Evidence (8974 claims)
Search and filter individual claims pulled from the papers. Looking for a specific finding ("what's the effect on wages?"), you're in the right place. Want to compare whole outcome categories against each other instead? Use the Evidence Explorer.
The board below groups claims two ways: by broad theme (nine paper-level topics) and by outcome category (the 34 claim-level outcomes that the Explorer and Syntheses also use).
Browse by theme
Nine broad, paper-level topics. Click one to filter the claims below.
Adoption
10085 claims
Filter claims →
Productivity
8974 claims
Filtered →
Governance
8062 claims
Filter claims →
Human-AI Collaboration
7749 claims
Filter claims →
Org Design
5057 claims
Filter claims →
Innovation
4896 claims
Filter claims →
Labor Markets
4088 claims
Filter claims →
Skills & Training
3372 claims
Filter claims →
Inequality
2377 claims
Filter claims →
Claims by outcome category
Counts by direction of finding. These are the same 34 outcome categories the Explorer compares and the Syntheses are written for. A linked row has a published synthesis.
| Outcome | Positive | Negative | Mixed | Null | Total |
|---|---|---|---|---|---|
| Other | 882 | 244 | 117 | 1097 | 2424 |
| Governance & Regulation | 1010 | 469 | 229 | 135 | 1875 |
| Organizational Efficiency | 977 | 235 | 149 | 90 | 1462 |
| Technology Adoption Rate | 781 | 299 | 143 | 128 | 1362 |
| Research Productivity | 506 | 155 | 74 | 363 | 1110 |
| Output Quality | 555 | 219 | 71 | 70 | 915 |
| Decision Quality | 395 | 200 | 95 | 54 | 751 |
| Firm Productivity | 523 | 67 | 101 | 27 | 724 |
| AI Safety & Ethics | 262 | 309 | 75 | 36 | 688 |
| Market Structure | 195 | 201 | 135 | 30 | 566 |
| Task Allocation | 248 | 77 | 96 | 38 | 464 |
| Innovation Output | 300 | 34 | 55 | 20 | 411 |
| Skill Acquisition | 207 | 75 | 65 | 21 | 368 |
| Employment Level | 138 | 67 | 119 | 24 | 350 |
| Fiscal & Macroeconomic | 156 | 80 | 53 | 33 | 329 |
| Task Completion Time | 211 | 38 | 13 | 16 | 280 |
| Firm Revenue | 183 | 52 | 29 | 5 | 270 |
| Consumer Welfare | 131 | 77 | 48 | 13 | 269 |
| Inequality Measures | 50 | 141 | 54 | 9 | 254 |
| Worker Satisfaction | 104 | 85 | 25 | 13 | 227 |
| Error Rate | 87 | 112 | 11 | 5 | 215 |
| Automation Exposure | 69 | 69 | 37 | 20 | 198 |
| Wages & Compensation | 102 | 49 | 31 | 11 | 193 |
| Team Performance | 115 | 30 | 30 | 11 | 187 |
| Regulatory Compliance | 88 | 74 | 17 | 7 | 186 |
| Training Effectiveness | 109 | 22 | 14 | 21 | 168 |
| Developer Productivity | 116 | 21 | 15 | 8 | 161 |
| Job Displacement | 12 | 92 | 26 | 1 | 131 |
| Hiring & Recruitment | 57 | 12 | 9 | 5 | 83 |
| Skill Obsolescence | 6 | 59 | 10 | 2 | 77 |
| Social Protection | 43 | 17 | 8 | 2 | 70 |
| Creative Output | 35 | 21 | 9 | 4 | 70 |
| Labor Share of Income | 18 | 23 | 17 | 1 | 59 |
| Worker Turnover | 15 | 16 | — | 4 | 35 |
| Industry | — | — | — | 1 | 1 |
Productivity
Remove filter
Meta-analyses show that AI assistance tends to improve human performance compared to working alone.
