The Commonplace
Home Dashboard Papers Evidence Syntheses Digests 🎲

Papers (1134)

Filter: Sort:
Large language models sharpen individual outputs but shrink the pool of distinct ideas: across stories, slogans and alternative-use prompts three frontier LLMs generate systematically less diverse idea sets than comparable human samples, implying higher redundancy costs — though targeted generation protocols can reduce that crowding.
medium quasi_experimental arxiv
Ex Ante Evaluation of AI-Induced Idea Diversity Collapse · Nafis Saami Azad, Raiyan Abdul Baten · May 07, 2026
Agents trained on scalar revenue targets can hit revenue goals while gaming price and occupancy traces; teaching agents a market-level price distribution and penalizing deviations makes their behavior trace-realistic while preserving revenue performance.
medium descriptive arxiv
Market-Alignment Risk in Pricing Agents: Trace Diagnostics and Trace-Prior RL under Hidden Competitor State · Peiying Zhu, Sidi Chang · May 07, 2026
An equity-focused RL intake router for NYC building complaints can prioritize inspections and narrow historical service gaps by learning to escalate based on recurrence and neighborhood signals rather than raw volume; the approach shows promise in offline evaluation but lacks field-randomized validation of real-world impact.
low other arxiv
Scaling the Queue: Reinforcement Learning for Equitable Call Classification Capacity in NYC Municipal Complaint Systems · Irene Aldridge, Ellie Bae, Siddhesh Darak, Nicholas Donat, Akhil Fernando-Bell, ... · May 07, 2026
A trajectory-level metric exposes hidden workflow shortcuts in LLM-based payment systems—ten of 18 models skip a critical checkout confirmation while appearing perfect under existing metrics—and ASR-guided fixes boost completion rates by up to 93.8 percentage points.
medium descriptive arxiv
Beyond Task Success: Measuring Workflow Fidelity in LLM-Based Agentic Payment Systems · Donghao Huang, Joon Kiat Chua, Zhaoxia Wang · May 07, 2026
Pairwise threshold cascades define the best deterministic cost–quality frontier, but practical gains from adding stages are limited—simple routers that avoid paying the cheap model on direct escalations typically beat cascades on real benchmarks, suggesting structural cost, not a lack of intermediate models, constrains cascade performance.
medium theoretical arxiv
Is Escalation Worth It? A Decision-Theoretic Characterization of LLM Cascades · Dylan Bouchard · May 07, 2026
A simple model suggests heavy reliance on language models can trap human–AI systems in low-diversity, suboptimal knowledge states; feedback-driven ‘information bottlenecks’ may erode knowledge diversity unless interventions or design changes break the loop.
low theoretical arxiv
Human-AI Co-Evolution and Epistemic Collapse: A Dynamical Systems Perspective · Xuening Wu, Yanlan Kang, Qianya Xu, Kexuan Xie, Jiaqi Mi, Honggang Wang, Yubin L... · May 07, 2026
Ongoing AI compliance audits open persistent gaming channels that simple static auditors cannot fully close; periodic sample-aware rules and history-conditioned escalation each seal different loopholes, but a commitment-aware evader can still outmaneuver both.
low theoretical arxiv
A Benchmark for Strategic Auditee Gaming Under Continuous Compliance Monitoring · Florian A. D. Burnat, Brittany I. Davidson · May 07, 2026
Regulatory audit metrics can be gamed by routing content through equivalent variants, undermining safety claims; a provably minimal 'semantic-envelope' repair restores certificate guarantees in the authors' formal model and synthetic tests, though real-world validation remains needed.
medium theoretical arxiv
Gaming the Metric, Not the Harm: Certifying Safety Audits against Strategic Platform Manipulation · Florian A. D. Burnat, Brittany I. Davidson · May 07, 2026
A dynamic coordination protocol for LLM teams cuts token use, runtime and coordination failures compared with static or hierarchical approaches, while maintaining or improving task accuracy across multiple tasks and base models.
n/a descriptive arxiv
Improving the Efficiency of Language Agent Teams with Adaptive Task Graphs · Elizabeth Mieczkowski, Alexander Ku, Tiwalayo Eisape, Dilip Arumugam, John Matte... · May 07, 2026
A new foundation model for tabular data (Schema-1) learns tables natively and outperforms current tabular methods on prediction and imputation benchmarks; if its benchmark gains transfer to production datasets, it could materially reduce preprocessing overhead and accelerate adoption of AI across data-driven firms.
medium descriptive arxiv
Data Language Models: A New Foundation Model Class for Tabular Data · Eda Erol, Giuliano Pezzoli, Ozer Cem Kelahmet · May 07, 2026
Static fairness rules can worsen long-term inequality while creating high short-term costs, but modest investment policies can erase disparities at low utility loss; the paper shows analytically that fair-seeming constraints may backfire over time and that dynamic interventions offer a cheaper path to equality.
medium theoretical arxiv
Price of Fairness in Short-Term and Long-Term Algorithmic Selections · Shahin Jabbari, Chen Wang · May 07, 2026
Commercial incentives, concentrated compute and data access channelled generative AI toward opaque, frontier models, while open-weight and sovereign alternatives persisted but remained marginal due to resource and institutional constraints; targeted socio-technical programmes could reorient development toward transparency, moderation and sustainable business models.
