Papers (1699)

Filter: Sort:

China’s green data‑centre pilots lower firms’ energy intensity by boosting ESG performance and green spending, with the biggest gains in resource‑based cities, competitive industries and mature firms led by managers combining green and digital vision. As successive pilot rounds roll out the aggregate payoff first rises and then fades, implying early concentrated benefits and later dilution.

medium quasi_experimental Full text openalex

How Does Urban Green Data Center Policy Empower Corporate Energy Utilization Efficiency? Evidence from China · Guoliang Wang, Haoxuan Sheng, Yue Lu · Mar 12, 2026

AI could sharply speed and standardize administrative decisions in Vietnam, but entrenched civil‑law protections around discretion, reason‑giving and review mean straightforward automation is legally fraught; a phased, risk‑based rollout with mandated oversight and procedural redesign is needed to unlock benefits without eroding rule‑of‑law safeguards.

low descriptive Full text openalex

ARTIFICIAL INTELLIGENCE AND ADMINISTRATIVE GOVERNANCE: A CRITICAL ANALYSIS OF TECHNOLOGICAL INTEGRATION IN VIETNAM'S LEGAL FRAMEWORK · Hoang Thanh Hanh, Vu Minh Chau, Tran Van Hoy · Mar 12, 2026

High-quality LLM suggestions markedly improve caseworkers’ SNAP accuracy—by about 27 percentage points—while incorrect suggestions substantially reduce accuracy; gains level off above roughly 80% chatbot accuracy, revealing an 'AI underreliance plateau' that complicates real-world deployment.

high rct Full text arxiv

LLMs in social services: How does chatbot accuracy affect human accuracy? · Jennah Gosciak, Eric Giannella, Zhaowen Guo, Michael Chen, Allison Koenecke · Mar 11, 2026

Making readers' trails visible — by logging navigation, highlights, links and notes — would expose previously invisible interpretive labor and create new signals for personalization and model training. But turning reading traces into a data asset risks privacy harms and market concentration unless accompanied by careful governance and empirical validation.

n/a descriptive Full text arxiv

Chasing RATs: Tracing Reading for and as Creative Activity · Sophia Liu, Shm Garanganao Almeda · Mar 11, 2026

Boson Sampling Born Machines can be trained on classical hardware yet retain quantum sampling hardness, enabling a 'train-classically, deploy-quantumly' model; modest architectural additions (ancillas and postprocessing) restore full expressivity while keeping training classically tractable.

n/a theoretical Full text arxiv

Universality of Classically Trainable, Quantum-Deployed Boson-Sampling Generative Models · Andrii Kurkin, Ulysse Chabaud, Zoltán Kolarovszki, Bence Bakó, Zoltán Zimborás, ... · Mar 11, 2026

Task-aware capability and risk signals constructed from human pairwise comparisons let users predict LLM performance more accurately and turn delegation into a visible, auditable choice; surfacing these primitives could reduce information asymmetry and enable risk-priced routing, auditing, and new supervisory roles.

medium descriptive Full text arxiv

Task-Aware Delegation Cues for LLM Agents · Xingrui Gu · Mar 11, 2026

How you apply PRF matters more than where it comes from: across 13 low-resource retrieval tasks, the feedback model drives retrieval gains more than feedback source; LLM-generated feedback is the cheapest high-return option unless a high-quality first-stage retriever supplies strong candidate documents, in which case corpus-derived feedback outperforms.

medium other Full text arxiv

A Systematic Study of Pseudo-Relevance Feedback with LLMs · Nour Jedidi, Jimmy Lin · Mar 11, 2026

Randomized 'human uplift' trials remain useful but are routinely compromised by frontier AI's nonstationarity, shifting baselines, diverse and evolving users, and spillovers; experts recommend adaptive designs, continuous post-deployment monitoring, mixed methods, and conservative interpretation rather than reliance on single trials for high‑stakes decisions.

medium descriptive Full text arxiv

RCTs & Human Uplift Studies: Methodological Challenges and Practical Solutions for Frontier AI Evaluation · Patricia Paskov, Kevin Wei, Shen Zhou Hong, Dan Bateyko, Xavier Roberts-Gaal, Ca... · Mar 11, 2026

AI tools automate routine software-engineering work—speeding delivery and improving defect management—yet benefits vary widely by team, domain and data maturity, and require new skills, governance and quality-assurance practices.

