Microsoft (MSFT.US) unveiled a series of new artificial intelligence upgrades on Monday, introducing a novel multi-model deep research system named "Critique." According to the company's Chief Executive Officer, Satya Nadella, the system's core utilizes an innovative dual-model collaborative architecture, which decomposes the traditional single generation task into two independent stages: "generation" and "evaluation." Within this framework, the first model is responsible for initial task planning, information retrieval, and draft composition, while the second model acts as a senior reviewer, specifically tasked with verifying factual accuracy, auditing logical coherence, and refining the final report. This self-correcting mechanism, achieved through multi-model interaction, aims to fundamentally overcome the prevalent "hallucination" phenomenon in existing AI models, significantly enhancing the depth and analytical quality of research reports. Notably, the Critique system demonstrates high openness and compatibility, as its operation is no longer confined to a single model from a single vendor. Based on disclosed technical details, the system achieves cross-vendor intelligent collaboration by integrating models such as OpenAI's GPT series, Anthropic's Claude series, and Microsoft's self-developed Phi series. For instance, the system can leverage the powerful creative generation capabilities of a GPT model to produce a first draft, which is then subjected to rigorous logical auditing by a Claude model. According to Microsoft's latest benchmark test data, Critique has outperformed comparable single-architecture products on the market in core dimensions such as the breadth and depth of research findings and the quality of expression, demonstrating strong competitive advantages. Concurrently with the release of Critique, Microsoft also launched a complementary functional system called "Council," further enhancing the application scenarios for multi-model collaboration. The Council system allows users to run multiple AI models from different sources in parallel under the same interface for a single complex instruction. By introducing an independent third model as an "impartial arbiter," the system can automatically identify and summarize the similarities and differences in the outputs of each model, helping professional researchers capture unique insights that a single model might miss, thereby substantially improving the comprehensiveness of decision support. "After selecting Model Council in the Researcher's model selector, this feature becomes available. Council runs both Anthropic and OpenAI's models simultaneously, with each generating a complete, independent report that presents facts, citations, and analytical frameworks that the other model might have overlooked or weighted differently," Microsoft explained. "After both reports are generated, a dedicated judging model evaluates the reports, distills a summary of key findings, and highlights meaningful consensus or divergence between the models—including differences in scale, framework, or interpretation—while also pointing out the unique contributions of each model." It is understood that these two cutting-edge technologies have been initially integrated into the "Researcher" toolkit within Microsoft 365 Copilot. According to Microsoft's market rollout plan, the Critique and Council features have now entered an early testing phase, with initial access limited to enterprise customers enrolled in Microsoft's "Frontier Program." Analysts suggest that with the implementation of this deep research system, Microsoft's moat in the enterprise productivity tools market will be further strengthened, while also signaling that AI competition is evolving from a mere contest of model parameters to a new phase focused on complex system integration and logical verification.
Comments