• The real bottleneck for video world models is broken physics, not image quality. PhysisForcing adds physical constraints at contact and deformation regions, lifting closed-loop success as a robot world model from 16% to 24% — right direction, still low.

• Reward went up while the image quietly got worse. NormGuard found that RL post-training inflates model norms by 5%-15%, a quality-loss signal the reward proxy can't catch but you can read off directly.

• Jailbreaks don't erase safety features — they suppress a handful of attention heads. Switch off a few early-layer "attack-vulnerable heads" and the model complies with harmful requests, while mid-layer safety heads stay active. Read them for training-free detection.

• A big lab turned a paper assistant into a pre-submission review tool. Google's PAT uses inference-time scaling to surface deep errors, recalling 34% more than zero-shot on SPOT, positioned as assisted verification rather than a verdict.

Also Notable

• Pulling 4D Multi-Object Interactions From Casual Monocular Video Into VLA Training — another path to real interaction data at scale, the "mining real data" counterpart to PhysisForcing's synthetic route.
• VLA Long-Horizon Error Accumulation Traces Back to Fixed-Weight Static Feature Fusion — S²-VLA treats it with state-space-guided dynamic fusion.
• Under Decentralized, Partially Observable Conditions, LLM Multi-Agents Misalign With Teammates and Environment State — LLawCo explicitly learns "cooperation laws" to model embodied multi-agent behavior.
• Neural Video Codecs Beat Classical on Efficiency but Resist Deployment, With Uncertain Cross-Hardware Results — MLVC builds a learned codec aimed at real multi-platform use.
• Adversarial Robustness Without Pruning or Masking — learn to both amplify and attenuate non-robust features; the method is lightweight.
• Fine-Grained Skill Assessment (Sports, Surgery) Needs Step-by-Step Visual Reasoning — combine latent visual diffusion with Monte Carlo tree search for stepwise judgment.
• Scene Text Detection Drops the Moment Distribution Shifts — TextDS does parameter-efficient representation alignment without large-scale pretraining.
• Two-Person Conversational Facial Animation Must Match Both High-Level Cognitive Intent and Low-Level Motor Reflex — existing methods handle neither end well; MindFlow coordinates both.
• Remote-Sensing Change Captioning Has Long Been Capped by Small-Model Capacity — RSICCLLM uses a multimodal large model to describe bi-temporal change.
• How Chemical Reaction Networks as a Biochemical Probabilistic Compute Substrate Can Be Reduced — a step toward cell-level adaptive programming.

Read the full edition →