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 |
Neuromodulatory systems and meta-decision circuits in animals provide analogies for implementing meta-control (M) in artificial systems.
Neuroscience analogy cited to motivate architectural choices; not empirically instantiated in the paper.
Developmental trajectories can scaffold gradual competence (from observation to exploratory action) and should be reflected in training curricula.
Argument from developmental biology and learning theory; proposed as a design principle rather than empirically tested here.
Evolution supplies inductive biases and slow structural priors that can be leveraged in artificial learners.
Biological analogy and theoretical suggestion; no empirical experiments presented to quantify effect in AI systems.
The taxonomy and measurement approach provide operational metrics to quantify empathic communication for economic analyses (productivity, customer satisfaction, retention).
Authors propose that their data-driven taxonomy and automated/coding measures can be used as metrics; the paper demonstrates derivation and use in trial outcomes but does not present direct economic outcome measurements.
LLM-generated responses frequently score as more empathic than human-written responses in blinded evaluations.
Blinded evaluations comparing LLM-generated replies to human-written replies using recipient/judge ratings of perceived empathy (reported in blinded tests described in paper). Exact blinded-test sample sizes not specified in the summary but derived from the study's evaluation procedures.
LLMs are more likely to complement human tacit skills than to replace explicit rule‑following jobs; value accrues to workers and firms that integrate model outputs with human judgment and tacit expertise.
Labor‑economics style argument and theoretical reasoning; no empirical labor market analysis provided.
Commoditization via rule extraction is limited; firms that can harness and deploy tacit LLM capabilities will retain economic rents.
Theoretical economic argument based on non‑rule‑encodability; no empirical firm‑level data included.
The highest‑value attributes of LLMs may be inherently non‑decomposable into simple, auditable rules, which increases the value of proprietary, black‑box models and strengthens economies of scale and scope for large model providers.
Economic reasoning and theoretical implications drawn from the central thesis; no empirical market analyses provided.
Some LLM capabilities are tacit, practice‑derived, or 'insight'‑like, akin to the Chinese concept of Wu (sudden insight through practiced skill).
Philosophical framing and analogy to the concept of tacit knowledge (Wu); argumentative rather than empirical support.
The economically valuable capabilities of large language models are precisely those that cannot be fully encoded as a complete, human‑readable set of discrete rules.
Formal, conceptual argument (proof by contradiction) plus qualitative historical case analysis comparing expert systems and LLMs; no new empirical datasets or experiments reported.
The paper reports quantitative improvements (registration accuracy and reduced inter-object penetration) and demonstrates generalization gains of the multi-object approach on multiple datasets.
Cross-dataset experiments and quantitative metrics reported in the paper comparing MOD to baselines, showing improved registration and reduced penetration as well as transfer/generalization performance across datasets.
The dataset and MOD produce far less inter-object penetration than prior datasets and single-object methods, with consistent improvements demonstrated across three benchmarks.
Reported empirical comparisons in the paper measuring inter-object penetration and showing substantially lower penetration for the proposed dataset+method relative to alternatives; experiments run on three benchmarks as stated in the paper.
MOD consistently improves multi-object reconstruction quality across three datasets/benchmarks compared to state-of-the-art baselines.
Experimental results presented across three datasets/benchmarks showing consistent improvements of MOD over SOTA baselines on multi-object reconstruction metrics. (The summary does not list the names of the three benchmarks or the per-benchmark metrics/numbers.)
The MessyKitchens dataset and MOD together yield materially better registration accuracy than prior datasets and single-object methods.
Quantitative evaluations in paper report improved registration accuracy when using MessyKitchens and/or MOD relative to prior datasets and methods; comparisons performed across benchmarks. (Exact numeric gains and sample sizes not included in the provided summary.)
MOD (built on SAM 3D) produces fewer inter-object penetrations and more physically plausible object configurations than single-object monocular methods.
Empirical evaluation reported in paper comparing MOD against single-object baselines (including SAM 3D) on inter-object penetration metrics; results show reductions in measured penetrations. (Specific numeric reductions and dataset sizes are not provided in the supplied summary.)
Distilling corrected decision trajectories into the model via supervised fine-tuning produces better recovery behavior than relying solely on reward signals or final-outcome optimization.
Comparative training setup where LEAFE uses supervised fine-tuning on corrected trajectories and is empirically compared to outcome-driven methods (e.g., GRPO) that optimize rewards; improved Pass@k reported.
LEAFE's gains occur across diverse interactive coding and agentic tasks with limited interaction budget.
Reported evaluation across a suite of long-horizon tasks (examples include multi-step coding problems and agentic tasks with rich feedback channels) with consistent improvements claimed.
LEAFE uses the same environmental interactions more effectively, improving sample efficiency under fixed interaction budgets.
Experimental regime with fixed interaction budgets demonstrating higher Pass@k for LEAFE relative to baselines given the same number of environment interactions; paper argues LEAFE converts richer feedback into targeted training signals rather than only final rewards.
