• The Maintenance Cost of Agents Is Shifting From the Model to the Harness. As models converge, the real engineering surface moves to the code that builds prompts, manages state, and calls tools — and this paper treats that layer as a first-class object worth its own methodology.

• An Image Stack That Owes Nothing to a Closed API, trained for $400K on 209M images, with four variants covering generation, editing, bilingual text, and fast inference — weights and code fully Apache 2.0.

• On-Policy Distillation Doesn't Raise the Ceiling; It's a Catalyst for Exploration. It steers a student onto paths it could already reach but wasn't taking, and prompt diversity beats resampling the same problem.

• Betting on the Non-Autoregressive Route? Read the Map First. Discrete diffusion's properties come down to how you construct the discrete state space, and this survey organizes scattered work into one framework from tokenization to generation.

Also Notable

• Pruning Shrinks the Model but Breaks Free Generation — pass@1 collapses to near zero while pass@k survives, so the ability is still there but the coherent output isn't; this uses "short-to-long" on-policy distillation to recover pruned models, a nice applied counterpart to today's OPD mechanism paper.
• Treating an Agent as "Foundation Model Plus Evolvable Scaffold": self-improving agents move from prototype to deployment, and this system-level survey is the other side of the Harness Handbook coin.
• Agent Evaluation Pipelines Are Fragmented and Tightly Coupled, so this work splits evaluation into three independent components and offers a lightweight, extensible base for it.
• Letting Understanding and Generation Correct Each Other Live in Masked Diffusion, like a teacher drawing while speaking, with the two modalities sharing latest decisions rather than only prior history.
• Turning Deterministic Emotion Recognition Into Ambiguity-Modeling One-to-Many Prediction: a conditional rectified-flow head carries uncertainty estimates for the blurry facial behavior of real scenes.

Read the full edition →