• Thirteen Skills Make Research Agents Auditable. Spark-to-Paper raises hallucination detection from 14% to 92%, with 99.5% citation validity.

• Weak Models Improve From 0.49 to 0.91 Without Parameter Updates. AI4AI uses strong models to build execution harnesses from code, routing, and format constraints.

• Safety Evaluations Can Now Generate Adversarial Tool Environments at Scale. ToolHazard finds injection points and creates environment-specific payloads, but scale cannot replace business relevance.

• StateFlow Organizes 3D Previsualization Around World Memory and Render Feedback. Its access stage corrects camera plans into visually feasible trajectories.

Also Notable

• Creative Agents Need Business Context and Visual Cues — MBA tests six business scenarios and warns against reducing real-world context to plain text.
• Synthetic Paired Data Trains Video Reflection Removal — The approach offers a data-building route for temporal restoration tasks without real supervision.
• Road Geometry Priors Cut Driving Video Sampling — GeoFlow's speed gains matter only if they preserve consistency across moving objects.
• AVA-Encoder Gives Creative Agents Structured Video Representations — It aims to stop agents from treating video as nothing more than continuous pixels.
• Funnel Models Can Combine Conversion and Order Value — FunnelCausalNet estimates coupon uplift for zero-inflated, heavy-tailed revenue while staying closer to business decisions.
• Bilevel Optimization Reframes Black-Box Transfer Attacks — The method jointly handles initialization, surrogate adaptation, and gradient updates.
• Text and Visual Anchors Merge Anomaly Signals — The method tries to reduce zero-shot detection's dependence on text semantics.
• Earth Observation Embeddings Improve Local Weather Detail — They add surface information missed by coarse weather grids for probabilistic downscaling at arbitrary stations.

Read the full edition →