Daily Research Brief 2026-07-24

📊 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: ‘voice becomes a first-class control surface for agents’ and ‘agent safety moves from optional to default’ landing at the same time. OpenAI stuffed full-duplex GPT-Live voice into the desktop client, letting you orchestrate multiple agents by voice; Anthropic gave Claude voice Gmail/Slack/Canva connectors — conversational orchestration formally moves from demo to product. The same day Anthropic shipped a free Claude Code security plugin, and the OpenAI model-escape fallout at Hugging Face continued — safety is being productized by vendors as a default capability at coding time. On the model side, DeepSeek V4 stable hard-switched and retired old aliases, and Black Forest Labs pushed image generation further with FLUX 3.

1. Latest arXiv Papers

1. FedAgentKE: Federated Semantic Knowledge Evolution for Heterogeneous Agents — https://arxiv.org/abs/2607.21361

2. ResearchArena: Evaluating Sabotage and Monitoring in Automated AI R&D — https://arxiv.org/abs/2607.19321

3. Semantic Cooperative Games for Contribution Attribution in LLM-Based Multi-Agent Systems — https://arxiv.org/abs/2607.18253

4. Probabilistic Concept-Aware Steering for Trustworthy LLM Inference — https://arxiv.org/abs/2607.18259

5. Copy Less, Ground More: Overcoming Repetitive Copying in Long-Context Reasoning via Evidence-Aware Reinforcement Learning — https://arxiv.org/abs/2607.19345

6. Fence: Specialized SLM Guardrails for LLM Applications — https://arxiv.org/abs/2607.18268

7. When JSON Is Not Enough: Semantic Reliability of Schema-Constrained LLM Ordering Agents — https://arxiv.org/abs/2607.18261

8. Measuring Reward-Seeking via Contrastive Belief Updates — https://arxiv.org/abs/2607.18966