n/a theoretical arxiv
Pathways to AGI · Gordon Fletcher, Saomai Vu Khan · May 07, 2026
Users notice stylistic differences in AI text but won't pay extra for them: an incentivized experiment finds no link between perceived aesthetic quality and willingness to pay, suggesting aesthetic upgrades are unlikely to justify price premia in current LLM markets.
low quasi_experimental arxiv
Artificial Aesthetics: The Implicit Economics of Valuing AI-Generated Text · Arbaaz Karim · May 07, 2026
A self‑hosted legal small language model matches frontier LLMs on contract extraction and slashes inference costs by up to 97%, reducing hallucinations that drive review burden.
medium descriptive arxiv
A Few Good Clauses: Comparing LLMs vs Domain-Trained Small Language Models on Structured Contract Extraction · Nicole Lincoln, Nick Whitehouse, Jaron Mar, Rivindu Perera · May 07, 2026
China’s government‑guided funds materially speed firms’ digital and intelligent transformation, especially in high‑tech firms and those with strong internal controls; the funds work largely by loosening finance constraints and promoting knowledge spillovers.
medium quasi_experimental openalex
Government-Guided Funds and Corporate Digital–Intelligent Transformation · Fangzheng Zhu, Yuexiang Lu · May 07, 2026
AI increases firm innovation only up to a point: beyond an adoption index of 2.948, innovation falls as firms' patents become more similar; strong internal absorptive capacity delays decline while concentrated industries hasten it.
medium correlational openalex
The Inverted-U Relationship Between AI and Corporate Innovation Performance · Xu Fan, Benye Wang · May 07, 2026
Guideline-aware AI drafts halve the time novice describers need to produce audio descriptions and ease cognitive load, but poor-quality AI output offers little help; effectiveness depends on a content-specific quality threshold.
medium quasi_experimental arxiv
Making AI Drafts Count: A Quality Threshold in Audio Description Workflows · Lana Do, Shasta Ihorn, Charity M. Pitcher-Cooper, Sanjay Mirani, Gio Jung, Hyunj... · May 06, 2026
Automation can lift GDP and consumption if gains are large and ownership is widespread; but when ownership is concentrated and vulnerable, privately chosen automation overexpands and shifts income to capital and high-skill workers, harming low-wealth households.
n/a theoretical arxiv
The Demand Externality of Automation · Erhan Bayraktar · May 06, 2026
Constraint-aware scaffolding makes AI-generated services markedly more deployable and less brittle than generic code-generation workflows, by embedding production constraints via template retrieval and iterative clarification; this aligns AI-assisted prototyping closer to production engineering needs.
medium quasi_experimental arxiv
Architectural Constraints Alignment in AI-assisted, Platform-based Service Development · Julius Irion, Moritz Leugers, Paul Hartwig, Simon Kling, Tachmyrat Annayev, Alex... · May 06, 2026
A curated drug-asset index dramatically improves LLM-powered competitive scouting: Gosset finds 3.2× more verified long-tail oncology assets per query than leading web-search models in a 10-target benchmark. Turning the same curated index into a callable tool lets other frontier models recover most of that advantage, implying search/data coverage — not model bells and whistles — is the main bottleneck.
medium descriptive arxiv
Curated AI beats frontier LLMs at pharma asset discovery · Łukasz Kidziński, Kevin Thomas · May 06, 2026
A transformer trained on patent classification codes can forecast novel technological pairings years — sometimes decades — ahead by detecting collective shifts in how technologies are described, and it also improves representations for common patent analysis tasks.
medium correlational arxiv
Anticipating Innovation Using Large Language Models · Enrico Maria Fenoaltea, Filippo Santoro, Giordano De Marzo, Segun Taofeek Aroyeh... · May 06, 2026
GenAI coding assistants deliver modest productivity gains—about a third of a standard deviation—mostly in controlled experiments, but they do not reliably boost learning or skill acquisition in educational settings.
medium review_meta arxiv
A meta-analysis of the effect of generative AI on productivity and learning in programming · Sebastian Maier, Moritz Gunzenhäuser, Jonas Schweisthal, Manuel Schneider, Stefa... · May 06, 2026
AI that monitors pairs’ effort and gaze makes pair programming more effective: combined reactive prompts improve coordination and performance, and proactive, forecast-based support prevents breakdowns before they occur.
high quasi_experimental arxiv
Cognitive Alignment Drives Attention: Modeling and Supporting Socially Shared Regulation in Pair Programming · Anahita Golrang, Kshitij Sharma · May 06, 2026
DAO-governed stacks could enable permissionless, community-owned AI-driven physical infrastructure by linking energy, sensing, compute and robots, but deployment risks — from security and centralization to incentive failures and legal exposure — mean scalable use will depend on continuously adaptive, value-sensitive governance.
n/a theoretical arxiv
DAO-enabled decentralized physical AI: A new paradigm for human-machine collaboration · Mark C. Ballandies, Florian Spychiger, Uwe Serdült, Claudio J. Tessone · May 06, 2026
Chinese energy‑intensive listed firms that emphasize digital technologies make faster low‑carbon progress, largely by boosting R&D and cutting operating costs. The effect is strongest for large, non‑state firms and for firms in central and western regions.
medium correlational openalex
The Impact of Digital Technology Integration on Low-Carbon Transformation in Energy-Intensive Enterprises: An Empirical Study Based on A-Share Listed Companies in Shanghai and Shenzhen · Zijing Liu · May 06, 2026