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
AI implementation during 2023–2025 was accompanied by a positive contribution to Russia’s potential GDP.
Analysis of aggregated industry data and a scenario approach using Russian-language sources (Ministry of Digital Development, HSE, Digital Economy ANO, analytical reviews).
For memory workloads requiring stable facts and stateful computation, architecture matters more than retrieval scale or model strength alone.
Conclusion drawn by the authors based on comparative experimental results reported in the paper (xmemory vs retrieval/model-strength baselines); excerpt provides aggregate benchmark comparisons but not full experimental details.
On the application-level task, xmemory reaches 95.2% accuracy, outperforming specialised memory systems, code-generated Markdown harnesses, and customer-facing frontier-model application harnesses.
Empirical evaluation on an application-level task reported in the paper showing 95.2% accuracy for xmemory and claiming it outperforms several classes of alternative systems; excerpt lacks details on the task, dataset size, or baseline numeric results.
On the end-to-end memory benchmark, xmemory reaches 97.10% F1, compared with 80.16%-87.24% across the third-party baselines.
Empirical evaluation on the paper's end-to-end memory benchmark reporting F1 scores for xmemory and a range for third-party baselines; the excerpt does not provide dataset size or statistical significance details.
On the structured extraction benchmark (judge-in-the-loop configuration) the system reaches 90.42% object-level accuracy and 62.67% output accuracy, above all tested frontier structured-output baselines.
Empirical evaluation on the paper's structured extraction benchmark in the judge-in-the-loop configuration; the excerpt reports the numeric accuracies and states they exceed tested frontier structured-output baselines. The excerpt does not specify dataset size or number of runs.
This iterative, schema-aware write-path design shifts interpretation from the read path to the write path: reads become constrained queries over verified records rather than repeated inference over retrieved prose.
Conceptual claim about how the proposed architecture affects system behavior; supported by the architectural description in the paper rather than explicit quantitative evidence in the excerpt.
We present an iterative, schema-aware write path that decomposes memory ingestion into object detection, field detection, and field-value extraction, with validation gates, local retries, and stateful prompt control.
Description of the proposed method/architecture in the paper (methodological contribution); no numeric evaluation attached to the description in the excerpt.
Reliable external AI memory must be schema-grounded (schemas define what must be remembered, what may be ignored, and which values must never be inferred).
Normative assertion supported by the paper's proposed design and subsequent experimental results (the paper introduces a schema-grounded approach and evaluates it against benchmarks), though the excerpt does not give full methodological details or sample sizes for this claim alone.
To manage AI legibility, creators perform four recurring forms of invisible authenticity labor: epistemic verification, linguistic naturalization, narrative restructuring, and performative embodiment.
Authors identify and name four recurrent practices from coding and analysis of 16 in-depth interviews with creators on Xiaohongshu and Douyin describing specific downstream repair and performance work.
Creators engage in 'AI passing': strategic efforts to conceal and humanize AI-assisted drafts so that outputs plausibly appear human-authored.
Concept introduced based on analysis of 16 in-depth interviews with creators on Xiaohongshu and Douyin describing tactics to hide AI involvement and present content as human-authored.
Claw-Eval-Live suggests that workflow-agent evaluation should be grounded twice, in fresh external demand and in verifiable agent action.
Conclusion/recommendation drawn from the benchmark design and experimental findings; conceptual claim advocating evaluation grounded in external demand signals and verifiable actions.
The release contains 105 tasks spanning controlled business services and local workspace repair, and evaluates 13 frontier models under a shared public pass rule.
Benchmark release statistics reported in the paper: explicit counts of tasks and evaluated models (105 tasks; 13 models).
For grading, Claw-Eval-Live records execution traces, audit logs, service state, and post-run workspace artifacts, using deterministic checks when evidence is sufficient and structured LLM judging only for semantic dimensions.
Grading methodology described in the paper: instrumentation and hybrid deterministic/LLM-judging approach documented by authors (procedural description).
