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
On BioML-Bench, spanning biomedical imaging, protein engineering, single-cell omics, and drug discovery, AutoScientists achieves a mean leaderboard percentile of 74.4% across 24 tasks, improving over the strongest AI agent by +8.33%.
Empirical evaluation on the BioML-Bench benchmark (24 tasks); reported mean leaderboard percentile and comparative improvement versus the strongest baseline agent.
Under matched experimental budgets, AutoScientists improves over prior AI agents across biomedical machine learning, language-model training optimization, and protein fitness prediction.
Empirical comparisons reported in paper across multiple benchmark suites and tasks (BioML-Bench, GPT training optimization experiments, ProteinGym).
AutoScientists is a decentralized team of AI agents that interpret a shared experimental state, self-organize into teams around promising hypotheses, critique proposals before using experimental compute, and share successes and failures to reduce redundant exploration.
System design and implementation described in the paper (architecture and agent protocols); qualitative description of agent behaviors and coordination mechanisms; demonstrated in experiments.
We describe the benchmark design, evaluation protocol, and quality-control pipeline, and position OR-Space as a benchmark for studying the reliability, failure modes, and practical readiness of LLM agents in industrial OR workflows.
Statement of the paper's contributions and contents (methodological description of what the paper includes).
By combining persistent workspaces with lifecycle-oriented tasks, OR-Space evaluates whether agents can perform reliable optimization work beyond end-to-end text generation.
Stated objective/claim in the paper about the benchmark's purpose and what it measures (conceptual/goal-oriented statement).
OR-Space defines an Explain task mode, where agents answer grounded questions about solutions, constraints, and business implications using evidence spread across workspace artifacts.
Definition of the Explain task mode provided in the paper (design/specification).
OR-Space defines a Revise task mode, where agents modify existing models under changing requirements or solver feedback while preserving valid prior logic.
Definition of the Revise task mode in the benchmark design (descriptive claim in the paper).
OR-Space defines three task modes: Build, where agents construct solver-ready optimization models from heterogeneous artifacts.
Definition of one of the benchmark's task modes as described in the paper (method/design description).
Each instance is an executable workspace containing business documents, structured data, optional code artifacts, solver outputs, and task-specific evaluators distributed across interdependent files.
Design specification of OR-Space provided in the paper (descriptive claim about benchmark instance structure).
We introduce OR-Space, a full-lifecycle workspace benchmark for evaluating industrial optimization agents across model construction, model revision, and grounded explanation.
Paper presents and names a new benchmark (methodological contribution described directly in the text).
Large language model (LLM) agents are increasingly used to assist with operations research (OR) modeling.
Statement in the paper asserting an observed trend; likely based on literature/context motivating the work (no empirical sample or quantitative citation provided in the excerpt).
A recommended organizational design for the AI era is the 'resonance protocol enterprise' in which structures are temporary crystallizations, AI governance protects adaptive openness, and legitimacy derives from sustaining recursive renewal.
Normative/proposal in the paper outlining a new organizational design paradigm; presented as conceptual design without empirical pilot or evaluation.
Digital transformation initially enhanced adaptability by fluidifying information flows and expanding relational connectivity, thereby improving some organizations' adaptability.
Theoretical claim supported by qualitative interpretation of digital transformation phenomena; no systematic measurement or reported sample.
Organizations capable of rapid relational reconfiguration, customer reconnection, and generative experimentation often proved more resilient during the pandemic.
Illustrative/theoretical interpretation of pandemic cases offered in the paper; no quantified sample or formal empirical evidence reported.
Although AI creates obstacles, it also has the potential to be an important tool for creating innovative opportunities and continued growth if managed with sound practices.
Concluding statement in the paper's abstract presenting a normative/conditional conclusion based on the paper's evaluation and synthesis of evidence (no primary quantified results provided in the supplied text).
AI leads to the creation of new jobs.
The paper explicitly states it examines the creation of new jobs as a ramification of AI (abstract); claim presented qualitatively without reported sample sizes or quantified effect in the provided text.
GENESIS is built on three composable primitives (agents, skills, hooks) and a knowledge layer (SYNAPSE) that doubles as the source of ground truth and the recipient of every artifact the framework produces, making capabilities compound across runs.
