Multimodal

Latest: Tencent Opens 5B Youtu-Parsing-Omni, Scores 96.96 on OmniDocBench With One Encoder Across All Modalities · huggingface.co, Oct 29, 2026

Thème Mots-clés suivis : multimodal, vision language, omni

14 AI Weekly issues and 91 live stories from the last 30 days mention Multimodal.

35 mentions au total 1/20 numéros récents -100% vs référence pic 1/numéro suivi depuis 2026-04-18

Tendance des mentions — 20 derniers numéros

Actus en direct (91)

Tencent Opens 5B Youtu-Parsing-Omni, Scores 96.96 on OmniDocBench With One Encoder Across All Modalities

huggingface.co · 2026-10-29 · Score : 25

SenseTime Open-Sources 27B SenseNova-6.8-Code2Video That Programs Each Shot Instead of Generating Pixels

huggingface.co · 2026-10-27 · Score : 24

Google Opens 744M EmbeddingGemma 2, a Unified Text-Image-Audio-Video Embedding Model Running in 191MB On-Device

huggingface.co · 2026-10-28 · Score : 26

Meituan LongCat Paper Tests Whether Discrete Tokens Can Losslessly Represent Visual Signals

huggingface.co · 2026-10-28 · Score : 17

inclusionAI's EvidenceConnect Lifts 4B-27B VLMs Across 25 Perception Benchmarks via RLVR

huggingface.co · 2026-10-27 · Score : 19

Alibaba DAMO Opens RynnWorld-Latent, a Cross-Embodiment Video World Model Trained on 553M Frames

huggingface.co · 2026-10-28 · Score : 20

Peking University and JD's PDRVLA Hits 99.3% on LIBERO With 57M-Param Physical Dynamics Module

huggingface.co · 2026-10-28 · Score : 22

Liquid AI Opens LFM3 Family of 2B-Class Multimodal Models With 256K Context and ~1GB 4-Bit Footprint

liquid.ai · 2026-10-27 · Score : 26

NoizAI's AMBIT Pairs Executable Scene Plans With a Self-Distilling Transfusion for First-Order-Ambisonics Spatial Audio

arxiv.org · 2026-10-26 · Score : 16

Google's StreamVibe Benchmarks Proactive Video Agents That Track Evolving User Intent, Lifts Response Quality 10.7%

huggingface.co · 2026-10-26 · Score : 19

Tencent Open-Sources WeVisOne-VL, a Unified VLM That Pairs Object Detection With Pixel-Level Segmentation and OCR

huggingface.co · 2026-10-22 · Score : 22

HKUST and Tencent's FAPL Forces Video DiTs to Use MLLM Latents for Factor-Aware Cinematic Control

huggingface.co · 2026-10-23 · Score : 21

TTR-Compressor Reframes Sequence Compression as Test-Time Regression, Matches MLP Merger With 55.6% Fewer Vision Tokens

huggingface.co · 2026-10-23 · Score : 17

USTC's SERA Edits VLM Planner Harness Between Rollouts, Hits 51.4% on RoboMME Long-Horizon Manipulation Tasks

huggingface.co · 2026-10-21 · Score : 20

SJTU's SpatialBrain Adds Memory, Skill Library and Cognition Layer to Frozen VLM, Beats Direct Answers on 4 of 5 Benchmarks

huggingface.co · 2026-10-21 · Score : 20

Slot-JEPA Factorizes V-JEPA Features Into 16 Object Slots, Beats V-JEPA 2.1 by 50 Points on Robot Pick-and-Place

huggingface.co · 2026-10-21 · Score : 20

Meta AI and Oxford Open OverSeer, a Multi-Agent World Model That Fixes Cross-Agent Consistency Using Static Views

arxiv.org · 2026-10-21 · Score : 20

Zero-Shot Qwen 3.8-27B VLM Beats Supervised 2B Rerankers With a Single System One Choice Readout

huggingface.co · 2026-10-21 · Score : 18

ETH Zürich's VPM Teaches World Models 4D by Jointly Forecasting Multi-View Video and 3D Point Trajectories

huggingface.co · 2026-10-21 · Score : 22

EvaOmni Benchmark Grades Audio-Visual Reasoning on 1,000 Evidence-Grounded Questions Across 10 Chinese Universities

huggingface.co · 2026-10-20 · Score : 18

Alibaba's FocusGUI Adapter Lifts Omni-Modal GUI Agents by 33.2 Type-Match Points With Frozen Backbones

huggingface.co · 2026-10-20 · Score : 20

MMEnterprise-Bench Caps Top Model GPT-5.6 Sol at 68% on Evolving Multimodal Workplace Records With Visual Reasoning Cliff

