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 |
Shifting to DSS changes the cost structure of AI: it lowers recurring OPEX per user by reducing inference energy and enabling local/device processing instead of centralized, inference-heavy cloud services.
Economic reasoning and proposed modeling approaches (capex/opex comparisons) described conceptually; no empirical economic model outputs or market data are included.
DSS societies can achieve much lower inference energy per task and enable easier on-device/edge deployment compared to monolithic LLM deployments.
Argument that smaller, domain-focused models require fewer compute resources and thus lower energy and are better suited to edge hardware; empirical measurements to support this claim are proposed but not supplied.
Architecturally, replacing single giant generalists with 'societies' of small, specialized DSS models routed by orchestration agents yields operational benefits (routing to experts, modular upgrades, specialization).
Conceptual architectural proposal describing specialized back-ends and orchestration/routing agents; the paper outlines recommended experiments but reports no empirical orchestration benchmarks.
A more sustainable and effective trajectory is to build domain-specific superintelligences (DSS) grounded in explicit symbolic abstractions (knowledge graphs, ontologies, formal logic) and trained via synthetic curricula so compact models can learn robust, domain-level reasoning.
Prescriptive proposal based on theoretical arguments about the benefits of symbolic abstractions, compact model training, and synthetic curricula; no experimental validation or empirical comparison is provided in the paper.
Standardizing these infra-level primitives could lower integration costs across ecosystems and accelerate enterprise adoption of agent-hosted services.
Policy/economic argument presented in the paper's implications and research directions; no empirical standardization impact study provided.
Missing infraprotocol primitives in MCP create opportunities for platform differentiation—providers implementing CABP/ATBA/SERF-like extensions can capture value by offering more production-ready agent tooling.
Strategic/economic reasoning stated in the implications section; not supported by empirical market-share data in the summary.
A concrete empirical test recommended by the paper is to run controlled comparisons of distribution-shift generalization between negative-only, preference-only, and hybrid-trained models across safety and usefulness metrics.
Methodological recommendation given in the paper; it is not an empirical result but an explicitly proposed verifiable experiment for future work.
Regulators could feasibly focus on certifying constraint datasets and testing model adherence to explicit prohibitions, since constraint compliance is empirically testable and verifiable.
Policy recommendation derived from the paper's epistemic argument about constraints being verifiable; presented as a plausible regulatory strategy rather than one already validated by policy experiments.
There is a commercial opportunity for startups and vendors to specialize in 'constraint datasets' and constitutional-rule libraries as tradable assets.
Market/economic inference made from the technical claim that constraints are verifiable and reusable; no empirical industry survey data provided—this is a forward-looking implication.
If negative/safety-focused signals are more sample- and compute-efficient for certain alignment goals, firms may reallocate labeling budgets away from costly preference elicitation toward collecting high-quality negative examples and rule sets.
Economic implication extrapolated from the paper's sample-efficiency claim; the paper reasons from technical sample-efficiency arguments and cited empirical parity but does not present market-level empirical data.
Improved alignment can reduce harms from misinterpretation (incorrect decisions, misinformation), lowering downstream liability and reputational risk for vendors and customers.
Paper's safety and externalities discussion argues this as a likely consequence; the claim is theoretical and not supported by empirical incident data in the paper.
Providers may charge a premium for alignment-enabled API tiers or incorporate C.A.P. into enterprise plans because of additional compute per interaction, affecting pricing and unit economics.
Paper's pricing and costs discussion predicts potential monetization strategies and pricing experiments (A/B pricing, willingness-to-pay studies) but does not report market data.
C.A.P. has potential economic effects: it can reduce time lost to misinterpretation, thereby increasing effective throughput and productivity, though net gains depend on trade-offs with pre-processing overhead.
Economic implications section provides conceptual cost–benefit arguments and recommends pilot measurements (time saved, reduced human review cost) but provides no empirical economic measurement.
C.A.P. shifts interactions from one-way command-execution to two-way, partnership-style collaboration, increasing perceived partnerliness.
Theoretical argument drawing on cognitive science and Common Ground theory and proposed human-evaluation measures (satisfaction, perceived collaboration); no empirical human-subject results reported.
C.A.P. improves long-term and dynamic dialogue alignment and reduces off-topic or mechanically incorrect responses.
Main argument of the paper based on the combined functions (expansion, weighted retrieval, alignment verification, clarification); the paper provides conceptual/theoretical justification but does not report large-scale empirical results.
Public archives of prompts and commits accelerate diffusion by lowering search/learning costs and enabling replication, thereby increasing adoption speed and lowering entry barriers.
Paper's asserted implication based on the existence of public artifacts and general reasoning about knowledge diffusion; this is an interpretive claim rather than an experimentally validated finding (argumentative, extrapolative).
Developing economic metrics linked to architecture (interoperability indices, expected upgrade cost, observability coverage, market concentration measures, systemic‑risk indicators) is recommended to guide policy and investment.
