Daily Research Brief 2026-05-21

📊 Token usage: estimated from retrieval and writing scale.

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

Editor’s Note

Google I/O 2026 shipped Gemini 3.5 Flash (4x faster, 40% cheaper); two AI research systems landed in Nature the same day, opening autonomous AI science; Nature warned AI is inflating paper counts while quality declines. On the paper side: HAGE external memory retrieval, 4x context ‘glasses’, and Frontier-Eng Bench for agent capability.

1. Latest arXiv Papers

1. HAGE: Making AI Retrieve External Memory More Intelligently — https://arxiv.org/abs/2605.09942

2. HKUST & ByteDance: ‘Glasses’ for Long Documents — 4x Context Window — https://arxiv.org/abs/2605.13831

3. Meituan CVPR 2026: U-Mind Real-Time Multimodal Interaction Framework — https://arxiv.org/abs/2602.23739

4. In-Context Learning: Textbook Order Matters More Than Content — https://arxiv.org/abs/2605.13511

5. Two AI Research Systems Land in Nature the Same Day, Ushering Autonomous AI Science

6. Nature Warns: AI Explodes Paper Count but Declines Science Quality

7. ShapeCodeBench: A Testbed for AI ‘Reading Images, Writing Code’ — https://arxiv.org/abs/2605.11680

8. Google I/O 2026: Gemini 3.5 Flash — 4x Faster, 40% Cheaper

9. Frontier-Eng Bench: Redefining Agent Capability Evaluation — https://arxiv.org/abs/2604.12290

10. New LLM Strategy: Proactive Capability Trade-offs Instead of Blindly Stacking Parameters