• Open omni finally hits closed-flagship scale. Qwen3.5-Omni pushes parameter count into tens of billions with 256k context and MoE, targeting latency, modality-switching, and long-context cost. Voice and vision teams should re-evaluate self-host plans.

• LLMs judge better than they generate. On three pragmatic tasks, the same model scores higher as listener than as speaker. Any benchmark or reward signal built on self-judge now has an asymmetric blind spot.

• Reasoning models redefine unlearning. Even when the final answer is scrubbed, the chain of thought walks the forbidden knowledge back step by step. CiPO extends "don't output" to "don't follow this reasoning path."

• How public information changes over time is itself a training signal. Milkyway freezes the base model, updates only an external forecasting harness, and lifts FutureX from 44 to 61.

Also Notable

• CBM With CLIP Inherits Two Old Problems — Pretraining bias and concept granularity. This CVPR paper answers with concept-wise attention.
• Test-Time Text-Side Learning for OOD Detection — A practical patch for VLM applications that can't retrain after deployment.
• Concept Erasure Moves Beyond the Text Side — Adding image-side cooperation removes unsafe concepts more precisely. The T2I safety stack is maturing.
• 3DGS Finally Gets a High-Frequency Surface Fix — Neural Gabor basis functions give Gaussians frequency structure. CVPR-level gains.
• Medical CT Report Generation Goes Multi-Agent — A single VLM gives way to hierarchical agents mimicking multi-physician review. Vertical agentification.
• A Large Human-Labeled Benchmark for Video Editing and VFX — HF Daily pick. Infrastructure for video-editing evaluation.
• Zero-Shot Long-Horizon UAV Vision-Language Navigation — Breaks out a fine-grained cognition module. More solid than "big model plus generic prompt."
• Agent RL Data Should Co-Evolve With Agent Behavior — CoEvolve proposes a mutual-evolution framework.
• Layer-Wise Hidden State Dispersion as an Uncertainty Signal — More stable than assuming how hidden states evolve. A new signal for hallucination detection.
• SNR-Timestep Bias Is an Overlooked Diffusion Training Problem — This CVPR paper delivers a systematic diagnosis and mitigation.

Read the full edition →