• Trimming context to save tokens can backfire when the cache misses. Cutting old text is the obvious way to slim a long-running agent, but TokenPilot shows unconstrained rewrites disturb the prefix and break the prompt cache. The real trade-off has two axes: text sparsity versus cache continuity.

• Picking the highest-reward data to distill a small model may hurt it. On math reasoning, Oracle-refined high-score data drifts away from the small model's native style. That distribution shift raises learning cost, and the model does worse than on trajectories it sampled itself.

• UniDDT unifies understanding and generation by decoupling, not sharing. Understanding wants abstract semantics; generation wants pixel detail. Forcing both through one pathway satisfies neither, so UniDDT splits them structurally with a decoupled diffusion transformer.

• A geometry-conditioned latent surrogate speeds up two-phase flow simulation 60,000x. Instead of the full flow field, it learns the AMR mesh density — where the solver concentrates resolution. Inference runs in 0.045 seconds per trajectory, turning simulation into an interactive query.

Also Notable

• VinQA Interleaves Visual Elements Into Document QA Answers — most document QA returns plain text, wasting the tables, charts, and photos; interleaved answers fit real documents better.
• What Happens When You Add Two Opposing Steering Vectors at Once — past activation steering injects a single direction; this studies the collision.
• Reliable Uncertainty Intervals for Annual-Total and Year-Over-Year Forecasts — a multi-step split conformal method using block bootstrap plus cross-validated residuals.

Read the full edition →