LEAFE converts rich environment feedback into actionable corrective supervision rather than optimizing only final success signals, which drives performance gains.
Algorithmic description: LEAFE summarizes error messages/intermediate observations into experience items, backtracks to causal decision points, explores corrective branches, and distills corrected trajectories via supervised fine-tuning. Empirical comparisons show improved Pass@k relative to reward-only/outcome-driven baselines.
Open dataset and code improve reproducibility and lower barriers for follow-up work on applied LLM tools and economic impact studies.
Release of SlideRL dataset (288 rollouts) and code repository; general statement about reproducibility benefits.
Parameter-efficient RL fine-tuning (0.5% of params) can yield large quality gains, implying a potentially high ROI for targeted fine-tuning versus full-model scaling.
Observed empirical gain of +33.1% for the tuned 7B over its untuned base and the 91.2% relative performance vs Claude Opus 4.6; implication drawn about cost-effectiveness of tuning few parameters rather than scaling model size.
The inverse-specification reward—where an LLM attempts to recover the original brief from generated slides—provides a holistic fidelity signal.
Reward design: inverse-specification component implemented and used as part of composite reward; claimed to measure fidelity via recovery accuracy.
Performance on this agentic slide-generation task is driven more by instruction adherence and tool-use compliance than by raw model parameter count.
Cross-model comparison across six models on the 48-task benchmark, with analyses showing instruction adherence and tool-use compliance better predict agent performance than parameter count.
Adoption will shift labor demand toward expertise in deterministic capture/replay tooling, trace analytics, and integration automation.
Economic/organizational implication discussed in the summary; no employment-data analysis provided—stated as an expected change in skill demand.
The approach improves utilization and ROI of expensive emulation/simulation resources by enabling reuse of deterministic traces across platforms.
Implication drawn from being able to replay identical traces on both simulator and emulator; no direct financial ROI calculation or utilization metrics provided in the summary.
Using replay-driven validation markedly shortens integration and debug cycles for the demonstrated chiplet subsystem, enabling end-to-end system boot and workload execution within a single quarter.
Reported outcome for the ODIN SoC building block: authors state they were able to reach full system boot and run workloads within one quarter of integration using the methodology. (Single-case timeline reported; no control/comparison group or statistical analysis provided.)
Replay-driven validation made previously hard-to-reproduce interactions and bugs deterministic and repeatable at system level, enabling more focused and efficient debug.
Authors report that deterministic capture/replay converted non-deterministic protocol interactions and transient bugs into repeatable traces that could be inspected and debugged; examples include complex GPU workloads and protocol sequences reproduced end-to-end. (Qualitative/process-level evidence from the demonstrator; no numerical bug-count reduction provided.)
A replay-driven validation methodology using deterministic waveform capture and replay from a single design database enables reliable, repeatable system-level reproduction of complex GPU workloads and protocol sequences for tightly coupled CPU–GPU chiplet subsystems.
Applied to a demonstrator SoC building block (ODIN chiplet architecture) integrating a CPU subsystem, multiple Intel Xe GPU cores, and a configurable NoC; deterministic waveform capture during execution and deterministic replay of those waveforms across targets was performed; same design database used to manage captures, traces, and replay sessions. (No large-sample statistical evaluation reported; demonstration limited to the described system.)
Overall conclusion: forecast-then-execute (anticipatory trajectory reasoning) is an effective principle for building multimodal agents capable of reasoning, planning, and acting in complex environments.
Paper's Conclusion in the provided summary asserts this, based on the reported experimental comparisons and the two-stage TraceR1 framework.
The paper reports improvements in planning stability (consistency of multi-step plans), execution robustness (success under environment/tool variability), and generalization (out-of-distribution tasks and unseen tool/environment states).
Reported outcomes in the summary explicitly list these three improvement categories; the specific metrics and magnitudes are not provided in the summary.
Compared to reactive agents that optimize actions stepwise without trajectory anticipation, TraceR1 yields better multi-step planning and execution.
Baselines & comparisons described in the summary include reactive agents; the paper reports improvements of TraceR1 relative to these baselines across the benchmarks (no numeric values in the provided text).
Explicit anticipatory (trajectory-level) reasoning is a crucial design principle for reliable multi-step task performance in complex real-world environments.
Paper reports comparisons between anticipatory (trajectory-forecasting) agents and reactive / single-stage baselines, concluding the anticipatory design yields better multi-step reliability; exact experimental details and statistics not included in the provided summary.
TraceR1 materially improves planning coherence, execution robustness, and generalization in multimodal, tool-using agents versus reactive or single-stage baselines.
Reported evaluation across seven benchmarks (online and offline computer-use, multimodal tool-use reasoning) comparing TraceR1 to reactive agents and single-stage RL baselines; summary states 'substantial gains' though no numerical results are provided in the provided text.
The proposed algorithm's performance is robust to heterogeneous populations in the synthetic experiments (i.e., it continues to find core alternatives under varying degrees of population heterogeneity).
