: Beijing’s defensive substitution and regulatory countermeasures are reshaping supply chains into a durable, politically segmented decoupling rather than dissolving interdependence.
low theoretical openalex
Weaponized Interdependence and Dynamics of Partial Decoupling: Chokepoints, Adaptation, and the Security Dilemma in Advanced Technologies · Mofei Shen · Apr 22, 2026
Making AI reviewable and usable in government can backfire: codified procedures improve oversight on paper but create a playbook for successors to legally entrench politicized decisions, increasing strategic exploitation and locking in automation expansions.
n/a theoretical openalex
AI Governance under Political Turnover: The Alignment Surface of Compliance Design · Andrew J. Peterson · Apr 22, 2026
AI traders reliably pool private signals in simple prediction markets but falter when information structures grow complex; smarter models do better, yet feedback on past performance unexpectedly undermines market accuracy and profits.
medium other arxiv
Information Aggregation with AI Agents · Spyros Galanis · Apr 21, 2026
AI agents often echo their human owners: autonomous Moltbook agents reproduce owners' linguistic style, topics and affect, and agents that more closely mirror owners are more likely to divulge owner-related personal information, raising privacy and governance concerns.
medium correlational arxiv
Behavioral Transfer in AI Agents: Evidence and Privacy Implications · Shilei Luo, Zhiqi Zhang, Hengchen Dai, Dennis Zhang · Apr 21, 2026
Generic AI assistants confidently decide when evidence is missing and fail most inconclusive unemployment-adjudication cases, scoring about 15% on insufficient-information instances; a structured prompting checklist (SPEC) boosts accuracy to 89% while correctly deferring decisions when facts are inadequate.
medium descriptive arxiv
Learning When Not to Decide: A Framework for Overcoming Factual Presumptuousness in AI Adjudication · Mohamed Afane, Emily Robitschek, Derek Ouyang, Daniel E. Ho · Apr 21, 2026
A system that jointly adapts task scheduling and robot motion to individual human behavior (RAPIDDS) makes human-robot teams more efficient and safer, outperforming non-adaptive planners in simulation and a 32-participant lab study; improvements hold across objective metrics (speed, proximity) and subjective measures (fluency, preference).
medium quasi_experimental arxiv
Multi-Cycle Spatio-Temporal Adaptation in Human-Robot Teaming · Alex Cuellar, Michael Hagenow, Julie Shah · Apr 21, 2026
Tool-augmented LLMs can reliably handle complex quantitative finance queries: in a 100-question benchmark, top agents achieved near-perfect tool selection and minimal hallucination, underscoring external tools as the path to dependable numerical reasoning.
medium descriptive arxiv
Time Series Augmented Generation for Financial Applications · Anton Kolonin, Alexey Glushchenko, Evgeny Bochkov, Abhishek Saxena · Apr 21, 2026
Scholarly metadata will not be solved by a binary choice between open access and commercial enclosure; a lasting 'innovation annulus' between free structured data and premium refinement is functional and must be calibrated, not eliminated, as AI lowers basic structuring costs but raises provenance and refinement thresholds.
n/a theoretical arxiv
Market Dynamics, Governance and Open Research Metadata in the AI Era · Daniel W. Hook · Apr 21, 2026
An audit of a college Early Warning System shows it misallocates support: younger, male and international students are over-flagged while older and female students are under-identified. Compressing continuous risk scores into percentile tiers further magnifies these inequities across the intervention pipeline.
medium descriptive arxiv
Fairness Audits of Institutional Risk Models in Deployed ML Pipelines · Kelly McConvey, Dipto Das, Maya Ghai, Angelina Zhai, Rosa Lee, Shion Guha · Apr 21, 2026
A simple evaluation-driven scaling strategy (SimpleTES) lets relatively modest LLMs outperform larger baselines across 21 scientific tasks — doubling LASSO speed, cutting quantum gate overhead by 24.5%, and finding new combinatorial constructions — and trajectory-based retraining improves generalization to new problems.
medium descriptive arxiv
Evaluation-driven Scaling for Scientific Discovery · Haotian Ye, Haowei Lin, Jingyi Tang, Yizhen Luo, Caiyin Yang, Chang Su, Rahul Th... · Apr 21, 2026
Researchers propose shifting AI agents from single-user task automation to permanent, identity-bound agents that represent people in multi-user collaboration, backed by layered identity, per-identity authorization, and action-level audit logs; they demonstrate the idea with a prototype framework, ClawNet, but provide no field evidence of real-world impact.
n/a theoretical arxiv
ClawNet: Human-Symbiotic Agent Network for Cross-User Autonomous Cooperation · Zhiqin Yang, Zhenyuan Zhang, Xianzhang Jia, Jun Song, Wei Xue, Yonggang Zhang, Y... · Apr 21, 2026
Augmenting tabular housing-placement models with fine-tuned LLM casenote summaries raises accuracy and narrows error disparities in a nonprofit dataset; zero-shot LLM use produces mixed fairness effects and findings are limited by short, redacted notes and single-site data.
low correlational arxiv
Auditing LLMs for Algorithmic Fairness in Casenote-Augmented Tabular Prediction · Xiao Qi Lee, Ezinne Nwankwo, Angela Zhou · Apr 21, 2026