Yi Zhang
12 papers · Latest:
ReCoVR: Closing the Loop in Interactive Composed Video Retrieval
ReCoVR introduces a dual-pathway architecture for interactive composed video retrieval, using reflexive perception to refine search with user feedback and retrieval history.
PRISM: Refracting the Entangled User Behavior Space for E-Commerce Search
PRISM disentangles user preference and item relevance in e-commerce search by explicitly modeling their interaction, improving robustness and semantic consistency.
Penalized Likelihood for Dyadic Network Formation Models with Degree Heterogeneity
This paper introduces a penalized likelihood method to robustly estimate dyadic network formation models, addressing existence and bias issues from degree heterogeneity.
ProMax: Exploring the Potential of LLM-derived Profiles with Distribution Shaping for Recommender Systems
ProMax uses LLM-derived profiles and distribution shaping to significantly improve recommender systems by guiding models to learn unseen item preferences.
Vision SmolMamba: Spike-Guided Token Pruning for Energy-Efficient Spiking State-Space Vision Models
Vision SmolMamba introduces spike-guided token pruning in a state-space model for energy-efficient spiking vision, achieving superior accuracy-efficiency.
FreeScale: Distributed Training for Sequence Recommendation Models with Minimal Scaling Cost
FreeScale optimizes distributed training for sequence recommendation models, reducing computational bubbles by up to 90.3% on 256 H100 GPUs.
Disagreement as Signals: Dual-view Calibration for Sequential Recommendation Denoising
DC4SR denoises sequential recommendations by calibrating semantic priors from LLMs with model learning dynamics to handle evolving user interests.
Supplement Generation Training for Enhancing Agentic Task Performance
SGT trains small LLMs to generate supplemental text, boosting large LLM performance on agentic tasks without costly retraining.
SID-Coord: Coordinating Semantic IDs for ID-based Ranking in Short-Video Search
SID-Coord enhances short-video search ranking by integrating trainable semantic IDs to balance memorization and generalization, improving long-tail item performance.
Arbitration Failure, Not Perceptual Blindness: How Vision-Language Models Resolve Visual-Linguistic Conflicts
VLMs often fail to act on what they see, encoding visual info but struggling with arbitration, not perception, which targeted interventions can improve.
DIAURec: Dual-Intent Space Representation Optimization for Recommendation
DIAURec optimizes user and item representations for recommendations by unifying intent and language modeling with a comprehensive optimization strategy.
Coarse Screening
This paper shows that sellers investigating buyers before pricing need only three signal outcomes per buyer, even with rich type spaces, due to limited liability.
📬 Weekly AI Paper Digest
Get the top 10 AI/ML arXiv papers from the week — summarized, scored, and delivered to your inbox every Monday.