SAGEAgent is a self-evolving agent designed to optimize the acquisition of clinical diagnostic modalities for cost-aware multimodal survival prediction in cancer patients.
The paper introduces CogniConsole, an architecture that externalizes inference-time control into a structured interface to improve the reliability of large language model interactions.
The authors provide a theoretical characterization of the trade-offs between fairness and regret in bandit algorithms used for decision-making.
The NodeImport method introduces a new approach to address class imbalance in graph neural networks by assessing node importance during classification tasks.
Researchers developed a multi-expert system using iterative fine-tuning to improve optical character recognition for historical Manchu documents with limited training data.
The proposed EMAGN architecture improves the scalability of traffic forecasting by using learned clustering to linearize spatial attention mechanisms.
Researchers introduced a diagnostic tool called operator-on-F to evaluate latent world models in reinforcement learning by comparing model predictions against environment observations.
The proposed Reward Transport framework utilizes optimal transport coupling during training to embed controllable structures directly into learned flow fields for molecular property control.
This paper explores the differences between current predictive AI training regimes and biological intelligence, arguing for the importance of grounded world models in future embodied AI development.
Researchers analyzed evasion strategies against white-box monitors used for LLM safety and proposed new defensive mechanisms to improve auditing reliability.
This study applies AI techniques to analyze academic curriculum patterns to help improve graduation rates in software engineering programs.
Lenovo has established a new wholly-owned subsidiary in Guizhou with a registered capital of 3 million yuan to focus on AI hardware sales and software development.
This paper proposes a knowledge tracing method that models student learning by disentangling ability and proficiency states from interaction sequences.
A new study proposes using inductive logic programming to improve the interpretability of reinforcement learning policies in safety-critical applications.
This theoretical paper introduces a geometric coding theorem that extends classical information theory concepts to symmetry groups.
A new benchmark, PhysAssistBench, evaluates the ability of large language models to assist physicians by coordinating clinical knowledge, EHR interactions, and patient communication.
This paper argues that large-scale text-to-video generation serves as an effective pre-training paradigm for developing general-purpose computer vision models.
A new co-evolutionary diffusion model is proposed to improve the accuracy of gene regulatory network inference from single-cell data.
Researchers introduced MedRealMM, a large-scale multimodal benchmark based on real-world Chinese online medical consultations to better evaluate LLMs in clinical practice.
The LLaMEA framework has been extended to automatically generate complete multi-objective Bayesian optimization algorithms using large language models as evolutionary operators.
A new explainable AI framework is proposed to assist banking internal audit teams in detecting and justifying anomalous transaction patterns.
A reinforcement learning approach is proposed to optimize signal coverage and energy efficiency in reconfigurable intelligent surfaces for edge computing.
The authors introduce Partially Observed Structural Causal Models to better account for upstream contexts in causal modeling frameworks.
This academic work explores the progression of social simulation techniques from traditional agent-based modeling to modern AI-driven digital twins.