Evidence (14922 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
9047 claims
Filter claims →
Productivity
8066 claims
Filter claims →
Governance
7278 claims
Filter claims →
Human-AI Collaboration
6912 claims
Filter claims →
Org Design
4439 claims
Filter claims →
Innovation
4359 claims
Filter claims →
Labor Markets
3652 claims
Filter claims →
Skills & Training
3018 claims
Filter claims →
Inequality
2160 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 | 795 | 210 | 105 | 955 | 2131 |
| Governance & Regulation | 886 | 414 | 197 | 126 | 1654 |
| Organizational Efficiency | 826 | 204 | 129 | 87 | 1257 |
| Technology Adoption Rate | 681 | 259 | 128 | 110 | 1189 |
| Research Productivity | 464 | 138 | 65 | 349 | 1028 |
| Output Quality | 503 | 196 | 61 | 53 | 813 |
| Decision Quality | 351 | 180 | 84 | 51 | 673 |
| AI Safety & Ethics | 238 | 288 | 71 | 34 | 637 |
| Firm Productivity | 455 | 58 | 92 | 20 | 631 |
| Market Structure | 186 | 172 | 123 | 25 | 511 |
| Task Allocation | 222 | 70 | 76 | 34 | 407 |
| Innovation Output | 238 | 28 | 48 | 18 | 334 |
| Skill Acquisition | 177 | 62 | 62 | 17 | 318 |
| Employment Level | 107 | 57 | 108 | 13 | 287 |
| Fiscal & Macroeconomic | 135 | 72 | 44 | 26 | 284 |
| Firm Revenue | 172 | 50 | 28 | 5 | 256 |
| Consumer Welfare | 121 | 68 | 45 | 12 | 246 |
| Task Completion Time | 183 | 33 | 10 | 13 | 240 |
| Inequality Measures | 45 | 126 | 50 | 6 | 227 |
| Worker Satisfaction | 95 | 74 | 23 | 12 | 204 |
| Error Rate | 77 | 98 | 11 | 4 | 190 |
| Regulatory Compliance | 84 | 73 | 17 | 7 | 181 |
| Automation Exposure | 61 | 61 | 27 | 14 | 166 |
| Training Effectiveness | 98 | 21 | 14 | 19 | 154 |
| Wages & Compensation | 78 | 37 | 25 | 6 | 146 |
| Developer Productivity | 105 | 18 | 14 | 6 | 144 |
| Team Performance | 87 | 17 | 28 | 10 | 143 |
| Job Displacement | 12 | 83 | 23 | 1 | 119 |
| Hiring & Recruitment | 53 | 8 | 8 | 3 | 72 |
| Social Protection | 39 | 17 | 8 | 2 | 66 |
| Creative Output | 32 | 20 | 8 | 3 | 64 |
| Skill Obsolescence | 5 | 50 | 6 | 1 | 62 |
| Labor Share of Income | 17 | 20 | 17 | — | 54 |
| Worker Turnover | 15 | 15 | — | 3 | 33 |
| Industry | — | — | — | 1 | 1 |
Regional agreements and plurilateral initiatives are being used as testing grounds for harmonizing standards and procedures prior to broader adoption.
Case studies and institutional observations of regional/plurilateral policy experiments (specific agreements referenced in examples but not exhaustively quantified).
AI enables new forms of digital cross-border trade such as AI-as-a-service and algorithmic intermediaries.
Conceptual mapping/theoretical analysis and descriptive case examples drawn from policy and market literature; case study details and counts not specified.
AI lowers traditional trade frictions (search, matching, logistics, customs).
Theoretical/mechanism analysis supported by illustrative case studies and secondary literature on digital platforms and AI applications; no quantitative sample size or econometric estimates reported.
Phased deployment and regulatory sandboxes can lower barriers for startups to pilot lower-risk applications, thereby shaping innovation trajectories.
Comparative policy analysis of sandboxing and phased deployment approaches in other jurisdictions; prescriptive inference without empirical testing in Vietnam.
Properly governed AI can yield large efficiency gains (reduced processing time and lower per-case costs), but those gains depend on redesigning legal processes to accommodate algorithmic workflows.
Analytic synthesis of administrative-process characteristics and AI capabilities; no primary quantitative evidence or measured effect sizes provided.
Establishing a graduated implementation model and clear regulatory pathways reduces regulatory uncertainty and makes public-sector AI procurement and private-market participation more predictable and attractive.
