• 3DZip Preserves 94.7% Performance With 128 Tokens. It reduces 3D scene redundancy through feature diversity and spatial constraints, while raising inference speed to 1.92×.

• Clinical Multimodal Evaluation Must Catch Silent Failures. Per-sample, per-modality analysis shows that missing inputs can make wrong predictions appear more reliable.

• Frequent Replanning Can Make Recovery the Bottleneck. BRACE and E-RECAP allocate budgets by decision value and prune context, cutting tokens per replanning call by 62% to 92%.

Also Notable

• World Models Must Predict in a Geometry-Aware Policy Space Aligned With Action Generation and Spatial Change. SG-WAM explores self-guided world modeling in such a policy space.
• Heterogeneous Robots Can Train Depth Estimation Together Without Centralizing Data, but Hardware and Data Differences Demand Explicit Treatment. FeDepth designs federated depth estimation for heterogeneous robots.
• Querying Only Selected Response Coordinates May Support General-Loss Regression in Single-Index Models With Unknown Link Functions. The paper develops the theory from an active sampling perspective.
• Randomized Candidate Search Preserves the Quality of Sequential Feature Selection Without Scanning Every Feature at Each Step. The method targets feature selection in ultra-high-dimensional data.

Read the full edition →