• All Intrinsic RLVR Is Just Sharpening the Initial Distribution. Model prior quality sets the training ceiling. Model Collapse Step can predict feasibility before you commit resources.

• Code Beats Natural Language as a Spatial Reasoning Chain. Structured layout benchmarks improve 68.83%, with the largest gains on dense text and multi-element scenes.

• Imitation Learning's Structural Flaw Is Missing Judgment Training. ACT uses RL to make models compare and evaluate candidate actions. The critical thinking transfers to out-of-distribution tasks.

• High-Noise Diffusion Steps Only Need a Thumbnail. The information content equals a downsampled low-res image — full-resolution processing is wasted compute. Theory is solid, but quality tradeoffs at high resolution need validation.

Also Notable

• Unified Editor Uses MoE Routing to Dynamically Allocate Condition Signal Weights — solves mutual interference from static multi-task fusion.
• New Fix for Error Accumulation in Autoregressive Long Video — hierarchical denoising finds a better balance between temporal continuity and frame quality.
• 400 Expert-Level Agent Tasks Spanning Law, Finance, and Medicine — directly benchmarks million-dollar real-world decision scenarios.
• Explicitly Guiding ViT Fine-Tuning Toward Semantic Concepts Over Background Cues — improves robustness under distribution shift.
• Test-Time Adaptive Learning of New Classes Without Retraining — practical capability for online streaming scenarios.
• Benchmarking VLM Reasoning on Subtle Visual Differences — targets industrial inspection and medical imaging.
• Understanding Diffusion Distillation Through Weight Direction — enables more stable one-step image generation.
• Prototype-Guided Erasure of Broad Concepts From Diffusion Models — can remove entire art styles, not just individual characters.
• LLMs Switch Behavior Modes via Conditional Tokens — intrinsic behavioral plasticity, like a chameleon adapting to its environment.
• Linear Compensation Recovers Blocks Skipped by Sparse Attention — speeds up video generation without quality loss.

Read the full edition →