VideoLLMs
PulseAugur coverage of VideoLLMs — every cluster mentioning VideoLLMs across labs, papers, and developer communities, ranked by signal.
- 2026-10-01 research_milestone A new method, Temporal Activation Injection (TAI), was proposed to improve temporal reasoning in VideoLLMs by reinforcing temporal representations at inference time. source
3 day(s) with sentiment data
-
Video-language models inherently encode evidence readiness, study finds
Researchers have discovered that frozen video-language models inherently encode a signal indicating whether sufficient evidence has been gathered to answer a question. This signal, termed 'evidence readiness,' is presen…
-
Video-language models inherently encode evidence-readiness signals
Researchers have discovered that frozen video-language models inherently encode an "evidence-readiness" signal, indicating whether sufficient evidence for a question has arrived in the video stream. This signal is linea…
-
New method reinforces temporal reasoning in VideoLLMs
Researchers have identified a key weakness in Video Large Language Models (VideoLLMs): their struggle to maintain temporal reasoning from intermediate layers to the output. They discovered that reversing the frame order…
-
Survey details high costs and efficiency gaps in video LLMs
A new survey paper examines the computational costs associated with video large language models (VideoLLMs), which are becoming increasingly sophisticated but also more resource-intensive. The paper categorizes and anal…
-
StateTrace framework enhances VideoLLM reasoning for invisible objects · 2 sources tracked
Researchers have developed StateTrace, a new object-centric framework designed to improve the spatiotemporal reasoning capabilities of VideoLLMs, particularly in scenarios involving long videos where objects may become …
-
New PoisonVID attack bypasses safety features in VideoLLMs
Researchers have developed a novel attack called PoisonVID that can bypass safety measures in Video Large Language Models (VideoLLMs). These models are used to moderate user-generated content by sampling key frames, but…
-
New GSTEP framework prunes VideoLLM tokens for efficiency
Researchers have developed GSTEP, a novel pruning framework designed to enhance the efficiency of Video Large Language Models (VideoLLMs). Unlike previous methods that prune tokens locally within segments, GSTEP models …
-
New dataset VideoNorms tests cultural awareness in VideoLLMs
Researchers have developed VideoNorms, a new dataset designed to evaluate the cultural awareness of Video Large Language Models (VideoLLMs). The dataset includes over 3,000 human judgments derived from popular US and Ch…
-
VideoLLMs can now watch and think simultaneously with new VST paradigm
Researchers have introduced Video Streaming Thinking (VST), a new paradigm designed to enable online Video Large Language Models (VideoLLMs) to process and reason about video content simultaneously. This approach aims t…
-
New benchmark NeMo challenges video LLMs on temporal understanding
Researchers have introduced NeMo, a novel task and benchmark called NeMoBench, designed to evaluate the temporal understanding capabilities of video large language models (VideoLLMs). The task, inspired by the 'needle i…