Yue Li

A unified framework reveals that existing LLM policy optimization methods often overlook compound failures that require simultaneous adjustments to both trajectory and reward components.

Existing VLMs misinterpret critical evidence in exploratory manipulation, but a novel distillation approach boosts action recovery accuracy by nearly half.

Continuous interaction with LLMs can skew grading standards, leading to systemic unfairness in academic assessments.

ASR-driven data augmentation boosts Alzheimer's detection accuracy by over 4%, showcasing the potential of synthetic speech in clinical diagnostics.

Robot RL training can be dramatically sped up (3-10x) by decoupling CPU-based simulation from GPU-based learning, challenging the assumption that GPU-resident physics is essential for efficiency.

A new dataset of 1,111 transvaginal ultrasound images with detailed annotations finally enables AI-powered diagnosis of Cesarean Scar Defects, a condition frequently missed by sonographers.

LLM safety classifiers can be made dramatically more robust against jailbreaks by teaching them to "think twice" via lightweight, self-reflection fine-tuning.

Despite advances in LLMs, even syntactically correct outputs often fail to achieve the intended state transitions when translating natural language into executable Ethereum transactions, revealing a critical gap in "reasoning-to-execution" capabilities.

Key contribution not extracted.

Freeing adversarial perturbations from pixel-level constraints by injecting them into the latent space of diffusion models unlocks significantly more robust defense against facial deepfakes.