huggingface.co · 2026-10-21 · Score : 22

PhysRSI-Eval Paper Finds Public Benchmark Scores Don't Predict How Fast VLM Agents Self-Improve on Robots

huggingface.co · 2026-10-19 · Score : 19

Holi-World Opens 4B Egocentric Vision-Language Model and 80K-QA Benchmark Built From Human-Ego Video

huggingface.co · 2026-10-19 · Score : 20

Stanford's Halcyon T2I Family Beats Internet-Pretrained Peers at 275x Less Compute, No Language Pretraining

huggingface.co · 2026-10-20 · Score : 26

Columbia BrainLoRA Scales PEFT to 104B Multimodal Backbones, Tops Brain-Encoding Leaderboard With 5% of Prior Params

huggingface.co · 2026-10-19 · Score : 18

Tsinghua Stateful Harness Lifts Video World-Model Rule Following From 32% to 92% PairFollow by Decoupling State From Rendering

huggingface.co · 2026-10-20 · Score : 19

Xiaomi Opens Xiaomi-Lens-0, a 4B MLLM That Decodes Pixel Masks From Qwen3.5's Own Visual Features

huggingface.co · 2026-10-19 · Score : 21

MIT's Crystal-CLIP Fuses 3D Atomic Structure With Text, Outperforms Text-Only Baselines on Materials Discovery

news.mit.edu · 2026-10-18 · Score : 22

Oxford's Gen4P Turns Video Generators Into Single-Pass 4D Perceivers, Hits SOTA on 3D Point Tracking

huggingface.co · 2026-10-16 · Score : 17

TypeSafe-Family MJev-Spatial Hits 65.64% on Spatial Benchmarks With Single-Pass Attention Matching, 2.5x Faster Than Autoregressive Baselines

huggingface.co · 2026-10-16 · Score : 21

Zhejiang's VLA-Δ Supervises Vision-Language-Action Models on Physical State Transitions, Beats π₀.₅ on Real Robots

arxiv.org · 2026-10-15 · Score : 15

3DCodeVerse Opens 2.03M Executable 3D Code Pairs Across Six Languages, 90% Win Rate Over Claude Design on Human Eval

huggingface.co · 2026-10-16 · Score : 17

NUS's Velox Routes Video World Model Attention via Frustum Slabs, 4.36x Attention and 2.43x Denoising Speedup at Chunk 48

huggingface.co · 2026-10-16 · Score : 18

Wisconsin-Madison's MaLoW Internalizes Streaming-Video History Into Evolving LoRA Weights, Hits 84.8% on StreamingBench

huggingface.co · 2026-10-16 · Score : 19

Shanghai AI Lab's Intern Lumina U2 Unifies Image, Video and 3D in a 16B Multi-Codebook Diffusion LLM

huggingface.co · 2026-10-16 · Score : 23

Swiss AI Opens Apertus 1.5 With 262K Context, Multimodal Weights and 4T Added Pre-Training Tokens

huggingface.co · 2026-10-15 · Score : 22

ALERTE Nvidia Open-Weights Nemotron 3.5 Super VL, a 120B MoE With 12B Active Parameters and 1M Context

venturebeat.com · 2026-10-15 · Score : 32

70+ Researchers' Survey Maps 297 Multimodal Agent Memory Architectures Across Three Lifecycle Tiers

huggingface.co · 2026-10-15 · Score : 17

NVIDIA's TacitVLA Matches Reasoning VLA Baselines With 3.4x Lower Latency by Decoupling Action From Text

huggingface.co · 2026-10-14 · Score : 24

Reka Ships Edge 2603, a 7B Multimodal VLM Needing Only 331 Tokens Per 1024x1024 Image

huggingface.co · 2026-10-09 · Score : 23

Cloudflare Launches Clef-Omni Multimodal, Cuts Clef-Flash Pricing 58% to $0.038/M

blog.cloudflare.com · 2026-10-09 · Score : 27

OneSearch-VL 8B Beats Qwen3-VL-8B by 20 Points on Multi-Image Research, 27 Points on Video DR

huggingface.co · 2026-10-09 · Score : 23

KAIST's ME-World Jointly Denoises Multi-Agent Egocentric Video for Fine-Grained Embodied Interaction

huggingface.co · 2026-10-08 · Score : 22

Zhejiang's SpaceCast-Bench Finds Top VLM Scores 58% on Predictive Spatial Reasoning vs 87% Human Baseline

huggingface.co · 2026-10-08 · Score : 20

ALERTE Perplexity Open-Sources pplx-embed-v2-late, 0.6B and 9B Multimodal Late-Interaction Embedders

perplexity.ai · 2026-10-07 · Score : 30

UCLA OpenSLA Unifies 79 Sensor Modalities, 60 Action Groups and 116K Patients With Language Bridge

