Daily Research Brief 2026-07-08

📊 Token usage: estimated from retrieval and writing scale.

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

Editor’s Note

OpenAI officially released GPT-5.6 (Sol/Terra/Luna) on July 8 — the first frontier flagship after US export restrictions lifted, with Max deep-reasoning and Ultra multi-agent modes. Reuters reveals DeepSeek has been secretly building its own inference chip for a year to cut dependence on Nvidia/Huawei Ascend — full-chain domestic autonomy advances. On the paper side, Tsinghua’s JustGRPO (ICML 2026 outstanding paper) overturns diffusion-LLM training assumptions at 89.1% on GSM8K.

1. Latest arXiv Papers

1. The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models — https://arxiv.org/abs/2601.15165

2. From Failing to Passing: Evolving Natural Language Prompt Optimization Rules for LLM Code Generation — https://arxiv.org/abs/2607.05121

3. Medi-Gemma: A Hybrid Clinical Decision Support System Integrating Deterministic EMR Analytics and Retrieval-Augmented Generation — https://arxiv.org/abs/2607.04907

4. ASSEMCAD: Production-Ready CAD Assembly Generation from Natural Language — https://arxiv.org/abs/2607.05123

5. VISTA: A Visual Spec-to-App Benchmark for End-to-End Web Coding Agents — https://arxiv.org/abs/2607.02514

6. Distributed Attacks in Persistent-State AI Control — https://arxiv.org/abs/2607.02502

7. DemoPSD: Disagreement-Modulated Policy Self-Distillation — https://arxiv.org/abs/2607.02512

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