Paint-Anything Unified any-color control for image generation and editing. Specify any object's target color with a 24-bit hex value, and the model paints exactly that shade. On FLUX.2-4B, +85.3% ACBench-T2I.

Ovis-Embedding Alibaba's new omni-modal embedding model maps text, images, video, and audio into one common space with a shared backbone, achieving state-of-the-art on MMEB-v3.

Grounded Action Models A new robot foundation model paradigm built on 3D grounding. Language, points, or boxes become a shared object-centric representation, mixed with robot state history to predict action chunks.

Ant Group’s Realtime-Venus is now on Hugging Face A full-duplex interaction system that listens while it speaks, sees audio-visual context, and delegates tasks to run in the background without pausing the conversation.

Alibaba's Wan team released WanPE A 397B-parameter prompt enhancement model that turns a short user prompt into a director-level, shot-by-shot cinematic screenplay, boosting video generator preference by up to 50.9 points at 30 seconds

Your LLM can hold two thoughts at once Averaging embeddings of two texts makes an LLM predict both continuations at once. This linear superposition is built into the Transformer, fades during pretraining, but can be restored with light fine-tuning. The authors show how to decode both streams from a single forward pass.

Agent-Editing World Model Rethinks world modeling for LLM agents by judging and editing noisy reasoning-action continuations before execution. AEWM hits 70.5% macro-F1 on Action Judge, +10.6 over the strongest baseline.

PKU-VaLuE-Lab releases OmniEcho A benchmark and omni-modal model for spatial audio-visual understanding in embodied agents, using first-order ambisonics to help agents localize sounds and navigate toward them.

Training Object Permanence in World Models Can video models learn object permanence? WROP offers 150 Blender-rendered cognitive tasks, 1.5M samples, and a 300-question exam. A 16B world model, PWM-WROP, ranks first among continuation models.

Tencent researchers just released RewardVerse A rubric-guided video reward framework that inserts a dynamic rubric between the evaluation query and the scorer, mitigating scalar drift for stable, interpretable RL rewards