Empirical robustness checks reported in the experiments where population heterogeneity is varied and performance (core-attainment frequency) is evaluated.
The authors compare their sampling algorithm against classical social-choice rules and LLM-based heuristics and report superior core-attainment frequency for their method.
Experimental comparisons described in the paper between the proposed algorithm and baseline methods (classical social-choice rules, LLM-based heuristics) on the synthetic dataset; results summarized in the experiments section.
On a synthetic text-preference dataset, the proposed algorithm reliably finds alternatives that lie in the proportional veto core.
Empirical experiments reported in the paper using a synthetic dataset of text preferences; evaluation metric reported as frequency (proportion) of runs where the returned alternative is in the proportional veto core.
Temporal grounding (restricting models to contemporaneous information) should be adopted as a methodological best practice in economic research using LLMs to avoid leakage and produce more realistic assessments of model forecasting ability.
Study methodology and rationale emphasize temporal grounding; authors recommend it as best practice based on the observed benefits in reducing retrospective contamination.
Because the conflict unfolded after the training cutoffs of contemporary frontier LLMs, the dataset and analyses provide an archival, hindsight-free benchmark for studying model reasoning.
Case selection rationale: the 2026 Middle East conflict was deliberately chosen because it occurred after the training cutoffs of the evaluated frontier models; dataset preserves contemporaneous queries and model outputs.
Frontier large language models (LLMs) can reason about an unfolding geopolitical crisis using only contemporaneous public information, often demonstrating strategic realism (inferring underlying structural incentives beyond surface rhetoric).
Evaluation across 11 temporally defined nodes during the early 2026 Middle East conflict using 42 node-specific verifiable questions and 5 exploratory prompts; results assessed via verifiability checks and qualitative coding for strategic reasoning of outputs from contemporary frontier LLMs constrained to contemporaneous information.
BATQuant establishes new state-of-the-art results across multimodal benchmarks for MXFP4-aware PTQ under aggressive quantization.
Comparative benchmark results reported in the paper showing BATQuant outperforming prior PTQ methods on the described multimodal benchmarks (specific benchmark names and quantitative margins not provided in the summary).
Ablation analyses show that each BATQuant component (block-wise transforms, orthogonality relaxation, GPK decomposition, block-wise clipping) contributes to robustness and efficiency.
Reported ablation studies isolating components and measuring their individual impact on performance and overhead in the paper's experiments (exact effect sizes and per-component numbers not given in the summary).
Block-wise learnable clipping suppresses residual outliers locally and contributes to robustness under aggressive MXFP4 quantization.
Method description and ablation experiments in the paper showing incremental improvement when adding block-wise learnable clipping layers versus not using them; improvements measured on benchmark metrics post-quantization.
Global and Private Kronecker (GPK) decomposition compresses transform parameters, keeping storage and runtime overhead low compared to dense per-block transforms.
Algorithmic contribution described in the paper with reported comparisons (storage/runtime overhead) versus dense per-block transform parameterizations; supported by experimental/implementation measurements (specific memory/runtime numbers not provided in the summary).
Relaxing orthogonality constraints on transforms (i.e., using non-strictly-orthogonal transforms) improves distribution shaping and better fits activations to the limited MXFP quantization range.
Design rationale and ablation studies reported in the paper showing that removing strict orthogonality yields better quantization fit and improved task metrics versus enforced orthogonal transforms.
Aligning transforms to MXFP block granularity using block-wise affine transformations prevents cross-block outlier propagation and avoids the severe collapse seen with rotation-based integer quantization techniques.
Methodological design plus ablation/empirical results in the paper showing improved activation statistics and preserved model accuracy when using block-wise affine transforms aligned to MXFP blocks versus global rotations.
Standardized runtime governance frameworks could lower per-deployment compliance engineering costs and increase diffusion of agentic systems.
Theoretical argument that standardization reduces transaction/engineering costs; suggested market dynamics; no empirical implementation evidence.
A market will develop for third-party governance tools, auditors, and insurers providing policy evaluators, risk calibration, and certification services.
Economic argument and analogy to existing markets (governance-as-a-service, insurance); no empirical evidence presented.
Benchmarking time-sensitivity (via V-DyKnow) can inform procurement decisions: buyers should assess models on their ability to handle temporally sensitive information, not just static benchmarks.
Paper's recommendations and implications section arguing for procurement practices informed by V-DyKnow evaluations.
The authors provide an operational inventory and conversation-analysis tool (the 28-code instrument) that can be reused for monitoring and mitigation by researchers, firms, and regulators.
Paper includes the codebook and describes its application as a re-usable monitoring/analysis instrument; proposed adoption discussed in implications.
This is the first empirical, message-level study of verified chatbot-related psychological-harm cases (as opposed to speculative discussion).
Authors' positioning in paper; claim of novelty based on review of prior literature and their message-level, verified-case approach.