XYZ AI Lab Releases Two Deep Search Agents, Achieving Multiple SOTAs and Exploring AI4AI Paradigm

XYZ AI Lab's release of XYZ-Aquila-mini (35B) and XYZ-Aquila-pro (397B) is not just about two Deep Search Agents; it is a systematic validation of the AI4AI R&D paradigm. This paradigm lets AI participate in driving the complete R&D loop of agent systems, including task construction, model training, runtime optimization, and evaluation verification.

This direction is not isolated. Google DeepMind's AlphaEvolve has been used for data center optimization and chip design; Anthropic disclosed in May 2026 that over 80% of new code in its codebase was written by Claude; OpenAI introduced the RSI Index in GPT-5.6 to measure models' ability to participate in AI R&D; early versions of Kimi K3 also handled most GPU kernel optimization work.

XYZ's uniqueness lies not in point optimization but in integrating large-scale compute resources, multi-agent collaboration, and the complete R&D process into a closed loop, turning massive experiments into verifiable, reproducible, and accumulative R&D capability.

Deep Search tasks require the model to go through problem decomposition, web retrieval, source judgment, cross-page comparison, computation, verification, and final answer. Any link in this long chain can fail. The model is only part of the system; task construction, tool returns, context management, answer routing, and evaluation isolation all affect results.

XYZ-Aquila-mini achieves best results on seven benchmarks: BrowseComp (78.8), BrowseComp-ZH (82.9), DeepSearchQA (89.5), GAIA (97.1), LiveBrowseComp (48.7), HLE (51.1), and WideSearch (80.8). XYZ-Aquila-pro achieves the highest public scores on BrowseComp-ZH, LiveBrowseComp, and WideSearch, and ties with Kimi-K2.6 on DeepSearchQA.

These results demonstrate that the system comprising model, data, training, tools, context strategy, runtime, evaluation, and infrastructure is competitive, not just the model alone.

XYZ's AI4AI system follows the principle 'humans set rules, AI finds paths; independent evaluation, full traceability.' Humans define R&D goals, evaluation standards, resource budgets, and risk boundaries; Agents decompose complex problems into executable tasks, with different Agent Workers analyzing failure trajectories, proposing solutions, modifying code, and running experiments.

After experiments, candidate changes are evaluated by an independent evaluator that generates verifiable evidence. A gating mechanism decides to accept, reject, or escalate to human review based on preset criteria. Every experiment's plan, process, results, and judgment basis are fully recorded and written into shared memory. Effective changes are accumulated as reusable experience; failed attempts retain reasons and boundaries to reduce repeated exploration.

Agents cannot modify evaluation standards, read private answers, or self-approve solutions. When evidence conflicts, suspected data leakage, high-risk changes, or permission overruns occur, the system pauses automatic progression and escalates to human review.

The value of the AI4AI system lies not only in executing more experiments but also in identifying improvement directions not preset by humans from a large number of failure trajectories. For example, the team found that an Agent located a key webpage but did not use it in the answer; the problem was that the information was buried in an overly long context.

AI proposed two candidate directions from the trajectories: active context management (having the model actively organize evidence when the state is clear) and a 'research notebook' (writing candidate answers and sources into an independent workspace and cross-validating). The former has been incorporated into the current system; the latter is retained as a stateful design.

This shows that AI is beginning to expand the solution space the team can actually explore, discovering solutions beyond human preset from massive experiments, and then feeding them into the same evaluation and acceptance process.

AI4AI optimization does not revolve around tuning a single checkpoint; instead, it seeks verifiable improvements across multiple coupled stages: task construction, training, runtime, and reinforcement learning. Task construction builds evidence graphs from reliable sources to ensure questions are retrievable and answers unique; training uses state-faithful SFT, supervising only the information the model actually saw at that time.

Runtime uses append-only audit logs and progressively visible states to distinguish whether evidence was never found, existed but not presented, or was obtained but not used in the answer. Reinforcement learning explores process rewards and shared prefix reuse under context management to improve credit assignment and training efficiency in long-horizon tasks.

Deep Search is only the first public validation ground for the AI4AI system. In the future, this method is expected to extend to Coding Agents, general multi-tool Agents, and scenarios in law, medicine, and scientific research, with the underlying closed loop remaining consistent: propose candidates, execute experiments, independently verify, and accumulate experience.