Qwen
20 articles found in this topic.
HKUST and Partners Introduce Robust-R1, a Degradation-Aware Reasoning Paradigm for MLLMs
HKUST and partners introduce Robust-R1, a novel degradation-aware reasoning paradigm for Multimodal Large Language Models (MLLMs). This approach enables MLLMs to diagnose and withstand real-world visual degradations, moving beyond implicit adaptation. Robust-R1 aims to significantly improve the robustness and quality of MLLMs for industrial applications.
Epoch AI Report: Chinese AI Models Lag US Counterparts by Seven Months
A new Epoch AI report reveals Chinese AI models lag US counterparts by an average of seven months in capability. This disparity is linked to open-source versus closed-source development, with US models showing rapid advancements while Chinese models exhibit a 'leapfrog' catch-up pattern.
Jensen Huang Showcases DeepSeek, Kimi AI Models for Next-Gen Chip Performance
Nvidia CEO Jensen Huang showcased Chinese AI models DeepSeek and Kimi K2 at CES 2026, demonstrating the capabilities of the upcoming Rubin architecture. These models, utilizing the Mixture of Experts approach, showed significant performance gains and cost reductions, highlighting the growing impact of Chinese AI on global benchmarks.
CAS Launches VTCBench to Evaluate Visual Language Models in Long Text Compression
CAS introduces VTCBench, a new benchmark to evaluate visual language models (VLMs) in long text compression. It assesses cognitive limits through retrieval, reasoning, and memory tasks, revealing a "U-shaped curve" in model performance and spatial attention bias in VLMs.
UIC and Adobe Introduce VGent Model for Multi-Object Localization, Boosting F1 Score by 20 Points
UIC and Adobe introduce VGent, a new modular AI architecture for multi-object localization. It significantly boosts performance in visual grounding tasks, improving F1 scores by over 18 points while maintaining consistent inference speed.
QuCo-RAG Improves Dynamic RAG Performance by 14 Points Using Pre-training Corpus Statistics
QuCo-RAG, a new Dynamic RAG approach, quantifies uncertainty using pre-training corpus statistics, improving performance by 5-14 points on various LLMs. It addresses hallucination by identifying low-frequency entities and zero co-occurrence, shifting from internal model signals to objective data.
DeepSeek and Other Major Tech Innovations Began as Side Projects
Many major tech innovations, including DeepSeek, Twitter, and Apple's Macintosh, originated as side projects rather than core strategic initiatives. This trend highlights how experimental endeavors often lead to groundbreaking products, contrasting with the struggles of heavily funded core projects like Windows Phone or Google Glass.
Chinese AI Models Qwen3 and DeepSeek R1 Lead "Meaning Test" Benchmark
Chinese open-source AI models, Qwen3 and DeepSeek R1, have topped a new U.S. benchmark, the "Meaning Test," designed to assess AI's understanding of complex human concepts. Qwen3 secured first place, outperforming major U.S. models, highlighting a growing capability of AI in philosophical and ethical engagement.