Qwen2.5-VL
PulseAugur coverage of Qwen2.5-VL — every cluster mentioning Qwen2.5-VL across labs, papers, and developer communities, ranked by signal.
4 day(s) with sentiment data
-
Vision-Language Models: How ViT, CLIP, and Qwen2.5-VL Process Images
A recent analysis delves into the internal workings of Vision-Language Models (VLMs), explaining their three-stage image processing pipeline. This process involves a Vision Transformer (ViT) for patch slicing and mergin…
-
New tuning methods teach LMMs skills without forgetting
Researchers have developed novel methods for fine-tuning large multimodal models (LMMs) to acquire new skills without degrading existing capabilities. By analyzing the output token distribution and identifying a 'counti…
-
AI models transfer preferences via curated near-duplicate images
Researchers have demonstrated a phenomenon called subliminal learning, where one AI model can transfer its preferences to another through data curation, even when the source data remains unchanged. By selecting near-dup…
-
New benchmarks challenge multimodal AI models in generation and chart-to-code tasks · 2 sources tracked
Two new benchmarks have been released to evaluate the capabilities of multimodal large language models. The first, MMMG, focuses on multitask multimodal generation across image, audio, and interleaved text and image/aud…
-
llama.cpp bug causes non-deterministic results for M-RoPE embedding batches
A bug in the llama.cpp library causes incorrect results when processing embedding batches for M-RoPE models like Qwen3.5 and Qwen2.5-VL. The issue stems from a heap buffer overflow where the library reads past the alloc…
-
Vision-Language Models Show Promise in Grading Handwritten Exams
Researchers have evaluated the effectiveness of various vision-language models (VLMs) in grading handwritten examinations for outcome-based education. The study compared configurations including Qwen2.5-VL, InternVL3, a…
-
StepPrune method accelerates multimodal LLM inference by adaptively selecting visual tokens
Researchers have developed StepPrune, a novel method for adaptively selecting visual tokens in multimodal large language models (MLLMs) to accelerate inference. Unlike previous top-K methods that treat tokens independen…
-
Open-source multimodal models challenge GPT-4o on cost and performance · 2 sources tracked
Open-source multimodal models are rapidly catching up to GPT-4o in performance and cost-effectiveness, with several models like Alibaba's Qwen2.5-VL and Mistral's Pixtral 12B offering competitive capabilities for tasks …
-
Hugging Face rebuilds AUTOMATIC1111 UI with Gradio Workflow
Hugging Face has developed a new tool called Workflow1111, which reconstructs the functionality of the AUTOMATIC1111 Stable Diffusion Web UI using Gradio's workflow capabilities. This new workflow is composed of eleven …
-
New benchmarks and methods tackle LLM hallucinations across modalities and domains
Researchers are developing new methods and benchmarks to detect and mitigate hallucinations in large language models (LLMs) across various modalities and domains. OmniHallu offers a unified framework for detecting hallu…
-
Video-FLAIR framework learns adaptive reasoning for multimodal queries
Researchers have introduced Video-FLAIR, a novel training framework designed to optimize reasoning strategies for multimodal queries. This system employs reinforcement learning to dynamically select the most appropriate…
-
New research tackles AI hallucinations in video and language models
Researchers are developing new methods to combat hallucinations in AI models, particularly in video-language and large language models. One approach, CounterVid, uses counterfactual video generation to create synthetic …
-
New VLM extracts document data without OCR, outperforming larger models
Researchers have developed a new method for extracting key-value pairs from document images without relying on traditional OCR preprocessing. They fine-tuned a compact 256M-parameter vision-language model called SmolDoc…
-
New benchmarks probe VLM spatial reasoning, revealing localization and relation understanding gaps · 5 sources tracked
Researchers are developing new benchmarks and methodologies to better understand and diagnose spatial reasoning failures in vision-language models (VLMs). One approach, GUI-Primitives, uses contrastive instruction pairs…
-
New framework Zero-MELO boosts MLLM micro-gesture recognition
Researchers have developed Zero-MELO, a novel framework designed to enhance the performance of Multimodal Large Language Models (MLLMs) in micro-gesture recognition (MGR). The framework addresses limitations in MLLMs' a…
-
New frameworks boost AI long video understanding efficiency
Researchers have developed two new frameworks, EcoFrame and EviSelect, designed to improve the efficiency of long video understanding by large language models. EcoFrame uses a training-free approach that adapts the fram…
-
New PhyCheck dataset evaluates Video LLMs' understanding of physical laws
Researchers have introduced PhyCheck, a new dataset designed to evaluate and improve the physical law understanding capabilities of Video Large Language Models (VideoLLMs). The dataset includes coarse-grained and fine-g…
-
New frameworks enhance VLM reasoning with visual tokens and self-diagnosis · 3 sources tracked
Researchers have developed new frameworks to enhance the reasoning capabilities of Vision-Language Models (VLMs). One approach, Chain-of-Visual-Thought (COVT), uses continuous visual tokens to capture dense perceptual i…
-
New methods enhance spatial reasoning in multimodal LLMs · 4 sources tracked
Researchers have developed new methods to improve spatial reasoning in multimodal large language models (MLLMs). SpatialCLI uses specialist vision models as tools to enhance MLLMs' perception and reasoning, achieving si…
-
New frameworks boost MLLM long-video understanding by adaptive frame processing · 3 sources tracked
Three new research papers introduce novel frameworks for enhancing the long-video understanding capabilities of multimodal large language models (MLLMs). These approaches aim to overcome the limitations of fixed context…