🏢 Peking University

SpikeReveal: Self-supervised learning unlocks sharp video sequences from blurry, real-world spike camera data, overcoming limitations of prior supervised approaches.

STIR: A novel spatio-temporal network reconstructs high-quality images from spiking camera data by jointly refining motion and intensity information for efficient and accurate high-speed imaging.

SPARKLE: A single-loop primal-dual framework unifies decentralized bilevel optimization, enabling flexible heterogeneity-correction and mixed update strategies for improved convergence.

SMART: a novel self-supervised model tackles missing EHR data, improving patient health status prediction via missing-aware attention and robust pre-training.

SILENCE, a novel lightweight system, protects user privacy in offloaded speech understanding on resource-constrained devices by selectively masking short-term audio details without impacting long-term…

SemanIR boosts image restoration efficiency by cleverly sharing key semantic information within Transformer layers, achieving state-of-the-art results across multiple tasks.

SfPUEL: A novel end-to-end SfP method achieves robust single-shot surface normal estimation under diverse lighting, integrating PS priors and material segmentation.

Deep Equilibrium Models (DEQs) outperform standard neural networks, but lack theoretical understanding. This paper provides general separation results showing DEQ’s superior expressivity and character…

SemFlow: A unified framework uses rectified flow to seamlessly bridge semantic segmentation and image synthesis, achieving competitive results and offering reversible image-mask transformations.

SegVol: A universal, interactive 3D medical image segmentation model achieving state-of-the-art performance across diverse anatomical categories.

Dissected Dynamics Modeling (DDM) excels at multi-modal visual reinforcement learning by cleverly separating and integrating common and unique features across different sensory inputs for more accurat…

Scalable MAPPO-L: Decentralized training with local interactions ensures safe, high-reward multi-agent systems, even with limited communication.

ZO-GDEGA: A unified algorithm achieves faster, more robust zeroth-order minimax optimization with lower complexity and weaker conditions, solving stochastic nonconvex-concave problems.

RoboMamba: a novel robotic VLA model efficiently combines reasoning and action, achieving high speeds and accuracy while requiring minimal fine-tuning.

Richelieu: a self-evolving LLM-based AI agent masters Diplomacy, a complex game requiring strategic planning and negotiation, without human data, by integrating self-play for continuous improvement.

Researchers developed Camera-Agnostic Patch (CAP) attacks, improving adversarial patch reliability by simulating camera image processing in attacks against person detectors.

ReVideo enables precise local video editing by independently controlling content and motion, overcoming limitations of existing methods and paving the way for advanced video manipulation.

Boosting time series forecasting accuracy, Retrieval-Augmented Diffusion Models (RATD) leverage relevant historical data to guide the diffusion process, overcoming limitations of existing models and d…

Reinforcement learning paradigms exhibit a representation complexity hierarchy: models are easiest, then policies, and value functions are hardest to approximate.

ReEvo, a novel integration of evolutionary search and LLM reflections, generates state-of-the-art heuristics for combinatorial optimization problems, demonstrating superior sample efficiency.