Daily Research Brief 2026-04-14

📊 Token usage: estimated from retrieval and writing scale.

Covers the latest AI research, open source and industry moves, updated daily.

Editor’s Note

The paper thread today: invariance in vision models, pair-based 3D scene generation, physics-informed state-space models, safety-violation detection across agent behaviors, and RL for physics olympiads.

1. Latest arXiv Papers

1. Who Handles Orientation? Investigating Invariance in Vision Models

2. Pair2Scene: Learning Local Object Relations for 3D Scene Generation

3. Physics-Informed State Space Models for Reliable Dynamics

4. Detecting Safety Violations Across Many Agent Behaviors

5. Solving Physics Olympiad via Reinforcement Learning

6. OmniShow: Unifying Multimodal Conditions for Human Animation

7. Budget-Aware Uncertainty for Radiotherapy Segmentation

8. SyncFix: Fixing 3D Reconstructions via Multi-View Consistency