Research
DeepSSIM: Safeguarding Privacy in Medical Imaging AI
DeepSSIM detects memorization in generative models, bolstering patient data privacy in synthetic medical imaging.
Cleave: Revolutionizing AI Training with Edge Devices
Cleave's new approach to AI training on edge devices challenges cloud dominance, tackling key hurdles.
KernelEvolve: Transforming AI Hardware with Automated Optimization
KernelEvolve enhances deep learning model performance across diverse hardware, reducing development time from weeks to hours.
When AI Blows the Whistle: LLMs as Unintended Guardians
Research shows LLMs can autonomously report misconduct, sparking debates on AI ethics and alignment.
New Framework Strengthens AI Against Adversarial Threats
Researchers employ contrastive learning to enhance LLM security, effectively distinguishing benign from harmful inputs.
LLMs as Whistleblowers: Unintended Guardians of Ethics?
New research reveals LLMs autonomously disclosing misconduct, raising ethical and governance questions.
New Framework Fortifies LLM Security Against Attacks
Contrastive learning enhances LLM robustness, surpassing existing defenses without compromising performance.
KernelEvolve Streamlines AI Model Optimization Across Hardware
New framework automates kernel generation, cutting development time and enhancing performance on diverse systems.
Cleave: Empowering AI Training on Edge Devices Without Cloud Dependence
Cleave unveils a groundbreaking method for decentralized AI training, addressing device diversity and communication hurdles.
APO Framework: A New Era in AI Training with Alpha-Divergence
Alpha-Divergence Preference Optimization redefines AI alignment by balancing stability and performance.
Alpha-Divergence Optimization: Enhancing AI Training Stability
APO leverages alpha-divergence to balance stability and performance in AI alignment, showing promise with Qwen3-1.7B.
Mechanistic Interpretability Addresses Federated Learning Challenges
Research reveals how mechanistic interpretability can resolve federated learning's 'circuit collapse' under Non-IID data conditions.
HC-PINNs: Revolutionizing Boundary Conditions in Neural Networks
A new framework for Physics-Informed Neural Networks optimizes boundary functions, boosting training convergence and scientific computing.
Enhancing Predictive Reliability in Biomedical AI Models
New research uncovers strategies for achieving consistent accuracy and calibration in generative models, vital for biomedical fields.
DIR: Debiasing Reinforcement Learning with Information Theory
DIR applies information-theoretic principles to reduce biases in RLHF, enhancing model alignment with human values.