• Unified image models don't fail at stacking abilities — the abilities fight each other. DanceOPD treats the conflict between text-to-image, local edits, and global edits as a training problem, reconciling them with on-policy field distillation and even absorbing inference tricks like CFG into the model.

• You don't need an external skill library to densify a sparse reward. OPID infers hierarchical skill supervision straight from finished on-policy trajectories, turning a trajectory-level reward into dense token-level self-distillation.

• A video model's most dangerous failure is not knowing it can't see clearly. Under motion blur, glare, and occlusion, accuracy drops 15-30 points while the model stays oblivious. Robust-TO scores each frame's reliability before orchestrating tools, and degrades the least across five perturbations.

• The real selling point of discrete visual tokens is efficiency. ViQ pairs text-aligned pretraining with split-head quantization to keep both semantics and detail, cutting multimodal training time by 20% to 70%.

Also Notable

• A Plan-Reason-Search-Memory Agent Layer Over T2I — fills in the vague, implicit, knowledge-dependent requests users actually make. Qwen-Image-Agent
• NVIDIA's Two-Tower Split for Context and Denoising — separate networks for context representation and iterative denoising loosen up diffusion language models. Nemotron-TwoTower
• Ternary Quantization Without Costly Quantization-Aware Training — and it still holds down the accuracy loss. CAT-Q
• Real-Time Streaming Video Editing — solves background stability and low latency, two long-standing headaches, at once. LiveEdit
• Filtering Reasoning SFT Data Without a Strong Reasoner — high-quality signal shows up surprisingly early. Reasoning Quality Emerges Early
• Test-Time Scaling for Robot Manipulation — studies how reasoning actually scales on embodied tasks. E-TTS
• Small Open Models as GUI Agents — autonomous experience exploration and after-the-fact reuse close the task-planning gap. GUI Agent
• Separating Visual-Spatial Reasoning from Language Priors — diagnoses whether a VLM truly understands or just recites priors. CRISP
• Simulating Mechanics on World-Coordinate 3D Meshes, Not Pixels — more physically credible. PhysiFormer
• Cambridge Applies a Radical-Interpretation Philosophy — inferring an AI system's beliefs and intentions from computational facts. Radical AI Interpretability

Read the full edition →