XYZAILab launches 35B and 397B DeepSearch agents built in an AI4AI loop
机器之心 · wechat · 2026-07-25
What XYZAILab released
- XYZAILab launched two DeepSearchAgent systems: XYZ-Aquila-mini (35B) and XYZ-Aquila-pro (397B).
- The lab says both models set new state-of-the-art results on several DeepSearch benchmarks.
The core idea: AI4AI
- The project is framed as AI4AI: using AI to help build, debug, and improve AI systems themselves.
- The team says it ran a two-month effort with 10+ researchers, about 1,000 GPU-equivalent cards, and more than 300 agent workers.
- The workflow follows a human-governed loop: humans set goals, rules, budgets, and evaluation criteria; agents generate tasks, run experiments, propose fixes, and record evidence.
What changed in the system
- The system integrates task construction, training, runtime optimization, and evaluation into one closed loop.
- It introduces mechanisms such as append-only audit logs, visible state reconstruction, state-faithful SFT, and independent evaluators/gates.
- AI also proposed new engineering ideas from failure traces, including proactive context summarization and a persistent “research notebook” concept.
Results and implications
- XYZ-Aquila-mini reportedly leads on seven public benchmarks including BrowseComp, BrowseComp-ZH, DeepSearchQA, GAIA, LiveBrowseComp, Humanity’s Last Exam, and WideSearch.
- XYZ-Aquila-pro hits top public scores on some benchmarks and matches Kimi-K2.6 on DeepSearchQA.
- The lab positions DeepSearch as a first public proof that AI can drive an AI R&D loop, but only under explicit rules and independent verification.
More from coding & agent
- Agent verified 'success' after double-voting: a lesson in harness vs model bugs — TejasKumar_ · 2026-10-26
- diginomica: agentic automation hype is overdone, and 'The Hackening' of enterprise security has begun — jonerp · 2026-10-26
- 'Vibe verification': ~60% of AI-written formal specs prove the wrong property — basedjensen · 2026-10-26
- Devin launches Security suite powered by Claude Mythos with $100M+ credits — DevinAI · 2026-10-26
- HackBench injects reward hacks into agent trajectories to stress-test monitors — DanielKhashabi · 2026-10-26
- Hugging Face ships hf CLI updates: deduped cache saves 19GB, faster buckets, retryable GPU jobs — yb2698 · 2026-10-26