Policy recommendation grounded in the paper's normative analysis; no pilot metric development or empirical validation presented.
The benchmark provides a testbed useful for studying strategic behavior, coordination failures, and market-like interactions among agents, which can inform economic research and policy.
Paper claims the benchmark's multi-agent, strategic tasks can be used as experimental environments for economic and policy research; this is a normative claim supported by the benchmark's design rather than by empirical studies in the paper.
Open-source orchestration lowers entry barriers, broadening participation and potentially compressing rents that would otherwise accrue to well-resourced incumbents.
Paper's discussion section argues that releasing orchestration and evaluation tools publicly reduces the technical overhead for entrants; this is a theoretical/observational claim rather than empirically measured in the paper.
The clear performance gaps indicate high returns to specialized efforts (RL, domain-specific engineering) relative to generalist LLM-only approaches, shaping where teams invest labor and compute.
Paper links benchmarking results (performance gaps between baselines and humans) to economic implications, arguing specialization yields higher returns; this is an interpretive claim based on reported performance differentials.
Benchmarks like PokeAgent will reallocate researcher and industry attention toward multi-agent, partial-observability, and long-horizon planning problems—likely increasing funding and compute investment in RL and hybrid LLM+RL methods.
Paper offers an economic/implication analysis arguing that introducing such a benchmark changes incentives and investment patterns; this is a reasoned projection rather than an empirical observation.
Public investment in open environments, robotics testbeds, and safety research can reduce concentration risks and externalities and democratize access to embodied AI research.
Policy recommendation based on anticipated strategic importance of shared infrastructure; not empirically validated here.
Value in the AI ecosystem may shift from passive text/image corpora toward rich interaction datasets and simulated/real environments; ownership and control of simulation platforms and testbeds could become strategically important assets.
Economic and strategic inference from the proposed technical emphasis on embodied/interaction learning; no supporting market data in the paper.
Increased sample efficiency and transfer will reduce compute and data costs, lowering barriers to entry for firms and broadening feasible AI applications.
Economic argument connecting technical metrics to cost and market effects; not empirically demonstrated in the paper.
More autonomous learners that can self-experiment and learn from observation will lower deployment costs for adaptable agents and accelerate automation across more occupations, especially embodied and social tasks.
Economic reasoning and projection based on expected technical improvements; speculative without empirical economic analysis in the paper.
Cross-cutting elements (hierarchical organization, curriculum/bootstrapping, intrinsic motivation, uncertainty estimation, memory consolidation, neuromodulatory analogs) are important for improving learning in the proposed architecture.
Conceptual recommendation based on known mechanisms from neuroscience and machine learning literature; not validated in the paper.
System M (meta-control) should generate internal signals that decide when to prioritize A vs B, allocate attention, consolidate memory, and trade off uncertainty, novelty, expected information value, and effort costs.
Design proposal motivated by biological meta-control and decision theories; no empirical tests presented.
System B (action-driven learning) should learn through intervention, consequences, and trial-and-error, using active exploration, reinforcement learning, and hierarchical/skill learning.
Architectural proposal aligning with RL and hierarchical learning literature; theoretical description without experimental evidence.
System A (observation-driven learning) should build models of others, social contingencies, and passive affordances through imitation, self-supervised representation learning, and inverse RL.
Architectural specification and mapping to existing algorithms (imitation, SSL, inverse RL); no empirical validation provided.
Integrating observation-driven and action-driven learning with meta-control and evolutionary/developmental priors should improve sample efficiency, robustness, transfer, and lifelong adaptation.
Conceptual argument and proposed integration of methods; suggested but untested experimentally in the paper.
A biologically inspired three-part architecture (System A: observation-driven learning; System B: action-driven learning; System M: internally generated meta-control) can address these limitations.
Theoretical proposal and analogy to biological systems; no empirical validation reported in the paper.
Embedding LLM coaching tools in platforms (employee onboarding, customer support, peer-support communities) could raise overall conversational quality by improving expressive outcomes rather than only informational accuracy.
Authors' implication drawn from trial results showing improved alignment to empathic norms after personalized coaching; no field deployment evidence provided in the paper.
LLM-driven personalized coaching can cheaply scale soft-skill training (empathy expression) that would otherwise require costly human trainers, suggesting a high-return application of AI in workforce development.
Implication drawn from observed efficacy of brief automated coaching in the trial and the scalable nature of LLM deployment; no direct economic field trial provided in the paper.
Barriers to entry may be larger for tacit‑capability‑driven systems than for rule‑based systems, potentially increasing market concentration.
Economic argument linking tacit capabilities to requirements for large data, compute, and specialized training dynamics; speculative and not empirically tested in the paper.
HindSight-style retrospective matching could underpin markets or contingent contracts for ideas by providing an objective payoff rule based on later publications and citations.
