Daily Research Brief 2026-07-05

📊 Token usage: estimated from retrieval and writing scale.

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

Editor’s Note

GitHub Copilot integrated open-source Kimi K2.7 Code for the first time — Chinese models entering the international mainstream; the industry logic is shifting from a compute arms race to fine-grained operations. On the paper side: multimodal search agents, chain-of-thought under limited supervision, failure-aware robot recovery and rare-disease medical reasoning.

1. Latest arXiv Papers

1. SimpleSearch-VL: An Efficient Multimodal Search Agent Framework — https://arxiv.org/abs/2606.31504

2. OPINE-World: Interactive Exploratory Procedural World Modeling with Ontology-Error Priority — https://arxiv.org/abs/2607.01531

3. Semi-Supervised Chain-of-Thought Learning: Re-Estimating Reasoning under Limited Supervision — https://arxiv.org/abs/2607.01511

4. VIDEO-MME-LOGICAL: A Benchmark for Temporal Logical Reasoning in Video — https://arxiv.org/abs/2606.27828

5. FAR: Failure-Aware Retry Lets Robots Recover Autonomously from Failures — https://arxiv.org/abs/2607.01111

6. Bayesian Uncertainty Propagation for Agentic RAG Pipelines — https://arxiv.org/abs/2607.00972

7. RareDxR1: An Autonomous Medical Reasoning System for Rare Diseases — https://arxiv.org/abs/2607.00147

8. Program-as-Weights: A Programming Paradigm for Fuzzy Functions — https://arxiv.org/abs/2607.02512