Daily Research Brief 2026-05-10

📊 Token usage: estimated from retrieval and writing scale.

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

Editor’s Note

Anthropic’s revenue jumped 80x quarter-over-quarter past $30B ARR; OpenAI shipped GPT-Realtime-2; Meta’s AI push is making employees miserable. On the paper side: SCPRM cumulative process rewards, embodied AI safety survey, and pixel-level visual attribution.

1. Latest arXiv Papers

1. When Is the Same Model Not the Same Service? A Multi-Provider Analysis — https://arxiv.org/abs/2605.02819

2. SCPRM: A Schema-Aware Cumulative Process Reward Model — https://arxiv.org/abs/2605.02815

3. FlexSQL: Flexible Exploration and Execution for Text-to-SQL — https://arxiv.org/abs/2605.02814

4. OGPO: Sample-Efficient Full-Finetuning of Generative Models — https://arxiv.org/abs/2605.03065

5. Safety in Embodied AI: A Survey of Risks, Attacks and Defenses — https://arxiv.org/abs/2605.02900

6. FINER-SQL: Boosting Small Language Models for Text-to-SQL — https://arxiv.org/abs/2605.03465

7. SIFT-VTON: Geometric Correspondence Supervision for Virtual Try-On — https://arxiv.org/abs/2605.01296

8. Chain of Evidence: Pixel-Level Visual Attribution for VLMs — https://arxiv.org/abs/2605.01284