Each release is constructed from public workflow-demand signals, with ClawHub Top-500 skills used in the current release, and materialized as controlled tasks with fixed fixtures, services, workspaces, and graders.
Description of release construction in the methods: uses public workflow-demand data and ClawHub Top-500 skills; tasks are materialized with controlled fixtures and graders (procedural detail from the paper).
We introduce Claw-Eval-Live, a live benchmark for workflow agents that separates a refreshable signal layer, updated across releases from public workflow-demand signals, from a reproducible, time-stamped release snapshot.
Methodological contribution described in the paper; design and architecture of the benchmark are presented by the authors (design description, no external sample needed).
LLM agents are expected to complete end-to-end units of work across software tools, business services, and local workspaces.
Framing/background statement in the paper describing expected capabilities of workflow agents; no empirical sample size reported for this expectation.
The proposed framework balances AI-driven productivity with the epistemic sovereignty necessary to manage increasingly opaque software ecosystems.
Normative/architectural claim about the proposed framework; presented conceptually in the paper without reported empirical testing in the excerpt.
To preserve long-term resilience, engineering leaders must move beyond prompt-based development to implement rigorous human-in-the-loop pedagogical standards.
Prescriptive recommendation based on the paper's conceptual analysis; no randomized trials or empirical validation of this intervention reported in the excerpt.
Commonly reported gains include the automation of trivial and repetitive tasks.
Multiple studies in the review report that LLM-assistants automate mundane programming tasks.
Commonly reported gains include minimized code search due to LLM assistance.
Synthesis of study findings noting reductions in developer time spent searching for code or answers.
Commonly reported gains from LLM-assistants include accelerated development (faster task completion).
Multiple included studies report faster development workflows and reduced time-to-complete tasks, as synthesized in the review.
The majority of reviewed studies report considerable benefits from LLM-assistants.
Synthesis of findings across the 39 included peer-reviewed studies as reported in the review.
Framework, metrics, baselines, and collection scripts will be released open-source on acceptance.
Author statement of intent to release code and assets upon paper acceptance.
The paper describes three reference architectures (ColPali, ColQwen2, agentic complexity-based routing) which are not yet integrated end-to-end.
Author statement listing three proposed reference architectures and noting they are not yet integrated end-to-end.
Factual accuracy on stated claims is 85.5%.
Reported accuracy measurement on 'stated claims' in generated outputs from systems evaluated on EnterpriseDocBench. Details on annotator process and sample size not included in excerpt.
Both hybrid and BM25 beat dense embedding (dense embedding nDCG@5 = 0.83).
Reported nDCG@5 values for three retrieval approaches on the benchmark (values quoted in paper).
Hybrid retrieval narrowly beats BM25 (nDCG@5 of 0.92 vs. 0.91).
Empirical evaluation on EnterpriseDocBench using nDCG@5 as reported metric in paper. Exact query count or folds not provided in the excerpt.
We ran three pipelines through it: BM25, dense embedding, and a hybrid, all using the same GPT-5 generator.
Method statement describing experimental pipelines evaluated on EnterpriseDocBench (three retrieval variants combined with a shared generator).
The corpus is built from public, permissively licensed documents across six enterprise domains (five represented in the current pilot).
Author description of corpus composition (number of domains and pilot coverage). No document-count supplied in provided text.
We built EnterpriseDocBench to evaluate parsing fidelity, indexing efficiency, retrieval relevance, and generation groundedness on the same corpus.
Description of dataset/benchmark creation and stated design goals in paper (author-developed benchmark covering four stages).
Most enterprise document AI today is a pipeline: parse, index, retrieve, generate.
Author assertion about prevailing architecture of enterprise document-AI systems (introductory observation in paper). No empirical sample size or systematic survey reported in text provided.
We release the full code base and a richly annotated dataset to support reproducible research on adaptive VCAs.
Paper statement announcing release of code and dataset.
The recommender achieved high relevance (MRR@1=0.75).
Reported offline/online recommender evaluation in the paper using Mean Reciprocal Rank at 1 (MRR@1) metric; presumably computed over recommendations in the study (711 conversations).
