Honghao Wang

VisionWeaver: From Phenomenon Recognition to Cause Diagnosis, Opening a New Chapter in AI Visual Hallucination Research

AI news

VisionWeaver: From Phenomenon Recognition to Cause Diagnosis, Opening a New Chapter in AI Visual Hallucination Research

VisionWeaver & VHBench-10 — Root Cause Diagnosis for LVLM Hallucinations Date: 2025-11-14 · Location: Shanghai The Bilibili User Technology Center has unveiled VisionWeaver and its diagnostic benchmark VHBench-10, offering a new paradigm for understanding and tackling hallucinations in large vision-language models (LVLMs). --- 📖 Preface For years, we’ve known LVLMs can misinterpret

Train the Model with 1.55 Million Simulated Videos: GVE Learns 9 Video Retrieval Skills at Once

AI news

Train the Model with 1.55 Million Simulated Videos: GVE Learns 9 Video Retrieval Skills at Once

Quantum Bit|QbitAI Breaking the Bottleneck in Video Retrieval Current video retrieval research has reached a closed-loop bottleneck: For years, narrow-domain benchmarks like MSRVTT dominated, optimizing models for coarse-grained text queries. This led to: * Biased training data * Limited capabilities * Poor handling of fine-grained semantics * Weak long-context understanding * Inability to handle