low descriptive Full text arxiv

Artificial Intelligence as a Catalyst for Innovation in Software Engineering · Carlos Alberto Fernández-y-Fernández, Jorge R. Aguilar-Cisneros · Mar 11, 2026

A unified Bayesian-surrogate framework cuts costly quantum-chemistry evaluations by about tenfold without losing accuracy, speeding up minima and saddle-point searches on potential-energy surfaces. The method—implemented in Rust and combining derivative-aware GPs, inverse-distance kernels, optimal-transport sampling, and trust-region controls—lowers compute and capital costs for computational discovery workflows.

medium other Full text arxiv

Bayesian Optimization with Gaussian Processes to Accelerate Stationary Point Searches · Rohit Goswami · Mar 11, 2026

A simulation-trained 'summary network' can learn finite-sample estimators and even mimic iterative solvers, offering fast amortized inference for parametric models; however, evidence comes from simulations only, so real-data validity and uncertainty quantification remain open challenges.

low other Full text arxiv

ForwardFlow: Simulation only statistical inference using deep learning · Stefan Böhringer · Mar 11, 2026

Evaluation metrics push image generators toward brighter, exaggerated colors; a new 1.3M‑image benchmark, a learned color‑fidelity metric, and a training‑free refinement method restore authentic color and let developers improve photorealism without costly retraining.

n/a descriptive Full text arxiv

Too Vivid to Be Real? Benchmarking and Calibrating Generative Color Fidelity · Zhengyao Fang, Zexi Jia, Yijia Zhong, Pengcheng Luo, Jinchao Zhang, Guangming Lu... · Mar 11, 2026

Feeding MCMC-drawn parameter posterior samples into deterministic neural surrogates reproduces model uncertainty at a fraction of the computational cost and without training on implausible parameterizations. The architecture-agnostic method includes theoretical bounds tying performance loss to distribution mismatch, giving practitioners a practical diagnostic for retraining or reweighting.

n/a other Full text arxiv

MCMC Informed Neural Emulators for Uncertainty Quantification in Dynamical Systems · Heikki Haario, Zhi-Song Liu, Martin Simon, Hendrik Weichel · Mar 11, 2026

A lightweight GNN trained via federated learning across orbital planes—coordinated through HAPS—improves beam-selection accuracy and stability in LEO satellite networks, cutting the need to centralize raw channel data. The approach promises lower operational overheads and easier on-board deployment, though real-world deployment and full cost savings remain to be demonstrated.

medium other Full text arxiv

Federated Learning-driven Beam Management in LEO 6G Non-Terrestrial Networks · Maria Lamprini Bartsioka, Ioannis A. Bartsiokas, Athanasios D. Panagopoulos, Dim... · Mar 11, 2026

Adaptive RL-driven contact control raises robotic vial-scraping performance by roughly 11% versus fixed-force policies after sim-to-real transfer; gains are shown on five material setups but rely on simplified material models and unspecified sensing details.

medium descriptive Full text arxiv

Learning Adaptive Force Control for Contact-Rich Sample Scraping with Heterogeneous Materials · Cenk Cetin, Shreyas Pouli, Gabriella Pizzuto · Mar 11, 2026

NSF workshop urges a shift to clinic-aware algorithm–hardware co-design and sustained public investment in shared data, compute, and validation ecosystems to de-risk and speed medical AI commercialization. Without standardized benchmarks, continuous monitoring, and governance reforms, promising medical AI risks slow adoption, regulatory friction, and market concentration.

low descriptive Full text arxiv

Report for NSF Workshop on Algorithm-Hardware Co-design for Medical Applications · Peipei Zhou, Zheng Dong, Insup Lee, Aidong Zhang, Robert Dick, Majid Sarrafzadeh... · Mar 11, 2026

Co-designed quantum–classical supercomputers are the missing link to scale useful hybrid algorithms: integrating QPUs with GPUs/CPUs and middleware can slash manual orchestration and speed materials and drug discovery, but the capital- and skill-intensity will favor well-resourced firms and cloud or national testbeds.

n/a descriptive Full text arxiv

Reference Architecture of a Quantum-Centric Supercomputer · Seetharami Seelam, Jerry M. Chow, Antonio Córcoles, Sarah Sheldon, Tushar Mittal... · Mar 11, 2026