huggingface.co · 2026-10-07 · Score : 16

ALERTE OpenAI Rolls GPT-6 to All ChatGPT Tiers With Intelligent UI That Renders Charts, Buttons and Forms in Replies

techcrunch.com · 2026-10-07 · Score : 33

Google Launches EmbeddingGemma 2, a 740M Multimodal Embedder Under Apache 2.0 With 9.92-Point MTEB Code Gain

blog.google · 2026-10-06 · Score : 26

HF Paper HLA-WM Uses Hybrid Linear Attention to Stabilize Long-Horizon Video World Models

huggingface.co · 2026-10-05 · Score : 18

OmniConfess Paper Scores Token-Level 'Confessions' to Catch Omni-Modal LLM Hallucinations on 3,540-Example Benchmark

huggingface.co · 2026-10-06 · Score : 16

ACG-Bench Paper Debuts AE-VLA Dual-Arm Policy, Lifts Simulation Success to 21.53% From 2.94%

huggingface.co · 2026-10-06 · Score : 17

HF Paper Where-OPD Uses Synthetic Scenes to Self-Distill MLLMs, Transfers 3.23-Point Gain to Real Benchmarks

huggingface.co · 2026-10-02 · Score : 19

HF Paper Omni-Embed-Mini Hits 49.57 nDCG at 0.9B Params by Freezing Text, Rivals Gemini Embedding 2 at 2.3B

arxiv.org · 2026-10-01 · Score : 19

NeurIPS Paper Shows VideoLLMs Lose Temporal Signal Mid-Network, Proposes Inference-Time Activation Injection Fix

arxiv.org · 2026-10-01 · Score : 17

HF Paper OneStreamer Trains Streaming Video Model on 1M-Record Dataset, Tops Eight Benchmarks at 4B

arxiv.org · 2026-10-01 · Score : 17

Stability AI's 4Director Lets Video World Models Take Rigid 3D Scene Control

huggingface.co · 2026-10-01 · Score : 19

Bilibili Open-Sources Index-Translate Covering 150 Languages, Tops Out at 35B MoE Under Apache 2.0

huggingface.co · 2026-09-30 · Score : 27

HF Paper: APM-Bench Benchmarks Cross-Session Persistent Memory for Egocentric Streaming Video Assistants

huggingface.co · 2026-09-30 · Score : 18

InfiniHand Paper Streams World-Space Hand Motion From Egocentric Video at 11 FPS

arxiv.org · 2026-09-28 · Score : 17

Paper: Encoder-Free Multimodal Models Predicted to Catch Encoder-Based at ~10^22 FLOPs

arxiv.org · 2026-09-28 · Score : 19

Stanford's HomeBody Wires GPT-6 Astra Directly to a Unitree G1, Skips the VLA Layer

stanford.edu · 2026-09-26 · Score : 26

Nvidia/Bar-Ilan Paper Adds KL Regularizer That Cuts Audio-Video Desynchronization From 0.804 to 0.752 in Joint Diffusion

huggingface.co · 2026-09-24 · Score : 17

RoboFollow Benchmark Shows Top VLA Policies Ace L0 but Fail Instruction Following at L1–L3

huggingface.co · 2026-09-22 · Score : 24

Qwen Team Releases OmniVChat: Native Audio-Video Dialogue Task, Bench and RL Recipe in One Drop

huggingface.co · 2026-09-21 · Score : 19

ALERTE Alibaba Ships Qwen3.8-Omni-Flash, a 1M-Context Omnimodal Model With 26%+ Average Gain Over 3.5-Omni-Plus

technode.com · 2026-09-18 · Score : 29

ActionPiece VLA Tokenizer Prioritizes Physical Rank Consistency, Hits 94.8% LIBERO With Qwen3-VL-4B

arxiv.org · 2026-09-17 · Score : 24

Snap Unveils $2,195 Standalone Specs AR Glasses and Free Specs Intelligence AI Assistant

theverge.com · 2026-09-16 · Score : 23

CMU's ModAR Ships First World-Action Model That Autoregressively Denoises Depth, Point Tracks and DINO Features

huggingface.co · 2026-09-15 · Score : 20

LynnReal-Omni Paper Unifies Video Generation With 32B Multimodal DiT, Ships 27B Flash at 377ms per 540p Clip

arxiv.org · 2026-09-14 · Score : 18

'Omni-Streaming Thinking' Paper Benchmarks Streaming Video Reasoning Against GPT-6 Astra and Sol

huggingface.co · 2026-09-15 · Score : 17

PhysBrain 1.5 Paper Extends Vision-Language Models Into Physical Foundation Models

huggingface.co · 2026-09-15 · Score : 18

Agnes-AI Ships Open-Weight 33B Multimodal Agnes-3.0-Flash With 262K Context Under Apache 2.0