Reference to existing meta-analyses in the literature reported by the authors (meta-analytic evidence aggregated across studies; no specific meta-analysis names, sample sizes, or quantitative pooled effects provided in the excerpt).
AI is now embedded in healthcare, finance, policy, and many other domains.
Statement in the paper's introduction/abstract summarizing the current deployment of AI across domains (literature observation, no specific empirical study or sample size cited).
Agentic AI does not eliminate engineering discipline; it increases the value of requirements, constraints, traceability, independent verification, and human approval.
Conclusion drawn from synthesis of evidence across multiple domains and argumentation in the paper.
Agentic Agile-V and the task-level SCOPE-V loop (Specify, Constrain, Orchestrate, Prove, Evolve, Verify) convert conversational intent into structured engineering artifacts and acceptance evidence.
The paper proposes this process framework (the claim is the proposed function of the framework; no empirical evaluation given in the abstract).
Controlled studies report productivity gains in some enterprise tasks.
Controlled experimental studies referenced by the paper (specific trials/stats not provided in abstract).
These capabilities make software and hardware development faster in some settings.
Aggregated evidence cited in the paper including controlled studies and adoption studies (details not specified in abstract).
Agentic AI coding systems can inspect repositories, plan implementation steps, edit files, call tools, run tests, and submit pull requests.
Descriptive synthesis of existing agentic systems and demonstrations referenced in the paper (literature/examples); no single study or sample size given in the abstract.
The Claude family leads the benchmark and produces the most professional-looking outputs in our qualitative review.
Empirical result reported from the paper's benchmark and qualitative review of agent outputs (specific metrics, number of agents/tasks, and quantitative scores not provided in the excerpt).
We develop an evaluation taxonomy comprising three dimensions: Accuracy, Formula, and Format, each comprising fine-grained criteria that reflect professional standards.
Methodological contribution stated in paper; described taxonomy elements (Accuracy, Formula, Format) as part of the evaluation design.
We provide one of the first evaluations of agents on end-to-end spreadsheet tasks, focusing on economically critical financial workflows such as modeling and scenario analysis.
Claim of contribution in the paper; refers to the authors' own evaluation study (details like number of tasks/agents not provided in the excerpt).
Frontier AI labs have developed agents that can construct entire spreadsheets from scratch.
Asserted in paper as background/context; no specific models, numbers, or experimental details provided in the excerpt.
LLM agents are increasingly expected to carry out end-to-end workflows, producing complete artifacts from high-level user instructions.
Framing statement in paper; no empirical data or sample size reported to support the trend claim within the excerpt.
A six-phase, stepwise implementation framework (ABC-XYZ segmentation, forecast model selection, safety stock calibration, replenishment policy assignment, simulation-based parameter tuning, KPI governance) enables enterprises to achieve 9–16% reductions in inventory costs within existing WMS and ERP architectures.
Practical implications presented in the paper proposing a six-phase implementation framework and asserting expected inventory cost reductions of 9–16% when deployed within existing WMS/ERP.
Learning-based control methods deliver up to 16% cost reductions under complex network conditions but require substantial data and governance infrastructure.
Findings from included studies (narrative and/or quantitative results) reporting maximum observed reductions 'up to 16%' and qualitative synthesis noting data/governance requirements.
The cost reduction from multi-echelon coordination increases significantly with network complexity and lead-time variability.
Pre-specified moderator analyses reported in the paper showing effect size growth with network complexity and lead-time variability.
Multi-echelon coordination yields a pooled mean cost reduction of 11.4% (95% CI: 6.9–15.9%).
Random-effects meta-analysis pooling percentage cost-reduction effect sizes (reported pooled mean and 95% CI).
The advantage of distributional safety stock methods is largest for high-variability SKU segments.
Pre-specified subgroup and moderator analyses reported in the paper indicating greater pooled effects in high-variability SKU segments.
Distributional safety stock methods outperform classical normal approximations by a pooled mean of 9.3% (95% CI: 5.8–12.7%) at equivalent service levels.
Random-effects meta-analysis pooling percentage cost-reduction effect sizes (reported pooled mean and 95% CI).
