Daily Research Brief 2026-07-15

📊 Token usage: estimated from retrieval and writing scale.

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

Editor’s Note

Three threads today all point at the same thing: pushing AI’s ‘cost’ and ‘data’ bottlenecks down. Xiaomi U0 uses generative models to mass-produce robot training data, the E3 framework uses ‘initial operating points’ to cut 91% of redundant tokens for coding agents — one adds data, one saves compute. Agent evaluation is changing too: MM-ToolSandBox uses real data to puncture the ‘LLM tool-calling is ready’ illusion, showing visual precision is the real bottleneck.

1. Latest arXiv Papers

1. Do AI Agents Know When a Task Is Simple? Toward Complexity-Aware Reasoning and Execution — https://arxiv.org/abs/2607.13034

2. Hy-Embodied-VLM-1.0: Efficient Physical-World Agents — https://arxiv.org/abs/2607.12894

3. Let RGB Be the Language of Vision (RINO) — https://arxiv.org/abs/2607.12450

4. Contrastive-Augmented Flow Matching for Style-Content Disentanglement (CAtFM) — https://arxiv.org/abs/2607.12404

5. How to Realize Recursively Self-Improving Agents and Personal Singularity — https://arxiv.org/abs/2607.12254

6. Adaptive Cross-Modal Fusion with Sparse Attention for Pedestrian Crossing Intention Prediction (ADAPT) — https://arxiv.org/abs/2607.12293

7. MM-ToolSandBox: A Unified Framework for Evaluating Visual Tool-Calling Agents — https://arxiv.org/abs/2607.11818

8. Agentic Routing: The Harness-Native Data Flywheel — https://arxiv.org/abs/2607.11399