huggingface.co · 2026-09-11 · Score : 22

SenseNova-U1.5: 8B Native Unified MoT Beats Nano-Banana-Pro on Vision Reasoning

huggingface.co · 2026-09-10 · Score : 20

Show-Harness Paper: Off-the-Shelf VLMs Can Control Robots Zero-Shot via a Semantic Action Interface

arxiv.org · 2026-09-09 · Score : 18

CoVeR Paper: Geometry-Only Token Pruning Retains 93.5% of Multi-View 3D VLM Accuracy at 8% of Tokens

huggingface.co · 2026-09-09 · Score : 18

TANGO Ships First Whole-Body 29-DoF VLA for Humanoid Navigation, Deployed Zero-Shot on Unitree G1

huggingface.co · 2026-09-09 · Score : 22

Tencent Ships Gander, an Open Omni-Interaction Agent With 100% Turn-Taking Accuracy on Full-Duplex Bench

huggingface.co · 2026-09-09 · Score : 21

OpenAI Ships ChatGPT Images 2.5 With Sketch Input, Cuts Latency 50% Over Images 2.0

openai.com · 2026-09-08 · Score : 23

Select, Compress, Reinvest Paper: Controlled Study of Visual-Token Allocation in Long-Video MLLMs Beats Uniform Budgeting

huggingface.co · 2026-09-04 · Score : 16

LatentStream Paper Replaces Retrieval With Progressive Latent Memory for Streaming Video Understanding

huggingface.co · 2026-09-03 · Score : 18

WorldReward Paper Unifies Action-Consistency and Visual-Quality Reward for Camera-Conditioned World Models

huggingface.co · 2026-09-04 · Score : 18

Hugging Face Paper: Puffin-World Unified 3D Multimodal Model Cuts Gravity Error to 1.32°

huggingface.co · 2026-09-04 · Score : 22

DroneCATS Benchmark Finds 2B MLLMs Fly Better Than Frontier Models But Fail to Terminate Episodes

huggingface.co · 2026-09-02 · Score : 18

ALERTE Fei-Fei Li's World Labs Unveils Atlas, a Multimodal World Model Generating 1-Minute 1440p Video

worldlabs.ai · 2026-09-01 · Score : 32

LightNav-0 Paper: Compact VLM Hits SOTA on 10 Navigation Sim Benchmarks With Dual-Channel Pointing

arxiv.org · 2026-08-31 · Score : 22

ALERTE Runway Unveils Solaris, Its First 'Interface World Model' That Generates UIs Frame-by-Frame in Real Time

runway.com · 2026-08-31 · Score : 29

DreamX-Creator Paper Pitches Native Audio-Video Generation at 2K Resolution

huggingface.co · 2026-09-01 · Score : 18

NavMCP Scaffolds VLM + Navigation Foundation Model, Gets 78.3% Success on Unitree Go2

huggingface.co · 2026-09-01 · Score : 16

ByteDance Seed Ships Lucida, a Parse-Generate-Place VLM Pipeline for Real-to-Sim Scenes

huggingface.co · 2026-09-01 · Score : 17

ALERTE DeepSeek Drops V4-Flash-Vision-Exp, a 305B MIT-Licensed Multimodal Agent Model

huggingface.co · 2026-08-31 · Score : 31

Numéros de la newsletter (14)

#528 What are companies building with AI? An Applied AI Deep Dive

2026-09-01 · 1 mentions

Frontier Research: circuits got cheap as consciousness got loud

2026-06-21 · 2 mentions

Generative AI: DeepSeek's 75% cut rewrites the cost math

2026-05-26 · 3 mentions

AI Regulation: Trump killed his own AI oversight order at the door

2026-05-26 · 3 mentions

AI in Marketing: Gemini eats the ad stack, agents take the keys

2026-05-26 · 1 mentions

Machine Learning: a reasoning model just cracked Erdős

2026-05-26 · 4 mentions

World Models: Google Ships Omni, and the LLM Monopoly Cracks

2026-05-26 · 6 mentions

AI Safety: the deepfake takedown clock starts ticking

2026-05-26 · 3 mentions

Machine Learning: nobody trained bigger, everyone shrank the cache

2026-05-21 · 3 mentions

World Models: Google's Genie now simulates real streets, not just dreams

2026-05-21 · 4 mentions

AI Safety: Washington floats an FDA for AI models

2026-05-21 · 1 mentions

Frontier research: the next architecture is not transformer-shaped

2026-04-25 · 1 mentions

Gen AI: $40B compute check, open-weight reset, software bloodbath

2026-04-25 · 1 mentions

AI Healthcare: Why pharma data is the real moat

2026-04-18 · 2 mentions