Microsoft Researchers excel in Perplexity's self-developed benchmark: dual-model review advances to Frontier, Cowork synchronously opens long-task proxy

According to 1M AI News monitoring, Microsoft has rolled out two new Microsoft 365 Copilot capabilities through Frontier (an enterprise user early access program where participants can try Copilot features that are not yet officially live).

Researcher (a built-in deep research Agent for Copilot) has added two multi-model collaboration modes: Critique and Council. Critique involves collaboration between Anthropic and models from OpenAI: one handles planning, retrieval, and drafting, while the other is responsible specifically for reviewing and refining. When Auto is selected, it is enabled by default. Council also runs two models in parallel; each generates a complete report, and then a separate evaluation model consolidates the similarities and differences. Microsoft uses GPT-5.2 as the judging model (the strictest among the three judging methods in the original paper) and tests Critique on the DRACO benchmark (100 complex research questions released by Perplexity researchers, covering 10 domains). Overall, the combined score is 7.0 points higher than the benchmark’s best system, Perplexity Deep Research (using Claude Opus 4.6), representing a relative improvement of 13.88%. The original DRACO paper did not include Critique. This is data Microsoft obtained by testing it themselves under the same evaluation protocol.

Copilot Cowork is designed for longer, multi-step work: it first generates a plan based on the goal, then advances step by step across tools and files, showing progress along the way, and users can step in at any time. Microsoft cites Capital Group as an early trial case, saying it has been used for project planning, scheduling, creating deliverables, and preparing executive review materials.