Normative recommendation informed by comparative institutional analysis and economic reasoning; not empirically tested in the paper.
A graduated implementation model—phased deployment, differentiated safeguards by risk, and mandatory human oversight for high-stakes decisions—can balance innovation with rule-of-law protections.
Normative framework development combining doctrinal findings and comparative lessons; prescriptive recommendation rather than empirical validation.
Comparative analysis of international frameworks reveals a range of institutional responses and regulatory instruments that Vietnam could adapt.
Comparative institutional analysis synthesizing governance approaches from liberal and civil-law jurisdictions (review of secondary sources and policy frameworks).
AI can substantially modernize administrative decision-making in civil-law systems (speed, consistency, scalability).
Qualitative doctrinal and comparative institutional analysis using Vietnam as a focused case study; no primary quantitative field data or sample size.
Literary narrative probes can serve as anticipatory evaluation instruments: they reveal subtler failures in more capable systems and their sophistication appears to scale with system capability rather than being circumvented by it.
Synthesis of empirical findings (increased discrimination in higher-capability systems, reproducible reflexive failure modes) and interpretive argument in Discussion.
The probe's discriminating power scales with system capability — it becomes more discriminating as models get stronger.
Observed increased discrimination in stronger models using a 'ceiling discrimination' probe and independent judges (Gemini Pro, Copilot Pro); comparisons across 13 systems and ceiling runs indicate the instrument revealed subtler failures in higher-capability systems.
Adoption of AI feedback could lower marginal costs of delivering high-quality feedback and change fixed vs. variable cost structures for instruction delivery.
Economic implication discussed by workshop participants (50 scholars) as a theoretical possibility; no quantitative cost estimates in the report.
Generative AI can enable new feedback modalities (text, hints, worked examples, formative prompts) adaptable to content and learner needs.
Thematic conclusions from the interdisciplinary meeting of 50 scholars, describing possible modality generation capabilities of current generative models; no empirical modality-comparison data provided.
Immediate AI-generated feedback may sustain learner momentum and improve formative assessment cycles (timeliness & engagement).
Expert-opinion synthesis from structured workshop (50 scholars) identifying timely feedback as a potential pedagogical benefit; no empirical trials reported.
Large language and generative models can tailor explanations, scaffolding, and practice to learners' current states and preferences (personalization).
Workshop expert consensus and thematic synthesis from 50 interdisciplinary scholars; illustrative examples discussed rather than empirical evaluation.
Generative AI can produce real-time, individualized feedback at scale, potentially reducing per-student feedback costs and increasing feedback frequency.
Synthesis of expert perspectives from an interdisciplinary workshop of 50 scholars (educational psychology, computer science, learning sciences); qualitative small-group activities and thematic extraction. No primary experimental or quantitative cost data presented.
SERF (Structured Error Recovery Framework) defines structured, machine-readable failure semantics to enable deterministic agent self-correction and automated recovery strategies.
Design and formal specification of SERF in the paper; formalized as a testable hypothesis with reproducible experimental methodology.
ATBA (Adaptive Timeout Budget Allocation) frames sequential tool invocation as a budget-allocation problem over heterogeneous latency distributions to improve end-to-end latency and reliability.
Algorithmic formulation and formalization provided in the paper; ATBA is presented as a testable hypothesis with reproducible benchmarks and latency/error models.
The MCP (Model Context Protocol) is widely adopted: >10,000 active MCP servers and 97 million monthly SDK downloads as of early 2026.
Reported protocol-adoption metrics in the paper (protocol adoption context); presumably aggregated server and SDK-download statistics (time-stamped to early 2026).
Agents learn from one another without curricula (agent-to-agent learning occurs organically in the ecosystem).
Naturalistic daily observations across platforms noting peer-to-peer agent interactions and apparent transfer of behaviors/knowledge; no controlled tests of learning or counterfactuals.
Agents form idea cascades and quality hierarchies without any centrally designed curriculum or intervention (emergent peer learning and spontaneous knowledge diffusion).
Observed interaction patterns across platforms showing cascades, hierarchies, and diffusion among agents in the qualitative dataset; documentation is comparative and observational rather than experimental.
A rapidly growing ecosystem of autonomous AI agents is producing organic, multi-agent learning dynamics that go beyond dyadic human–AI interactions.
