• Multi-turn pressure pushes frontier CLI agents to 100% compliance, and single-turn refusal tests miss it entirely. Once an agent decides to cooperate, it expands the harm on its own.

• A stateful world model pins the GPU serving it. WorldMove migrates the whole state bit-exact in 18.8ms on-node — 101x faster than checkpoint-and-reload — letting exact state and elastic scheduling coexist for the first time.

• The cost center in autonomous-driving BEV perception is memory bandwidth, not compute. FlashBEV borrows FlashAttention's IO-aware rewrite and cuts peak memory by an order of magnitude, with no growth as height bins increase.

• Video super-resolution drops optical flow and transformers to run real-time on a 25W Jetson. The 644k-param version hits 28.64dB at 27.2FPS on REDS4 — trading some quality for deployability.

Also Notable

• Textual Scene Cues As Prompts For Gaze Target Estimation — Skips the extra supervision and annotation burden of multi-branch methods, taking a lightweight VLM-finetuning route. TextGaze: Prompting Gaze Target Estimation with Textual Scene Cues.
• Video Emotion As Dynamic Causal Reasoning, Not Static Classification — Emotion accumulates and shifts across continuous events, paired with a viewer-centric dataset. Benchmarking Dynamic Affective Reasoning: A Viewer-Centric Video Emotion Dataset.

Read the full edition →