Step-by-step guidance improved pleasantness and reduced user burden.
User-reported measures collected in the controlled study (likely subjective ratings across participants/conversations).
Device-level evidence increased correct resolutions from about 50% to over 90% relative to an LLM-only baseline.
Controlled study comparing SecMate with device-level diagnostic evidence to an LLM-only baseline; reported results across 144 participants / 711 conversations.
Service specificity is achieved through a proactive, context-aware recommender.
System description and recommender component evaluation in the paper.
User specificity relies on implicit proficiency inference and profile-aware troubleshooting.
System design and algorithmic description in the paper explaining user-proficiency inference and profile-aware components.
Device specificity is provided by a lightweight local diagnostic utility.
System design and implementation details reported in the paper describing the diagnostic utility component.
We present SecMate, a multi-agent VCA for cybersecurity troubleshooting that integrates device, user, and service specificity from conversational and device-level signals.
System description and architecture presented in the paper (design and implementation of SecMate).
The framework produces a list of testable empirical questions that we leave as open problems.
Statement in the paper that it derives testable empirical questions from the theoretical framework; no empirical tests are executed in the paper itself.
The framework operationalizes aspects of earlier qualitative work on supervisory control (Sheridan, 1992), common ground (Clark & Brennan, 1991), and mixed-initiative interaction (Horvitz, 1999) within a single normative ratio.
Conceptual synthesis and mapping of prior qualitative literature into the new per-task leverage formalism presented in the paper; this is a theoretical linkage rather than empirical validation.
The per-task ceiling does not bind the windowed measure, though both remain bounded: L_task by per-task novelty, L_window by the stock of accumulated planning investment that pays out within the window.
Theoretical derivation/argument in the paper distinguishing bounds on per-task leverage (L_task) and windowed leverage (L_window) and identifying their respective limiting factors; no empirical evidence provided.
We extend this per-task analysis to a windowed leverage measure that accommodates recurring tasks, spawned subtasks, and amortized system-design investment.
Conceptual/theoretical extension in the paper defining a windowed leverage metric and describing how it accounts for recurring tasks, subtasks, and amortized design investments; no empirical tests reported.
The asymptotic behavior of leverage decomposes into two scaling axes (capability and memory) with a non-zero floor on the planning term set by irreducible task novelty bounded by human throughput.
Mathematical/theoretical asymptotic analysis within the paper; conceptual derivation linking capability and memory as scaling axes and asserting a lower bound on planning cost due to task novelty and human throughput.
Information density itself is directional and bounded by separate ceilings on human-to-agent and agent-to-human flow.
Theoretical argument/derivation in the paper establishing directional information-density and distinct upper bounds for each flow direction; no empirical validation reported.
The denominator decomposes into three channels through which a conserved per-task information requirement must flow, each with its own time-cost scalar (specify the task, resolve mid-run interrupts, and review the result).
Analytic decomposition within the paper's theoretical framework; conceptual argument rather than empirical measurement.
We propose a per-task leverage ratio for human-agent collaboration: human work displaced by an agent, divided by the human time required to specify the task, resolve mid-run interrupts, and review the result.
Theoretical/conceptual proposal and formal definition provided in the paper; no empirical sample or experimental data reported.
PD--RSAC maintained zero feeder-limit violations in the experiments.
Empirical reporting in the paper's experimental results on the simulator that PD--RSAC had zero feeder-limit violations while operating under the formulated constraints.
Experiments on a large-scale EV fleet simulator built from NYC taxi data show that PD--RSAC achieves the highest net profit, reaching $1.22M.
Empirical results reported from experiments run on a large-scale EV fleet simulator constructed from NYC taxi data; PD--RSAC reported net profit of $1.22M in these experiments (paper's experimental section).
In clinical utility evaluation across three abstraction tasks, semantic search reduced time-to-completion by 24 to 89% compared to clinician-performed chart review.
Clinical utility assessment compared chart abstraction efficiency across three tasks and reported percentage reductions in time-to-completion ranging from 24% to 89%.