Daily Research Brief 2026-07-17

📊 Token usage: estimated from retrieval and writing scale.

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

Editor’s Note

Today’s main thread: open-source models formally launch a frontal assault on closed-source flagships. Moonshot dropped Kimi K3 the night before Google’s Gemini 3.5 Pro release — 2.8T parameters, 1M context, open weights imminent — raising the ‘world’s largest open model’ bar to the trillion scale in one move and forcing the closed camp to answer ‘what is the premium for’. On the paper side, one engineering question dominates: where exactly do long-horizon agents get stuck?

1. Latest arXiv Papers

1. Hierarchical Denoising For Multi-Step Visual Reasoning (HDR) — https://arxiv.org/abs/2607.15278

2. Answer-Conditioned Chains of Thought Degrade Verifiable-Reasoning Distillation in LLMs — https://arxiv.org/abs/2607.14552

3. Stop Thinking, Start Looking: Efficient Post-Training for Multimodal Document QA via Reasoning-Free Alignment — https://arxiv.org/abs/2607.14682

4. HyMobileAgent: Data-Environment Co-Scaling for Efficient GUI Agents — https://arxiv.org/abs/2607.14548

5. Beyond Single Expert: Harmonizing Diverse Visual Priors in MLLMs for Spatial Understanding (ViPS) — https://arxiv.org/abs/2607.15054

6. Multi-Head Latent Control: A Unified Interface for LLM Agent Decision Making — https://arxiv.org/abs/2607.14277

7. Branching Policy Optimization: Sandbox-Native Language Agent Reinforcement Learning — https://arxiv.org/abs/2607.14171

8. TRACE: Turn-level Reward Assignment via Credit Estimation for Long-Horizon Agents — https://arxiv.org/abs/2607.13988