Architectural description in the paper; claim about knowledge base acting as ground truth and enabling capability compounding (design-level claim). No quantitative evaluation given in the abstract.
GENESIS is an agentic AI framework that converts intents (e.g., a specification clause, a telemetry anomaly, or a research hypothesis) into solutions validated with over-the-air experiments, fed back into a persistent knowledge base.
System design / implementation claim presented in the paper (description of proposed framework). The abstract does not report empirical evaluation metrics or sample size.
Large Language Models (LLMs) have compressed comparable R&D work in general software engineering from days to minutes.
Paper's stated comparison/claim (likely based on prior reports or authors' experience); no experimental details or sample size provided in the abstract.
Agentic Technical Debt and Stochastic Tax are related but distinct: debt can amplify the tax.
Theoretical relationship asserted in the structural model; the note states debt can amplify the recurring Stochastic Tax and provides model expressions and discussion (and illustrative simulation) to substantiate the relationship.
Combining both levers yields a 502% improvement on single-cell RNA denoising over the initial baseline.
Reported experimental result in the paper comparing SIA to the initial baseline on the single-cell RNA denoising task (denoising metric unspecified in abstract).
Combining both levers yields a 91.9% runtime reduction on GPU kernels over the initial baseline.
Reported experimental result in the paper comparing SIA to the initial baseline on the low-level GPU kernel optimisation task (runtime measured).
Combining both levers yields a 56.6% gain on LawBench (Chinese legal charge classification) over the initial baseline.
Reported experimental result in the paper comparing SIA to the initial baseline on the LawBench task.
Combining both levers (harness updates and weight updates) outperforms scaffold iteration alone on all three benchmarks.
Empirical comparison reported in the paper: experiments across the three domains comparing SIA (combined harness+weight updates) against scaffold-iteration-only baseline.
These results show that per-query configuration of the full retrieval pipeline is a practical alternative to static workload-level tuning.
Authors' conclusion drawn from the reported empirical results on MuSiQue, BrowseComp-Plus, and FinanceBench demonstrating BRANE's performance advantages.
BRANE outperforms LLM-routing, rule-based, and fine-tuned Qwen3-4B baselines.
Empirical comparisons against specified baselines (LLM-routing, rule-based approaches, and fine-tuned Qwen3-4B) across the reported benchmark sets. The text does not provide numeric performance metrics or sample sizes in the excerpt.
BRANE matches the best fixed configuration's accuracy at up to 89% lower cost.
Empirical result reported by the authors based on experiments on the named benchmarks (MuSiQue, BrowseComp-Plus, FinanceBench). The provided text states the magnitude ('up to 89% lower cost') but does not give sample sizes or confidence intervals.
Across MuSiQue, BrowseComp-Plus, and FinanceBench, BRANE consistently pushes the cost-quality Pareto frontier.
Empirical evaluation reported on three benchmark suites: MuSiQue, BrowseComp-Plus, and FinanceBench. The claim is based on experimental comparisons across these datasets; the paper does not state numeric sample sizes in the provided text.
Hybrid Fusion significantly accelerated the recovery of smaller Slow AI teams (+6.9% at N=4).
Reported intervention result: Hybrid Fusion produced a +6.9% acceleration in recovery for smaller Slow AI teams, reported at N=4.
Integrating these isolated veridical signals via Hybrid Fusion successfully rescued the Fast AI team (+7.6% at N=8).
Reported intervention result: application of Hybrid Fusion integration produced a +7.6% improvement in Fast AI team performance, reported at N=8.
The Riemannian Oracle adapted to task states by heavily restricting temporal windows (< 0.8s) to intercept fast reflexive compliance and widening windows (> 1.2s) to capture delayed cognitive conflict.
Reported algorithmic behavior of the 2D Adaptive Riemannian Oracle in response to measured spatial covariance: window sizes described as <0.8s for fast states and >1.2s for slow states.
In the Slow AI condition, behavioural teams (N=8) eventually recovered to 100.0%.
Reported team performance metric for behavioural teams in Slow AI condition with N=8; team performance reported to reach 100.0%.