Paper's implications section proposing that retrospective matching can be used as an objective payoff rule for markets; this is a proposed application rather than an empirical finding.
Physically-plausible reconstructions reduce unsafe behaviors in deployed agents (e.g., collisions) and lower simulation-to-real failure modes.
Argument in paper tying reduced inter-object penetration and realistic contacts to fewer failures in simulation-to-real pipelines and safer agent behavior; not an empirical claim directly validated in real-world deployments within the provided summary.
Open release of a high-quality 3D dataset and pre-trained models will lower entry barriers and intensify competition in robotics, AR/VR, and 3D content markets.
Paper discussion posits that public benchmarks and models reduce dataset/compute barriers and enable broader research and product development. This is a policy/economic implication stated by the authors, not tested empirically in the paper.
Better monocular multi-object 3D reconstruction can lower perception costs for robots and embodied agents (fewer sensors, less calibration) and accelerate deployment in logistics, household service robots, inspection, and manipulation tasks.
Discussion/implications section in paper arguing that improved single-image multi-object reconstruction reduces reliance on extra sensors and calibration, with downstream benefits for robotic deployment. This is presented as implication/argument rather than empirical evidence in the paper summary.
By extracting more training value from the same environment interactions, LEAFE reduces marginal data/interaction costs and shifts the cost curve of deploying agentic systems (improves returns-to-sample-effort).
Economic implication argued in the paper based on reported increased sample efficiency under fixed budgets; no formal economic modeling provided—argumentative inference from performance gains per interaction.
The methodology enables modular chiplet economics by removing a key validation bottleneck, which could support modular upgrade paths and lower manufacturing cost via mixed-node IP blocks.
Authors propose this as an implication of improved integration and repeatability; argumentative claim without accompanying manufacturing-cost or economic-case studies in the summary.
Replay-driven validation can reduce engineering labor hours spent chasing non-deterministic bugs, lowering validation cost per project and decreasing risk of late-stage silicon respins.
Economic implication presented by authors: deterministic, repeatable debugging is argued to reduce manual effort and risk; no empirical labor-hour or cost-savings data provided in the demonstration.
Replay-driven validation is positioned as a scalable pre-silicon validation strategy for future chiplet-based heterogeneous systems.
Authors articulate scalability as a key positioning argument and present the methodology applied to a non-trivial CPU+multiple-GPU-core+NoC demonstrator; however, no large-scale or multi-project scalability study or quantitative scaling metrics are provided.
Surrogate-assisted inverse design reduces the marginal cost and time of exploring high-dimensional, discrete hardware design spaces by replacing costly EM simulations with fast ML inference, increasing R&D productivity and shortening design cycles.
Argument provided in implications: surrogate replaces EM simulations enabling faster iteration; no quantitative cost or time savings, or economic measurements, are presented in the summary.
A successful, stable parallel Newton software stack could spawn middleware and tooling ecosystems (sequence-parallel training/inference libraries), changing how cloud compute is sold and optimized for long-sequence workloads.
Forward-looking implication argued in the thesis based on observed algorithmic improvements and typical software-market dynamics; no empirical market studies provided.
Higher utilization efficiency and lower memory footprints from the proposed methods can reduce energy per computation on sequence tasks, moderating environmental impacts of large-scale sequence modeling.
Argument based on measured reductions in runtime and memory in experimental results combined with standard relations between runtime/memory and energy; no direct energy-measurement experiments reported.
If effective, these methods raise the value of parallel hardware (GPUs/TPUs) for sequence-heavy tasks and could increase demand for massive-parallel accelerators over specialized sequential hardware.
Economic and systems-level reasoning extrapolating from algorithmic speedups and memory reductions; no market-deployment experiments presented.
Enabling parallelization across sequence length can substantially increase GPU utilization and throughput for workloads previously dominated by sequential bottlenecks, reducing amortized compute cost per inference/training pass on long sequences.
Analytical argument based on observed runtime/parallelization improvements and the structure of GPU hardware; no large-scale economic deployment experiments reported in the thesis (argumentative/implicational evidence).
There is a market opportunity for scalable 'control-as-a-service' offerings and curated urban traffic datasets enabled by this data-driven control approach.
Authors' market and policy discussion extrapolating from technical results to business models and data infrastructure value; conceptual reasoning rather than empirical market analysis.
Reductions in travel time and CO2 emissions translate into measurable economic benefits (lower fuel consumption, productivity gains, reduced pollution-related health costs).
Economic implications discussed qualitatively in the paper as extrapolation from measured reductions in travel time and emissions; no direct empirical economic quantification within the traffic simulation experiments.
Benchmarks and standards are needed for evaluating high-frequency time series performance to guide procurement and contracting decisions.
Paper recommends establishing standards and benchmarking protocols specifically for high-frequency time series, motivated by observed TSFM brittleness on millisecond data. This is a policy/research recommendation rather than an empirical result.