SPLIT-RL, a staged post-training approach that trains visual reasoning and language reasoning in disjoint phases, improves average accuracy over GRPO by 1.4 to 6.1 points across Qwen3-VL models from 2…
NVIDIA researchers found that vision language models become significantly less likely to refuse harmful requests when given tools, with relative refusal failure rates rising up to 68.7% across all ele…
FastH3 V2, an eight-step video-and-audio generation model whose weights previously required 138 GiB and data-center GPUs, now runs on a single consumer machine — an RTX 5090, RTX 4090, RTX PRO 6000, D…
Alibaba's Qwen released Qwen-Image-2.1, a 7B-parameter image generation and editing model that combines a 32-layer single-stream DiT with a Qwen3-VL 8B text encoder and a 64-channel RGBA VAE, but ship…
NVIDIA's Vera Rubin NVL72 delivered up to 3.7x higher throughput than the GB300 NVL72 on the Qwen3-VL model in MLPerf Inference v6.1 preview submissions, using vLLM with the NVIDIA Dynamo framework, w…
NVIDIA's Vera Rubin NVL72 system debuted in MLPerf Inference v6.1 with up to 3.7x higher throughput than GB300 NVL72 on the Qwen3-VL benchmark and up to 2.5x higher throughput on DeepSeek-R1, NVIDIA r…
Researchers at HKUST-KnowComp introduced PACE (Progressive Acquisition of Critical Evidence), a factor-guided framework for long-video question answering that improves evidence recovery and answer acc…
A developer contributed a patch to ai-toolkit's MiniMax-H3 diffusion model extension that fixes reference-to-video-alignment (ref2va) handling of control images and videos in IC-LoRA-style video-to-vi…
Researchers from NEC Laboratories America introduced CAViAR (Causal Accident Video and Incident Analysis Repository), a human-annotated dashcam benchmark of 2,249 real-world accident videos from CarCr…
A new study from arXiv (2608.17095v1) introduces inference-time attention steering for vision-language-action (VLA) driving models, applying a bounded additive pre-softmax attention bias to visual tok…
Amplifying the weight difference between a reasoning model and its non-reasoning counterpart—creating an 'overthinking model'—surfaces hidden secrets up to 10× more often than the original reasoning m…
An ICML 2026 paper submitted July 9 reports that amplifying the weight-space difference between Qwen3-VL instruct and reasoning checkpoints made planted secrets or unintended behaviors surface up to 1…
A new tutorial from StarTrail-org introduces PixelRAG, a pixel-native retrieval-augmented generation pipeline that renders web pages and PDFs as images, divides them into overlapping tiles, and genera…
NVIDIA released Cosmos 3 Edge, a 4-billion-parameter open world model designed to run on-device for robots and vision AI agents, enabling real-time reasoning and local action generation. The model is …
A new evaluation technique called reversal-drop reveals that temporal video models often rely on positional encoding rather than understanding visual sequences, with Molmo2 failing to distinguish even…
Hugging Face released LeRobot v0.6.0, introducing world model policies (VLA-JEPA, FastWAM, LingBot-VA) that learn to imagine future states, new vision-language-action models (GR00T N1.7, MolmoAct2, EO…
Developers integrating Qwen3-VL vision-language models into production pipelines face three silent failures: the thinking budget trap where the model consumes output tokens on internal reasoning, leav…
Krea AI released Krea 2, a series of foundation models for image generation focused on creative exploration, featuring a diffusion transformer architecture and systems for prompt expansion and style r…
The Qwen team released three embodied AI models—Qwen-RobotManip, Qwen-RobotWorld, and Qwen-RobotNav—as part of Qwen-Robot-Suite, each built on a Qwen vision-language backbone and targeting different r…
Researchers introduced ToolGate, a lightweight external controller that predicts whether to execute or skip tool calls proposed by vision-language agents, addressing the pre-call control problem where…