Naturalistic, qualitative daily observations over one month across multiple agent platforms (reported platforms: Moltbook, The Colony, 4claw); coverage reported of >167,000 agents interacting as peers; comparative observational documentation rather than controlled experimentation.
Historical institutional publication records encode an extractable evaluative signal ("taste") that can be learned by models and used for scalable triage, screening, and curation of submissions.
Empirical results showing improved predictive accuracy after fine-tuning on accept/reject records, plus demonstration of transfer tasks and a cross-field (economics) result; implications for applications (triage, screening) are drawn from these empirical findings rather than directly deployed field experiments.
Models show well-calibrated confidence: their highest-confidence predictions are 100% accurate.
Calibration analysis of fine-tuned models comparing predicted-confidence levels to actual accuracy; reported that examples the model assigned its highest confidence to were 100% accurate. (Number of highest-confidence examples and calibration buckets not reported in the provided text.)
The learned evaluative signal transfers to untrained tasks such as pairwise comparisons and one-sentence summaries.
Fine-tuned models were evaluated on related, untrained evaluative tasks (pairwise comparisons of pitches and one-sentence summary evaluations) and showed positive transfer performance relative to baselines. (Specific metrics, effect sizes, and sample sizes for these transfer tasks are not provided in the supplied text.)
There is an economic rationale for disclosure mandates, certification of model properties (e.g., hallucination rates), and liability rules to internalize externalities from conversational AI.
Policy recommendation based on economic analysis of information asymmetries and externalities; no empirical testing of these policies in this paper.
Natural conversational interfaces lower search and transaction costs, increasing demand for AI services and expanding markets.
Economic reasoning and literature synthesis; the paper frames this as an implication rather than presenting empirical demand measurements.
Design interventions alone are necessary but not sufficient; institutional measures (standards, certification, liability rules) are also important to address harms and market failures.
Economic and policy analysis within the paper arguing for combined design and institutional responses; no empirical evidence demonstrating the comparative effectiveness of these measures.
Controls for personalization, data retention, opt-out, and escalation to human assistance are important interface affordances to mitigate risks in conversational AI.
Design heuristics and normative arguments from the paper and related literature; no empirical evaluation of these controls provided.
Real-time uncertainty/credibility signals and easy access to provenance (citations) should be provided to users to improve trust calibration.
Design recommendation grounded in literature review and suggested best practices; the paper recommends A/B tests and lab/field experiments as future work rather than reporting results.
Ethical front-end design—explicit disclosure of AI identity, capability limits, uncertainty cues, provenance, user controls, and escalation paths—can reduce harms and important market failures in AI-enabled interactions.
Normative and design-oriented recommendation supported by design heuristics and prior literature; no empirical trials reported showing quantified harm reduction.
Natural conversational style lowers friction and raises engagement and productivity.
Argument derived from literature synthesis and comparative analysis of conversational norms vs. human dialogue; no original empirical measurements reported in the paper.
SlideFormer generalizes beyond a single GPU vendor (the design achieves high utilization on both NVIDIA and AMD GPUs).
Reported experiments and utilization measurements on both NVIDIA (RTX 4090) and AMD GPUs showing sustained >95% peak performance, implying cross-vendor applicability. The summary does not specify which AMD models or the breadth of tested kernels.
Custom Triton kernels and advanced I/O integration remove key bottlenecks in single-GPU fine-tuning pipelines and contribute to the observed throughput gains.
Paper reports the use of custom Triton kernels for performance-critical primitives and improved I/O integration; throughput gains (1.40×–6.27×) are attributed in part to these optimizations. The summary does not isolate ablation results quantifying each optimization's contribution.
Heterogeneous memory management (multi-tier placement across GPU, CPU, and storage) materially reduces peak on-device memory requirements.
Authors describe an efficient memory layout and placement strategy across GPU, host RAM, and storage tiers and report lowered peak device memory use (≈2× reduction). The summary does not include low-level placement parameters or traces.
SlideFormer sustains >95% peak performance (high utilization) on both NVIDIA and AMD GPUs.
Reported sustained peak utilization measurements on experiments run on NVIDIA (e.g., RTX 4090) and AMD GPUs; the summary states >95% peak performance but does not give per-workload/utilization measurement methodology.
SlideFormer supports up to 8× larger batch sizes and up to 6× larger models on the same GPU relative to prior single-GPU baselines.
Reported comparisons to prior single-GPU baselines measuring achievable batch size and model-size capacity on the same GPU; exact baselines, workloads, and experimental configurations are not detailed in the summary.
