Yifan Lu

Compact, automatically selected anatomical regions can drastically reduce hallucinations in medical VLMs without requiring expert annotations.

RA-CLIPScore reveals spatial biases in generative models, offering a more interpretable evaluation that aligns with human perception of visual diversity.

Re-ranking control alone boosts key performance metrics by over 2%, but extending it to fine ranking unlocks even greater gains without sacrificing system stability.

LSA slashes indexing overhead while maintaining full attention performance, enabling efficient long-context processing for models with up to one million tokens.

By cutting end-to-end serving resource consumption by over 50% while boosting user engagement metrics, RecGPT-V3 redefines efficiency in large-scale recommender systems.

GMM-EVA slashes visual token budgets in long video understanding by intelligently prioritizing keyframes based on event-level structures.

A real-time generative world model can synthesize complex driving scenarios that traditional simulators struggle to capture, enabling safer and more effective evaluation of autonomous vehicle policies.

MONA unlocks faster LLM pretraining and superior downstream performance by turbocharging the Muon optimizer with Nesterov-style acceleration, leaving AdamW in the dust.

Text-to-image models can now generate megapixel images 6x faster and with better quality by replacing traditional decoders with a pixel diffusion-based upsampler.

Channel-wise adaptive learning rates in Gated Delta Networks unlock superior long-context recall, rivaling softmax attention without the quadratic cost.

Achieve full-attention accuracy with 10x operator speedup and 4.7x throughput improvement in long-context LLM inference by overlapping KV cache transfers with computation.