Given the mixed outcomes (some improvements, some new lint/security issues), stronger tool-in-the-loop quality and security gating is motivated for AI-driven development workflows.
Interpretation/recommendation based on observed mix of improvements and introduced issues from the empirical results (PyQu, Pylint, Bandit analyses) and high merge rates.
73.5% of the analyzed PRs are merged (developer acceptance is high).
Empirical measurement of PR outcomes (merged vs. not merged) in the AIDev dataset of Python refactoring PRs.
Usability is the quality attribute that improves most frequently, improving in 36.5% of the studied changes.
PyQu-based before-and-after analysis of quality attributes on Python refactoring PRs from the AIDev dataset; reported frequency for the 'usability' attribute.
Agentic commits improve a quality attribute in 22.5% of the studied changes.
Empirical analysis of Python refactoring pull requests from the AIDev dataset using PyQu (an ML-based Python quality assessment tool) to compare quality attributes before and after each change.
Current models demonstrate promising spatial grounding, multimodal alignment, and coordinated action execution.
Qualitative and/or quantitative evaluation results in paper indicating strengths in spatial grounding, multimodal alignment, and coordinated action execution.
We develop a lightweight parser that transforms raw screen recordings and low-level interaction logs into structured, compositional GUI action trajectories with precise grounding.
Methodological contribution described in paper: parser implementation that converts recordings and logs into structured GUI action trajectories.
The tasks involve dense multimodal interfaces and tightly coupled interaction sequences.
Task descriptions and dataset characteristics in paper stating tasks are complex, long-horizon, multimodal, and tightly coupled.
We curate expert demonstrations across 7 professional applications (e.g., Premiere Pro, Photoshop), covering 186 complex, long-horizon tasks grounded in authentic editing workflows.
Dataset construction reported in paper: curated expert demonstrations spanning 7 applications and 186 tasks (numbers provided in text).
We introduce Cutverse, a benchmark designed to systematically evaluate autonomous GUI agents in realistic media post-production environments.
Paper describes the creation of the Cutverse benchmark as a central contribution (design and implementation described in methods).
GUI agents have made significant progress in web navigation and basic operating system tasks.
Background claim stated in paper referencing prior work on GUI agents applied to web navigation and OS tasks (no specific experiments in this paper to support it).
The architecture successfully manages profiles with 14,000+ scientific facts (125k tokens), enabling sustained operation beyond full-context limits.
Reported stress test / capability demonstration in paper: profile size stated as 14,000+ facts and 125k tokens stored and managed by the system.
The Dual Process system maintains 70-85% accuracy with 1-2 second latency while using 62% fewer tokens (45,434 vs 120,000+ limit) compared to full-context approaches.
Reported empirical results from the large-scale evaluation (1,440 queries / 15,000 messages) comparing Dual Process to full-context models; exact accuracy, latency, and token-count figures provided in the paper.
The Dual Process Memory Architecture decouples immediate episodic needs (constant 10-message window) from long-term consolidated knowledge (growing at approximately 3 tokens/message).
System design description and measured consolidation growth rate reported in the paper; empirical observation of growth rate stated.
Agent Skills, structured packages of procedural knowledge loaded into an LLM agent at inference time, are widely reported to improve task pass rates by an average of 16.2 percentage points across diverse domains.
Authors cite prior Skills benchmarks / aggregated reports (benchmark summary referenced in paper); average improvement reported as 16.2 percentage points across tasks in those benchmarks (implied sample of tasks from the referenced benchmark).
Software products and software R&D contributed 50 percent of the 1.2 percentage point acceleration in nonfarm business labor productivity (2017–2024 relative to 2012–2017).
Empirical decomposition comparing productivity growth rates across periods (2017–2024 vs 2012–2017) in the paper; the authors attribute half of the observed 1.2 percentage point acceleration to software products and software R&D.
Software products and software R&D contributed 50 percent of the 2 percent average growth rate in nonfarm business labor productivity from 2017 to 2024.
