• The bottleneck for agentic science moves from "is the model smart enough" to "can the data be governed." NVIDIA's NAIS wraps permissions, audit trails, and workflow orchestration in an institution-grade skeleton, then reproduces a hypertension GWAS behind the firewall of a hospital holding 286,000 patients — and the first pass was wrong until human review caught a phenotype bias.

• Let one MLLM act first, critique second, and a self-consistent loop replaces the external reward model. CycleGRPO turns the "region → text → region" duality into a native reward, needs only region inputs with no text ground truth, and lifts captioning, QA, grounded dialogue, and referring segmentation together.

• τ=K_t·I breaks under fast direction changes. MIT's NeuralActuator models servo dynamics with a Transformer directly, validated across three actuator families from the $500 OpenManipulator-X to the $30,000 Franka Panda, with sensorless force perception and motor health monitoring thrown in.

• Swap in a synonym and CLIP's grounding drifts. SynCLIP fixes this at the pretraining stage with spatial attention alignment and a synonym-augmented corpus, rather than patching downstream.

Also Notable

• A VLM's Logits Shouldn't Be Compared Directly Across Cross-Domain Candidate Classes — MED-DSLC restores comparability with domain supervision and logit calibration.
• Take Skeleton Kinematics Out of the Hardcoding — EquiFusion uses equivariant latent diffusion to let human motion prediction train across datasets without retargeting for the first time.
• Static Prompt Sets Miss T2I's Subtle Failures — DynEval switches to in-the-wild dynamic evaluation that tracks real usage.
• Apple Built a Stateful Benchmark for Visual Tool-Calling Agents — MM-ToolSandBox covers 500+ tools, 16 domains, multi-image and multi-turn.
• When Should an Assistant Speak Up on Its Own — Vinci2 learns proactive help from continuous first-person video instead of answering only when asked.
• Training-Free Vision-and-Dialog Drone Navigation — CoT plus structured spatial memory fills the gap in directional grounding.
• Bridge Monocular Relative and Metric Geometry — FoundationGeo uses DINOv3 initialization for a two-stage bridge.
• The Drift Problem in Long Autoregressive Diffusion Video — Cycle-World constrains error accumulation with cycle consistency from reverse prediction.
• Control Motion Transfer Through Attention Heads Inside a DiT — generation follows the reference motion while aligning to the target prompt.
• 360-Degree Spherical Video Frame Interpolation — SVI360 handles the severe distortion standard methods can't.

Read the full edition →