Daily Research Brief 2026-07-13

📊 Token usage: estimated from retrieval and writing scale.

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

Editor’s Note

Today is a ‘release day’ main thread: OpenAI rolled out GPT-5.6 broadly, merged Codex into ChatGPT, and launched Work agents that can run for hours; Anthropic postponed Claude Fable 5 subscriptions a second time, while OpenAI removed Codex rate limits — the two giants’ ‘agent entry point war’ is at white heat. Even more notable is the concentrated domestic open-source burst: SenseTime’s unified vision model, Ant’s agent safety guardrails, Meituan’s 1.6T LongC…

1. Latest arXiv Papers

1. Wrong Before Right: ‘Late Repairs’ and Interface Failures in Aligned Language Models — https://arxiv.org/abs/2607.04640

2. Gen4U: Unifying Video Generation and Understanding via Diffusion — https://arxiv.org/abs/2607.06856

3. Tool-Making and Self-Evolving LLM Agents in Low-Latency Systems — https://arxiv.org/abs/2607.08010

4. Next-Generation Agentic RL Systems Enable Self-Evolving Agents (AReaL 2.0) — https://arxiv.org/abs/2607.01120

5. Is One Layer Enough? Single-Layer RL Matches Full-Parameter RL Training — https://arxiv.org/abs/2607.01232

6. Vidu S1: A Real-Time Interactive Video Generation Model — https://arxiv.org/abs/2607.03118

7. SciReasoner: Deep Native Structural Reasoning for Science — https://arxiv.org/abs/2607.07708

8. A Survey of World Models: Toward a Trinity Architecture for Physical AGI — https://arxiv.org/abs/2607.06401