Video LLMs
PulseAugur coverage of Video LLMs — every cluster mentioning Video LLMs across labs, papers, and developer communities, ranked by signal.
3 day(s) with sentiment data
-
Video LLMs struggle with event counting due to readout failures, new research finds
Researchers have identified a significant readout failure in Video Large Language Models (LLMs) that hinders their ability to accurately count events, such as objects crossing a screen. While the models possess the nece…
-
New Qwen model, video LoRA, and SVU research released; Anthropic users report faster limits
A new, smaller version of the Qwen model, Qwen3.5-4B-3D-Coder, has been released on Hugging Face, offering instructions for integration with various tools like llama.cpp, Ollama, and Docker. Separately, LiconStudio has …
-
New framework tackles video LLM hallucinations
Researchers have developed a new training-free framework called Attention-guided Representation Perturbation for Contrastive Decoding (ARP-CD) to address hallucinations in Video Large Language Models (Video LLMs). This …
-
New methods boost efficiency for Video Large Language Models · 4 sources tracked
Researchers have developed new methods for compressing visual tokens in Video Large Language Models (VLLMs) to improve inference efficiency. One approach, Video Codec-Guided Training-Free Token Compression, leverages ex…
-
New MotionBlind benchmark reveals Video-LLMs struggle with motion understanding
A new benchmark called MotionBlind has been developed to test the motion understanding capabilities of Video Large Language Models (Video-LLMs). Researchers found that most open-source Video-LLMs perform poorly, often f…
-
New TempCloze benchmark tests Video-LLMs' temporal reasoning
Researchers have introduced TempCloze, a new benchmark designed to evaluate the temporal reasoning capabilities of Video-LLMs. This benchmark aims to mitigate linguistic shortcuts by presenting models with the beginning…
-
New TempCloze benchmark reveals temporal alignment bottleneck in Video-LLMs
A new benchmark called TempCloze has been developed to evaluate the visual temporal reasoning capabilities of Video-LLMs. This benchmark presents models with the beginning and end of a video and asks them to identify th…
-
Huawei IJCAI 2026 papers highlight efficiency gains in AI models · 1 source tracked
Huawei presented four papers at IJCAI-ECAI 2026, shifting focus from scaling model size to optimizing efficiency and design. One paper details a hierarchical Vision Transformer (ViT) scaled to 30 billion parameters, ach…
-
New SlotNarrative Interface Boosts Video-LLM Token Efficiency
Researchers have introduced SlotNarrative, a novel interface designed to make Video Large Language Models (Video-LLMs) more token-efficient. This system organizes videos into persistent object narratives using compact o…
-
Two new methods tackle Video-LLM token compression for efficiency
Two new research papers propose novel methods for compressing video tokens in Video Large Language Models (Video-LLMs) to improve efficiency. The first paper introduces NovaCov, a set-wise token compressor designed for …
-
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…
-
Video-HOCA benchmark reveals Video-LLMs struggle with anomaly explanation
A new diagnostic benchmark called Video-HOCA has been introduced to evaluate the physical anomaly reasoning capabilities of Video Large Language Models (Video-LLMs). This benchmark utilizes an Ontological-Causal taxonom…
-
New benchmark diagnoses visual grounding in video LLMs
A new paper introduces the Visual Dependency Gap (VDG) to assess the visual grounding capabilities of video large language models (LLMs). The VDG measures the difference in accuracy between models processing original vi…
-
New E-VQA Task Aims to Make Video LLMs More Transparent
Researchers have introduced Evidence-Backed Video Question Answering (E-VQA), a new task designed to make Video Large Language Models (Video LLMs) more transparent. Current models often provide answers without clear vis…
-
New SLVMBench benchmark reveals video LLMs struggle with skill learning from long memory
Researchers have introduced SLVMBench, a novel benchmark designed to evaluate the ability of video large language models (video-LLMs) to learn skills from extended video memory and apply them in real-time scenarios. The…
-
GeoTrace framework compresses video tokens for efficient Video LLMs
Researchers have introduced GeoTrace, a novel framework designed to enhance the efficiency of Video Large Language Models (Video LLMs) by compressing visual tokens. This training-free method decomposes video evidence in…
-
TimeThink framework enhances temporal reasoning in Video LLMs · arXiv paper
Researchers have introduced TimeThink, a novel reinforcement learning framework designed to enhance the temporal reasoning capabilities of Video Large Language Models (Video-LLMs). This approach focuses on optimizing th…
-
Denoising Attention (DnA) improves visual task performance
Researchers have introduced Denoising Attention (DnA), a novel method designed to improve the performance of attention-based models in visual tasks. DnA addresses the issue of noisy attention patterns produced by standa…
-
New benchmarks and frameworks enhance video temporal grounding
Researchers have introduced new benchmarks and frameworks for improving temporal grounding in long-form videos. One study posits that hour-scale video grounding is primarily a search problem, not a recognition one, and …
-
New MACD method combats video LLM hallucinations
Researchers have developed a new inference strategy called Model-Aware Contrastive Decoding (MACD) to combat hallucinations in video language models. MACD leverages the model's own feedback to identify and target specif…