Is Human R&D Over? XYZ’s Top Search Agents Sweep Seven Benchmarks and Close the AI‑4‑AI Loop

Introduction

XYZ AI Lab recently released two Deep Search agents, XYZ‑Aquila‑mini (35 B parameters) and XYZ‑Aquila‑pro (397 B parameters), which achieved state‑of‑the‑art results on multiple Deep Search evaluations.

Benchmark Results

On six publicly disclosed benchmarks, Aquila‑mini achieved the best scores across the board: BrowseComp 78.8, BrowseComp‑ZH 82.9, DeepSearchQA 89.5, GAIA 97.1, LiveBrowseComp 48.7, HLE 51.1, and WideSearch 80.8. Aquila‑pro, while still trailing some closed‑source models on certain sets, holds the highest public scores on BrowseComp‑ZH, LiveBrowseComp and WideSearch, and matches a 1 TB Kimi‑K2.6 model on DeepSearchQA.

AI4AI Research Paradigm

The authors argue that the breakthrough lies not only in model size but in the AI4AI research paradigm, where AI is no longer merely a training target but an active participant that drives the entire agent‑based R&D loop. This approach aligns with the broader concept of Recursive Self‑Improvement (RSI).

Examples of AI4AI in practice include Google DeepMind’s AlphaEvolve for data‑center efficiency, Anthropic’s codebase where >80 % of new code was generated by Claude, and OpenAI’s GPT‑5.6 which incorporates RSI Index metrics for debugging, kernel optimization, and training‑loop improvements.

System Architecture

The XYZ system defines goals, permissions, resources, and acceptance criteria by humans. Within these constraints, AI agents autonomously decompose tasks, generate experiments, run training or optimization, and produce verifiable evidence. Over two months, a core team of ten‑plus researchers coordinated more than 300 Agent Workers using kilocalorie‑level compute to integrate task construction, model training, runtime optimization, and evaluation into a single pipeline.

The workflow follows the principle: Human defines rules, AI finds the path; independent evaluation, full traceability. After an experiment, an Evaluator creates audit‑able evidence, and a Gate decides acceptance, rejection, or escalation to human review. All steps are logged, enabling replay, verification, and experience accumulation.

Case Study: Context Management

During long‑range searches, the team observed that agents sometimes retrieved key webpages but failed to incorporate the evidence into the final answer because the context became too long or was overwritten. The AI4AI system proposed two remedies: (1) proactive context summarization while the context is still clear, preserving confirmed evidence, open questions, and next steps; (2) a “research notebook” where agents record candidate answers, source links, and intermediate conclusions for later cross‑checking.

These proposals were implemented, leading to improved token‑length distribution and reduced tool‑call overhead, as illustrated in the accompanying figures.

Training and Reinforcement Learning

Training uses state‑faithful supervised fine‑tuning (state‑faithful SFT) , where the model learns from the exact visible state of successful trajectories, including failures, timeouts, and recoveries. Reinforcement learning with process‑reward signals and context‑management mechanisms further refines long‑term decision making for Deep Search tasks.

Conclusions

The XYZ‑Aquila results demonstrate that a large‑scale, AI‑driven research loop can produce verifiable, reproducible improvements without handing over rule‑making or final judgment to the model. Hundreds of agents act as specialized executors while humans retain strategic control, enabling systematic, auditable progress toward recursive self‑improvement.