Daily Research Brief 2026-06-22

📊 Token usage: estimated from retrieval and writing scale.

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

Editor’s Note

Sakana’s Fugu cuts inference cost 60%; Apertus opens a sovereign-AI foundation architecture; AutoJack shows a single web page can take over an agent host — agent security architecture is on notice; and a math AI proves a new theorem in 5 days. No industry section in the original issue; the paper thread stands on its own.

1. Latest arXiv Papers

1. LLM-as-Environment-Engineer: Letting AI Design Its Own Training Data — 4B Model Surpasses GPT-5.4 — https://arxiv.org/abs/2606.17682

2. SpatialTree: A Spatial Understanding Framework for Multimodal LLMs — CVPR 2026 — https://arxiv.org/abs/2606.22107

3. AutoJack Attack: A Single Web Page Can Remotely Control an AI Agent Host — Architectural Safety Concerns — https://arxiv.org/abs/2606.21894

4. Small Models, Big Impact: 0.6B Fine-Tuned Model Hits 95% Accuracy on Professional Classification — https://arxiv.org/abs/2606.22015

5. Apertus: An Open-Source Foundation Model Architecture for Sovereign AI — https://arxiv.org/abs/2606.21763

6. Sakana Fugu: Japanese Team’s Efficient Low-Cost Inference Architecture — 60% Lower Inference Cost — https://arxiv.org/abs/2606.21942

7. Math AI Breakthrough: AI Proves a New Math Theorem in 5 Days, a Human PhD’s Two-Year Workload — https://arxiv.org/abs/2606.22089

8. Edge LLM Deployment Framework: 35B MoE Model Runs Smoothly on Phones — https://arxiv.org/abs/2606.21837