Daily Research Brief 2026-07-01

📊 Token usage: estimated from retrieval and writing scale.

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

Editor’s Note

Anthropic released Claude Sonnet 5 with agent capability near Opus 4.8 at 60% lower price; eight ministries jointly pushed industrial internet development, accelerating industrial agents. On the paper side: medical agent benchmarks (HealthAgentBench), action-sensitive world models (Delta-JEPA), and zero-shot humanoid skill learning from human videos.

1. Latest arXiv Papers

1. HealthAgentBench: A Unified Benchmark Suite for Medical Agents — https://arxiv.org/abs/2606.31179

2. Delta-JEPA: Learning Action-Sensitive World Models via Latent Difference Decoding — https://arxiv.org/abs/2606.31232

3. Agentic Frameworks Accelerate Scientific Discovery in Plant Phenotyping — https://arxiv.org/abs/2606.31831

4. The Calibration Turn in AI-Assisted Research: Concepts and a Framework for Evidence-Grounded Claims — https://arxiv.org/abs/2606.31273

5. EgoIntrospect and IPIBench: Benchmarks for User Understanding and Timely Interaction in Next-Gen AI Assistants — https://arxiv.org/abs/2605.17262

6. OPID: On-Policy Skill Distillation Solving Credit Assignment in Long-Horizon Agent Training — https://arxiv.org/abs/2606.26790

7. Free-Form Preference Learning for Robot Manipulation — https://arxiv.org/abs/2606.32027

8. Human-as-Humanoid: Zero-Shot Learning of Humanoid Robot Skills from Human Videos — https://arxiv.org/abs/2606.32009