Quantum Circuit Vision is a new cost-aware benchmark designed to evaluate how well multimodal AI agents can interpret quantum circuit diagrams and generate executable code.
This empirical study analyzes containerization practices in open-source machine learning projects to understand how iterative workflows affect build performance and container size.
The Co4ICF framework couples a physics-informed surrogate model with a reinforcement learning optimizer to prevent out-of-distribution errors in inertial confinement fusion simulations.
A new study analyzes the systemic risks posed by large language models in misinformation ecosystems, proposing a framework to categorize vulnerabilities and defense strategies.
Researchers introduced DiffUE, a method using diffusion autoencoders to improve the effectiveness of unlearnable examples in protecting images from unauthorized AI training.
This study explores the potential for malicious actors to automate scientific fraud by weaponizing AI systems to generate misleading research data.
This study proposes using Laguerre Geometry to model and analyze how concepts are structured and separated within large language models.
This paper introduces a theoretical framework for quantifying uncertainty in AI models by decomposing subjective risk based on loss functions.
A new approach to agentic routing is proposed to optimize model selection within execution harnesses by leveraging the specialized strengths of different AI models.
A new loss function designed for time-series forecasting aims to improve the prediction of rare demand spikes by prioritizing extreme-value accuracy.
A technical report details a CVPR 2026 challenge focused on testing the robustness of autonomous driving vision-language agents against adversarial multimodal attacks.
A study into the mechanistic interpretability of LLM-as-a-judge models reveals that scoring biases can be identified and analyzed within the internal hidden states of the models.
This paper proposes a formal safety framework for an AI predictor designed to provide honest outputs by conditioning on epistemically contextualized data.
Alibaba and Tencent received significant net inflows of Southbound capital, totaling approximately 3.09 billion HKD and 1.81 billion HKD respectively.
Researchers developed a method called RPSG that utilizes private seeds and differential privacy to generate synthetic data while balancing privacy and utility.
A study demonstrates that small-scale, targeted edits to Wikipedia articles can measurably influence the values and outputs of large language models trained on that data.
A new optimization method called Deep-Feature-Anchored Preference Optimization resolves the trade-off between audio quality and speech intelligibility in streaming target speaker extraction models.
NaviCache introduces a calibration-free caching method to reduce the computational costs associated with generating video via diffusion models.
Researchers have introduced SETA, a framework designed to scale the training of terminal-based language model agents by generating diverse and coherent task instructions.
The JEEVHITAA platform introduces a mobile system designed to facilitate coordinated, multi-actor information sharing and workflows within healthcare settings.
Researchers developed SWIFT, a framework that incorporates structural priors from traffic networks to improve trajectory prediction and generalization in autonomous driving.
This paper introduces an execution-gated self-distillation method that uses a deterministic launch filter to improve the cross-family generalization of code generators in game development.
ReportMedSAM utilizes natural language radiology reports to guide medical image segmentation, improving scalability and adaptability to new anatomical structures.
This paper explores the application of BERT-based models to improve dialogue state tracking in task-oriented conversational AI systems.