• NVIDIA Crams Language, Image, Video, Audio, and Action Into One Set of Weights. Cosmos 3 bets a single mixture-of-transformers can do every modality, and third parties rated it best open model in text-to-image, image-to-video, and robot policy.

• The Same KV Quantization Looks Fine in Prefill and Falls Apart in Long Decoding. KVarN shows the error compounds across timesteps, uses variance normalization to tame outlier token-scales, and takes 2-bit KV quantization to a new SOTA — calibration-free, with a vLLM implementation.

• Writing What You Learned In Context Back Into the Weights. "Language models need sleep" drops the metaphor: the mechanism is distillation plus self-rehearsal on synthetic data. But the abstract dodges the two hard questions — what to write back, and how to avoid forgetting.

• Sampling Budget Goes From Hand-Tuned Threshold to Learned Policy. Framing "how many samples to draw" as an MDP, an RL-trained controller small enough to run on CPU beats strong baselines on the "fewer samples, no accuracy drop" tradeoff.

Also Notable

• A Second KV-Cache Line the Same Day: Evict Instead of Quantize. Finds a few value states with abnormally large magnitude that can't be dropped, confirming outlier token-scale as the shared pain of long reasoning.
• NVIDIA OmniDreams Runs Autonomous-Driving Closed-Loop Sim With a Real-Time Generative World Model. Targets the long-tail scenarios reconstruction-based simulators can't reach.
• World Models and MLLMs Are Complementary, So Learn the Tradeoff Instead of Asking Which Wins. Judges when a visual rollout is trustworthy and when to discard it.
• OVO-S-Bench Does Online Spatial Reasoning From a Continuous First-Person Stream. A layered benchmark that often needs evidence beyond the current field of view.
• VSTAT Moves Video Understanding From Recognizing Isolated Moments to Tracking Entities and States. Aimed straight at the weak spot in MLLMs.
• Wide-Baseline Matching as a Test Bed for Spatial Reasoning. Layered by viewpoint shift and matching granularity, it forces MLLMs to handle geometry and occlusion.
• PaddleOCR-VL-1.6 Refines the Last Generation's Weak Regions Instead of Blindly Scaling Data. Does region-aware refinement.
• Economy of Minds Uses Hayek's Decentralized Coordination to Let Agents Self-Organize by Bidding. Stronger collective intelligence emerges without central control.
• AUDITFLOW Builds an Executable Symbolic Environment for Financial-Report Auditing. Lets agents link facts to taxonomic concepts, recompute expected values, then decide.
• SynCred-Bench: AI Can Now Generate Images With Realistic Text and Layout, Creating a "Synthetic Credibility" Threat. A new kind of visual deception.

Read the full edition →