• The unit of capability is dropping from "model" to "skill." The same Microsoft group made both ends of the research workflow into skill suites — Idea for problem selection and grounding, Reel for turning a paper into a poster, video, and blog. The shared bet: slice capability into thin, editable contracts instead of training one end-to-end model to do everything.

• Cross-platform GUI agents are blocked by data and conflict, not model size. UI-MOPD shows executable cross-platform trajectories are scarce and platform conventions fight each other. It treats "learn a new platform without forgetting the old ones" as an explicit target through multi-teacher on-policy distillation. OSWorld and MobileWorld success rates of 38.2% and 12.0% are still low.

• Latency stays put, resolution climbs. Wan-Streamer v0.2 leaves the modeling untouched and pushes the interactive output stream from 192x336 to 640x368, holding roughly 200ms latency and 25 FPS. Pose, gaze, and hands become legible — the difference between a demo and a deployable avatar.

• A third path for teaching reasoning to diffusion language models. SFT has exposure bias, and RL rewards are too sparse while sequence likelihood is uncomputable. dOPSD gives the teacher its edge through later, more-decoded steps in the student's own denoising trajectory rather than external ground truth, improving both math and code on Dream and LLaDA.

Also Notable

• Goals Reached by Accident Get Relabeled Into Success Demos via Hindsight — recycling wasted rollouts into supervision eases the sparse-supervision problem in long-horizon, partially observable tasks.
• Qwen2.5's "Misaligned Persona" Is a Latent Direction, and It's Causal in Open Weights — transplantable, reversible, and preventable, giving an actionable switch for the broad misbehavior narrow harmful data induces.
• A Self-Driving Lab's Real Bottleneck Is Costly Wet-Lab Validation, and Agents Waste Rounds on Low-Value Experiments — this one targets that validation bottleneck directly.
• In Unified Multimodal Models, Does Training "Understanding" Actually Help "Generation"? — it runs transferability as an experiment instead of assuming it holds.
• Autonomous-Driving VLMs Hallucinate Under SFT and Stay Overcautious — CritiqueDriveVLM treats it with verifier-guided RL plus latent thought distillation, distilling a verification signal into reasoning.
• Gaussian Primitives From Feed-Forward 3D Reconstruction Are Highly Redundant, and Hard Pruning Creates Artifacts — AdaptiveSplat uses texture-aware controllable allocation to densify where it matters and save where it doesn't.
• Two Black Boxes Opened on Neural VRP Solvers: Encoder-Side Linear Probes, Decoder-Side Attribution — dispatchers can verify, accept, or compare solutions step by step instead of getting only a total cost.
• Evaluating Driving Planners Needs Controllable Safety-Critical Scenes, and Soft Guidance Only Gives Probabilistic Preference, Not a Guaranteed Collision — CCFM switches to hard-constrained flow matching to force the scenes out.
• Fine-Grained Calligraphy-Style Recognition Is Still Hard for LVLMs — HCSU adds a dataset and benchmark that stops mixing modalities and flattening labels, serving cultural heritage.
• Unsupervised Syllable Tokenizers Often Learn Speaker Identity as Content — this one uses speaker-disentangled chunk-wise regression to separate "who's speaking" from "what's said."

Read the full edition →