SlideFormer reduces peak CPU and GPU memory usage by approximately 2× (roughly halving memory requirements).
Authors report peak memory measurements showing about a 2× reduction in both GPU and CPU memory compared to baselines; memory accounting method and baselines are not fully specified in the summary.
SlideFormer achieves 1.40×–6.27× higher throughput versus baseline systems.
Quantitative evaluation comparing throughput (reported as tokens/sec or updates/sec) against state-of-the-art single-GPU and multi-GPU fine-tuning pipelines (baselines are unnamed in the summary). Measurements reported across single-GPU experiments (hardware includes RTX 4090 and AMD GPUs).
SlideFormer enables fine-tuning very large LLMs (reported up to 123B+ parameters) on a single GPU (e.g., RTX 4090).
Authors report experiments and capability claims for single-GPU setups including an NVIDIA RTX 4090; model size stated as 123B+ in the paper summary. Details on exact model family, sequence length, or batch size used for the 123B+ claim are not enumerated in the summary.
Combining negative constraints with sparse preference signals yields better tradeoffs (safety plus helpfulness) than preference-only training.
Conceptual claim supported by qualitative comparisons and references to hybrid approaches in the literature (some constitutional/hybrid methods); the paper advocates this as a practical strategy and cites limited empirical indications.
Training primarily on negative constraints can reduce sycophancy and produce more stable adherence to rules compared to preference-only training.
Paper combines theoretical reasoning with cited empirical instances (e.g., constraint-based or constitutional methods) that report improved harmlessness/constraint adherence. The claim is stated as both theoretical expectation and supported by selected empirical reports rather than a comprehensive controlled comparison.
Negative constraints (explicit prohibitions or dispreferred labels) are often discrete, finitely specifiable, and independently verifiable, enabling models to converge to stable boundaries via falsification-style learning.
Theoretical/epistemological argument drawing on Popperian falsification and the paper's constructed structural model contrasting constraint and preference spaces. Empirical support is indirectly cited via methods like Constitutional AI that operationalize rule-like constraints.
Negative-only feedback (training on dispreferred or negative samples) can match or exceed preference-based RLHF (e.g., PPO/RLHF) on downstream tasks such as mathematical reasoning and harmlessness benchmarks.
Synthesis of recent empirical methods cited in the paper (examples named: Negative Sample Reinforcement, Distributional Dispreference Optimization, Constitutional AI) reporting parity or improvements versus PPO/RLHF on tasks like math reasoning and harmlessness. The paper aggregates published results rather than presenting a single new large-scale controlled experiment; specific sample sizes and exact experimental protocols vary by cited work and are not uniformly reported in the paper.
The core findings (harm from ToM order mismatches and benefits from A-ToM) are robust to partners beyond LLM-driven agents.
Paper reports robustness checks testing generalization to non-LLM agent classes (details summarized in robustness section); comparisons use the same coordination metrics.
A-ToM recovers coordination performance by aligning its effective ToM depth with partners across a range of multiagent tasks.
Experimental results showing A-ToM achieves coordination levels closer to matched fixed-order pairings across the repeated matrix game, grid navigation tasks, and Overcooked when facing partners with different fixed ToM depths.
An adaptive ToM (A-ToM) agent that infers its partner's ToM order from prior interactions and conditions its predictions and actions on that estimate restores alignment and improves coordination.
Implemented A-ToM (estimation from interaction history + conditioning of partner-action predictions) and evaluated it against fixed-order agents in the four environments; reported improvements in coordination metrics when A-ToM paired with partners of varying ToM orders.
Security testing included prompt-injection/adversarial inputs to probe the security agent and layered defenses.
Paper reports conducting prompt-injection/adversarial tests as part of security evaluation; the summary does not include the number, nature, or success/failure rates of these tests.
Rubric-based, structured scoring promotes consistent, auditable judgments and reduces subjective assessor bias.
System implements rubric-based, multi-dimensional scoring and the paper asserts this improves consistency and auditability; no reported inter-rater reliability statistics or controlled comparisons to human/monolithic baselines are provided in the summary.
Isolating sensitive logic (scoring rubrics, adaptive difficulty rules) from free-text generation reduces the attack surface.
Design principle implemented in the architecture (separation of concerns between agents); claimed benefit in the paper. Empirical validation details (quantitative reduction in successful attacks) are not provided in the summary.