EVAThink trains video models to reason when frame evidence is sparse
Using 6.7K curated samples, it improves the base model under both 32- and 64-frame budgets.
Stories per day · 30 days
4 RSS
Relatedmultimodal3reinforcement-learning3huggingface2memory2streaming-video2adaptive-reasoning1agentic-ai1agents1benchmark1benchmarks1chain-of-thought1cyclegrpo1deltastreamer1grounding1hallucination1hypernetwork1lfm31liquid-ai1
24 stories, newest first · one card per story
Stories tagged #video-understanding
SleuthVLM outperforms Gemini on omni video benchmarks
SleuthVLM reaches 75.0% accuracy on omni benchmarks via transcription and agentic reasoning.
AskStream benchmark tests when video models should ask
The strongest model scores only 52.18% Ask-F1 across 438 videos labeled for wait, ask, and answer.
VideoLVR adds latent visual evidence to MLLM video reasoning
Latent tokens reconstructed from the input video let models revisit local frames without large context growth.
Visible-State Supervision targets under-specified streaming video recognition
Under a matched recent-frame setup, ViSS improves OVO-Bench and helps hallucination detection most, while staying comparable on StreamingBench.
CycleGRPO couples region captioning and localization with cycle rewards
On GroundingSuite, SAMTok overall gIoU rises from 57.5% to 68.2%; video tIoU on Charades reaches 51.9% versus 45.6% for Qwen3-VL.
Liquid AI releases LFM3-VL-2B on-device vision-language model
The 1.92B model takes text, images, and video, reasons step by step, and uses the LFM Open License v1.0.
1
Memory as Weights: Internalizing Long-Term History for Streaming Videos
MaLoW stores streaming video history as LoRA weights, beating prior memory methods on OVO-Bench.
DeltaStreamer: Towards Human-Like Agentic Streaming Video Understanding
DeltaStreamer combines change-aware memory and targeted replay for streaming video question answering.
FastBench: Can Streaming VLMs Perceive High-Dynamic Real-World Streams?
FastBench shows streaming VLMs miss fast events; Gemini-3.5-Flash scores only 50.7%.
DataVista: Diagnosing Multimodal LLMs on Data Video Understanding
DataVista finds Gemini-3.1-Pro best of 19 MLLMs at 70% on data-video understanding.
OneSearch-VL: Unified Multimodal Deep Research Agent for Image and Video
OneSearch-VL-8B beats tool-using Qwen3-VL-8B by up to 20.2 points on new visual research benchmarks.
VepAgent: Bridging Causal-Transition via Tool-Augmented Reinforcement Learning for Video Event Prediction
VepAgent uses tool-augmented reinforcement learning to predict video events by modeling unobserved causal transitions.
Rethinking Long-Video Efficiency: A Joint Allocation Perspective on Frames, Pixels, and Front-End Latency
LoHi pairs dense low-resolution video with sparse high-resolution frames, raising long-video accuracy and cutting decode latency.
Foresight: planning future perception in streaming VLMs without retraining
FORESIGHT lets frozen Qwen3-VL-8B plan future perception in streaming video without retraining.
Before It Fades: Reinforcing Temporal Representations at Inference Time in VideoLLMs
Temporal Activation Injection restores fading mid-layer timing signals in VideoLLMs without training.
CoEvoWhen: Policy-Tool Coevolution for Ultra-Long Video Temporal Grounding
CoEvoWhen jointly evolves VLM policies and media tools for ultra-long video temporal grounding without updating parameters.
Video-RSI: Recursive Self-Improvement of Video Understanding Agents via Harness Evolution
Video-RSI lets video agents rewrite their own harnesses, improving accuracy while inspecting fewer frames.
LLMs are General Asynchronous Agents
Qwen 3.x models act as general asynchronous agents for video, games, and monitoring without task-specific training.
TRACE: Temporal Audit and Condition-aware Evaluation of Streaming Video Understanding
TRACE benchmarks streaming video models on timing, triggers, workload, and reliability, not accuracy alone.
TrackEverything: Long Horizon Dense Tracking via De-Duplicating 3D Scene Representations
TrackEverything tracks all visible 3D points in videos beyond 1,000 frames within 40 GB by de-duplicating scene geometry.
Semantic Action Graph: A Shared Representation for Agent Grounding and Human Interpretation of Sports Highlights
SportSAGE uses a semantic action graph schema to ground agent-generated sports highlights and let viewers query, navigate, and verify narratives.
BVB: Benchmarking Agentic Video Understanding via Programmatic Reconstruction in Blender
Blender-VideoBench evaluates agentic video understanding by having agents programmatically reconstruct real videos in Blender scenes.
Introducing agentic video understanding with Gemini
Google DeepMind launches agentic video understanding for Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite, cutting video-analysis tokens up to 88%.
8