The proposed policy framework contributes to establishing a foundation for Vietnam to proactively embrace the Agent Economy safely and effectively.
Claim in abstract about the intended contribution/impact of the proposed framework; no empirical evaluation or measured outcomes presented.
The Agent Economy promises substantial gains in productivity and innovation.
Asserted in paper abstract as an anticipated outcome; no empirical measurement, sample size, or quantified effect provided.
Future evaluations should use artifact-level denominators, reproducible parsing rules, correction taxonomies, and independent coding of governance events.
Authors' recommendations based on methodological lessons from this structured self-observed implementation case study and observed parsing/governance challenges.
AI assistance can stabilize an overloaded workflow only when (i) the fraction of tasks handled by AI exceeds a critical threshold, and (ii) the human attention required for review and expected rework is lower than the attention required for manual completion.
Formal analytical conditions derived from the paper's queueing model (model-based theoretical result; no empirical sample reported).
The paper calls for action by stakeholders to consider human and environmental moderators when adopting AI.
Policy/recommendation statement in the paper's conclusion/abstract; normative recommendation rather than empirical finding.
We revise the existing framework to redefine effective organizational determinants and shed light on practical implications including industry and education.
Authors' proposed theoretical revision of an existing framework and discussion of implications; presented as a conceptual contribution within the paper.
Most practitioners assume that AI brings productivity boosts owing to enhanced technical capabilities.
Statement of common practitioner belief reported by the authors in the paper's framing; no supporting survey or sample reported in the abstract.
Adoption of Claude Code increases cumulative lifetime languages used by +0.51.
Panel analysis of 5,838 developers over 28 months using the Callaway & Sant'Anna estimator; treatment = first Claude-co-authored commit.
Adoption of Claude Code increases the count of newly-used languages by +0.31.
Same dataset and staggered-rollout estimator (Callaway & Sant'Anna), treatment = first Claude-co-authored commit; not-yet-treated controls.
Adoption of Claude Code increases Shannon language entropy by +0.14.
Estimated with the doubly robust Callaway & Sant'Anna approach on the 5,838-developer panel over 28 months, using first Claude-co-authored commit as treatment.
Adoption of Claude Code increases the number of distinct programming languages used by a developer by +0.83.
Same panel and staggered-rollout estimation as above (Callaway & Sant'Anna), treatment = first Claude-co-authored commit.
Adoption of Claude Code increases the number of repositories a developer contributes to by +1.5 (monthly).
Same panel (5,838 developers, 28 months) and estimator (Callaway & Sant'Anna). Treatment = first Claude-co-authored commit; not-yet-treated controls.
Adoption of Claude Code is associated with an increase of +41 monthly commits per developer.
Analysis of a panel of 5,838 GitHub developers observed monthly over 28 months, exploiting staggered rollout of Claude Code (May 2025–Jan 2026). Treatment defined by developer's first Claude-co-authored commit; not-yet-treated developers used as controls. Estimates from the doubly robust Callaway and Sant'Anna (2021) staggered-difference-in-differences estimator.
Structured AI-based interventions provide causal evidence that they can transform access to scientific feedback from a largely private advantage into a more widely distributed resource.
Causal inference based on randomized field experiment showing increased revision likelihood and broader uptake of LLM tools across diverse regions and author groups.
Effects were strongest among teams with lower h-indexes and earlier career stages.
Heterogeneous treatment effects by team-level metrics (h-index) and career stage reported in the randomized experiment.
Effects were strongest for manuscripts less embedded in the scholarly literature.
Heterogeneous treatment effects reported by manuscript-level embedding in literature (e.g., referencing/citation context) within the randomized experiment.
Effects of AI feedback were strongest among authors from non-English-dominant research regions.
Heterogeneous treatment effects reported in the randomized experiment stratified by authors' geographic / language-dominance region; sample includes authors from 133 geographic regions.
Exposure to AI feedback increased authors' subsequent use of LLM tools in their future papers, suggesting longer-run shifts in scientific practice.
Follow-up measurements in the randomized field experiment tracking authors' later behavior (use of LLM tools in subsequent papers); comparison between treatment and control authors.