• SimWAM Removes the Video Expert From Deployment: It learns temporal patterns through video generation during training, keeps only a lightweight action branch at inference, reaches 91.5 PDMS on NAVSIM, and transfers zero-shot to nuScenes.

• YOLO-PEFT Turns Adapter Placement Into Constraint Planning: RS-LoRA beats full-model fine-tuning on two YOLO models while cutting peak memory by 43.9%. It also identifies architectures unsuited to PEFT.

• Multi-Agent Forensics Gets Its Value From Dividing Evidence: Four roles inspect texture, lighting, motion, and physical consistency. Generalization to unknown synthesis methods still needs validation.

Also Notable

• Long-Term Role Consistency Does Not Mean a Personality Never Changes. This Benchmark Tests Whether Agents Change Plausibly After Life Events. — The study focuses on life-event-driven personality changes in LLM agents..
• A Restored Aggregate Score Can Hide Per-Question Damage After Contamination Mitigation. The Paper Audits Whether “Unlearning” Truly Restores Original Ability Through Probability Shifts. — Its method combines hierarchical per-question probability evaluation with staged mitigation..
• C4 Narrows Hard-to-Measure Creative Understanding Into Cross-Concept Decoding. It Tests Whether Models Can Infer Hidden Links Between Seemingly Unrelated Concepts. — The benchmark measures cross-concept understanding in multimodal large models..
• Capek 0.5 Puts Embodied Intelligence Back Into the Execution Loop. It Perceives, Reasons, and Verifies Again After Every Action Changes the Environment. — The model organizes vision-language reasoning around execution instead of producing a complete plan upfront..

Read the full edition →