• Hierarchical Communication Removes Full Connectivity. Multics routes information through player, group, and global layers. It scales directly from four-player training to ten-player inference, though systematic evaluation mainly covers four to six players.

• DiT Attention Can Contract During Denoising. AReA adds no parameters and requires no retraining. It targets both high-resolution structural distortion and quadratic compute, but exact speedups still need confirmation.

• Client Isolation Must Cover the KV Cache. OLMo-3-7B-Instruct departed from the current client’s goal in 22% of eligible decisions across 20 controlled two-client scenarios. Probes and targeted ablations provide causal evidence, but deployment risks need testing across more models.

Also Notable

• RL for Masked Image Models Falls Into a “Visual Exploration Trap” — MaskSense addresses the visual exploration trap at both the sampling and optimization stages. It raises GenEval accuracy from 54% to 86% and HPS from 28.89 to 36.39..
• Single-Page Retrieval Cannot Find Every Cross-Document Clue — VL-DocIR finds cross-document retrieval much harder. Document-context scoring especially helps queries whose evidence spans several pages in one document..
• PDE Training Data Need Not Be All High-Fidelity — Mixing fidelity levels by numerical error laws matched accuracy with up to 6.3 times less data-generation compute. Tests covered six PDE problems and four surrogate architectures..
• Valid Correlation Matrices Are Only the Starting Point — MAGNET enforces manifold constraints while preserving clinical variation within broad disease labels. Across three cohorts, class-conditional fidelity improved by 7% to 21% over prior manifold-aware methods..

Read the full edition →