• SEED's Teacher Isn't a Bigger Model. It's the Same Model With Hindsight. It reads the trajectory it just ran, writes the lesson up as a natural-language skill, then uses the probability gap between with-skill and without-skill contexts to fill the supervision void outcome rewards leave mid-trajectory.

• Search Agents Spin in Circles and Burn the Budget; a Stronger Model Isn't What's Missing. SearchOS pulls task progress out of conversation history into a relational table you can query for coverage, and lets middleware decide when to interrupt.

• Language and Visual Imagination Share One Plan Sequence. Every step anchors to a physical state, which removes the integration work between a VLM and a world model. The cost is chained error: imagine one step wrong and the language reasoning that follows sits on a false premise.

• Splitting a Population Into Subgroups and Aggregating Beats Asking About the Whole. The identity needs no ground truth to check, so any pipeline estimating distributions with an LLM can self-audit today.

Also Notable

• Generating Self-Driving Test Scripts From Regulation Text, Where Compile Rate Is the Real Metric — treat it as mechanically verifiable DSL code generation and iterative RAG's job becomes fixing compile errors round by round; the evaluation approach generalizes further than the scenario generation does.
• Freezing a Small Model and Grafting Verified Knowledge Back as KV State: bit-exact (SHA-256 match, zero KL) describes the determinism of the restoration, not that the semantics survive splicing across contexts. Draw that boundary before reading claims like this.
• A Continuous-Space Diffusion Language Model Giving Every Token Its Own Timestep, mapping deterministically from Gaussian noise to a token canvas in one pass instead of sampling repeatedly. It slots into the non-autoregressive route map from a few days ago as one concrete bet.

Read the full edition →