A dominance-based RLHF method cuts rare harmful outputs without sacrificing helpfulness. Using entropically regularized optimal transport to enforce stochastic dominance gives tunable, provable control over tail risk that can be aligned with firm or regulator risk preferences.

medium theoretical Full text arxiv

Safe RLHF Beyond Expectation: Stochastic Dominance for Universal Spectral Risk Control · Yaswanth Chittepu, Ativ Joshi, Rajarshi Bhattacharjee, Scott Niekum · Mar 11, 2026

An AI trained on ~7,400 paired CBCT studies produces draft oral/maxillofacial reports at intermediate-radiologist quality and, as a co-authoring tool, systematically improves final report quality—helping novices reach intermediate standards, lifting intermediates toward senior quality, and cutting omission-related errors among seniors.

medium descriptive Full text arxiv

Bridging the Skill Gap in Clinical CBCT Interpretation with CBCTRepD · Qinxin Wu, Fucheng Niu, Hengchuan Zhu, Yifan Sun, Ye Shen, Xu Li, Han Wu, Leqi L... · Mar 11, 2026

A KL-based shrinkage meta-analysis adaptively pools across heterogeneous studies, producing closed-form estimators that lower mean-squared error and deliver valid confidence intervals without assuming parameter homogeneity. The method offers a principled way to combine estimates across firms, regions or deployments—useful for evaluating AI impacts when settings differ.

n/a theoretical Full text arxiv

Redefining shared information: a heterogeneity-adaptive framework for meta-analysis · Elizabeth M. Davis, Emily C. Hector · Mar 11, 2026

Luck gets you first through the door, but not the lock: rare fast discoveries driven by extreme-value first-passage times explain fragile early leads, while only one-sided interaction advantages (exclusive access, preferential treatment or asymmetric reinforcement) can convert those leads into stable monopolies; moreover, small spatial/network disadvantages demand exponentially more scale to match early-discovery probabilities.

n/a theoretical Full text arxiv

Macroscopic Dominance from Microscopic Extremes: Symmetry Breaking in Spatial Competition · Stuti Guha, Shawn D. Ryan, Bhargav R. Karamched · Mar 11, 2026

Optimizing for average reward can hide persistent failure: when environment dynamics are non-ergodic a policy that looks optimal in expectation can lock a deployed agent into long-run low-reward regimes, so evaluators and regulators should prioritize time-average or distributional guarantees rather than ensemble means.

n/a theoretical Full text arxiv

Ergodicity in reinforcement learning · Dominik Baumann, Erfaun Noorani, Arsenii Mustafin, Xinyi Sheng, Bert Verbruggen,... · Mar 11, 2026

A predictive prompt-selection method halves (or more) the costly rollouts needed for RL finetuning and improves reasoning accuracy across math, planning and visual-geometry benchmarks; by making iterative finetuning cheaper and faster, the technique could materially lower the marginal compute cost of model improvement for practitioners.

medium descriptive Full text arxiv

Dynamics-Predictive Sampling for Active RL Finetuning of Large Reasoning Models · Yixiu Mao, Yun Qu, Qi Wang, Heming Zou, Xiangyang Ji · Mar 11, 2026

Sentiment judgments of short, context-free team messages depend more on the message than the rater: individuals' labels are only moderately stable and ambiguous wording drives most disagreement, while mood traits produce only small positive biases. The result warns that automated sentiment signals—often trained on pooled human labels—can be noisy and systematically biased unless annotator and item heterogeneity are explicitly modeled.

medium correlational Full text arxiv

Exploring Indicators of Developers' Sentiment Perceptions in Student Software Projects · Martin Obaidi, Marc Herrmann, Jendrik Martensen, Jil Klünder, Kurt Schneider · Mar 11, 2026

Growing agents through ongoing conversations with domain experts turns tacit know-how into reusable knowledge assets, lowering per-task effort and improving consistency while shifting costs from one-time engineering to continuous practitioner time; the approach scales through iterative crystallization but depends on sustained engagement and risks overfitting to individual practices.

low theoretical Full text arxiv

Nurture-First Agent Development: Building Domain-Expert AI Agents Through Conversational Knowledge Crystallization · Linghao Zhang · Mar 11, 2026

« Prev 1 2 3 … 61 62 63 … 67 68 Next »