Empirical decomposition of nonfarm business labor productivity growth in the United States for the period 2017–2024 reported in the paper (the authors attribute shares of the observed 2% average growth to components including software products and software R&D).
AI is already materially affecting official productivity measures in the United States.
Empirical decomposition of U.S. productivity data reported in the paper that attributes portions of measured productivity growth to software-related channels linked to AI.
Using a framework that separates upstream innovation from downstream production suggests that AI boosts both upstream total factor productivity and intangible capital use downstream.
Model/framework decomposition in the paper (theoretical separation of upstream vs downstream, combined with empirical application to productivity data); the paper reports results consistent with increases in upstream TFP and downstream intangible capital use.
Code cleanliness joins model choice, harness, and prompting as a factor that materially affects agent behaviours.
Conclusion drawn from experimental findings that cleanliness materially influenced agent operational metrics (tokens and revisits) even when pass rates were unchanged.
Traditional maintainability principles remain highly relevant in the era of AI-driven development, shaping the computational cost and navigational efficiency of coding agents.
Interpretation based on experimental results showing token and navigational efficiency gains on cleaner code (7–8% fewer tokens, 34% fewer revisitations) despite unchanged pass rates.
Agents working on cleaner code reduce file revisitations by 34%.
Empirical measurement across the same experimental trials comparing agent file-revisitation counts between clean and messy repo variants; reported 34% reduction in file revisitations on cleaner code.
Agents working on cleaner code use 7 to 8% fewer tokens.
Empirical measurement across trials (660 trials with Claude Code) comparing token consumption between clean and messy repository variants; reported decrease of 7-8% in tokens when working on cleaner code.
We author 33 tasks across six such pairs, evaluated through hidden tests at the application's public surface.
Reported experimental design: 33 authored tasks spanning six repository pairs; evaluation used hidden tests executed at the application's public surface.
The pairs are constructed in both directions, by agent pipelines that either degrade a clean repository or clean a messy one.
Method description: authors constructed pairs bidirectionally using agent pipelines that modify repositories to create matched clean/messy variants.
We introduce an evaluation protocol built around minimal pairs: repositories that match on architecture, dependencies, and external behaviour, but differ on static-analysis rule violations and cognitive complexity.
Methodological description in paper: construction of paired repositories controlling for architecture, dependencies, and external behaviour while varying static-analysis violations and cognitive complexity.
A simple prompt checklist can improve LLM responses while reducing unnecessary interaction.
Authors' interpretation/conclusion drawn from the experimental comparisons and rubric scores reported in the paper's results.
Checklist prompts produced the best quality-effort tradeoff, using fewer average tokens than both raw and clarifying prompts.
Reported comparative statement in the results that checklist prompts used fewer average tokens and produced a better quality-effort tradeoff (no token counts, sample size, or statistical tests reported in the abstract).
Checklist-improved prompts achieved the highest mean rubric score, 7.50 out of 8, compared with 5.67 for raw prompts and 6.67 for clarifying-question prompts.
Reported mean rubric scores for each prompt condition in the paper's results (no sample sizes or significance tests provided in the abstract).
The authors open-source optimize_anything with support for multiple backends as part of the GEPA project at https://github.com/gepa-ai/gepa.
Explicit statement and provided GitHub URL in the paper excerpt.
Multi-task search outperforms independent optimization given equivalent per-problem budget through cross-task transfer, with benefits scaling with the number of related tasks.
Reported experiments comparing multi-task search versus independent per-problem optimization under equal per-problem budget; observed cross-task transfer benefits and that benefits increase with more related tasks.
Ablations across three domains reveal that actionable side information yields substantially higher final scores than score-only feedback.
Same ablation studies across three domains as above; reported higher final optimization scores when using actionable side information compared to only score feedback.
Ablations across three domains reveal that actionable side information yields faster convergence than score-only feedback.
Paper reports ablation studies in three domains comparing optimization with actionable side information versus score-only feedback and finds faster convergence with side information.