• Meaning-Preserving Rewrites Shift 23 of 29 Review Metrics. Researchers turned 674 ICLR and NeurIPS reviews into 4,044 versions, with broadly consistent results across two LLM judges.

• G6D Solves 6D Pose Estimation Without Training. It combines instance masks, camera intrinsics, and CAD models for geometric matching and refinement, with a CPU-only configuration available.

• Q-TIE Separates Time Constraints From Semantic Similarity. It maps a query's temporal intent to start and end dates, then uses that interval as an independent RAG reranking signal.

Also Notable

• Spatial Audio Adds Semantics and Directional Cues — OmniEcho covers spatial audiovisual perception and embodied navigation, but fine-grained localization and distance estimation remain clear weaknesses..
• Continual Learning Changes Sharply Across Tasks and Viewpoints — CE⁴L systematically evaluates representative methods across four task categories under first-person, third-person, and cross-view continual learning..
• Preference Alignment Can Make Outputs Less Human-Like — The study recommends measuring what humans prefer separately from how humans answer. Standard DPO also shows this gap..
• With a Fixed Cache, Changing Only the Retrieval Encoder Raises ImageNet-A Gains From at Most 0.44 to 19.7 ± 0.4 Points — A comparison of 16 retrieval spaces makes retrieval space itself a first-class design variable in cache adaptation..

Read the full edition →