Microsoft launches 'Critique' for M365 Copilot
Microsoft released 'Critique' for M365 Copilot — a multi‑model research component that benchmarks and analyzes generated outputs to surface higher‑quality suggestions. The new tool signals growing emphasis on multi‑LLM evaluation and agent‑style research features inside productivity suites. (x.com)
Microsoft announced on March 30, 2026 that Researcher—the deep research agent inside Microsoft 365 Copilot—now includes two multi‑model capabilities named Critique and Council. (techcommunity.microsoft.com) Critique separates generation from evaluation by assigning a generator role (planning, retrieval, drafting) and a reviewer role that focuses on structure, sources, and citations, and Microsoft says the system uses models from Frontier labs including Anthropic and OpenAI. (techcommunity.microsoft.com) Microsoft reported Researcher with Critique improved aggregated DRACO benchmark performance by +7.0 points (SEM ±1.90), a +13.88% gain over the Perplexity Deep Research result cited in the DRACO paper across 100 complex tasks spanning 10 domains. (techcommunity.microsoft.com) Council (Model Council) runs the same question through multiple deep‑reasoning agents in parallel, preserves each model’s standalone full report, and adds a concise synthesis “cover letter” that highlights where models agree, diverge, and what each uniquely contributes. (mc.merill.net) Auto Critique and Model Council are rolling out to organizations enrolled in the Microsoft 365 Copilot Frontier program with a public‑preview rollout by March 30, 2026, and feature access and third‑party models (for example, Claude) are controlled by tenant administrators; a Microsoft 365 Copilot (Premium) license is required. (mc.merill.net) Microsoft says Critique’s review step is rubric‑based—mirroring academic and professional review workflows—and that emphasizing structured evaluation alongside generation is intended to raise factual accuracy, analytical breadth, and presentation quality in long‑form research outputs. (techcommunity.microsoft.com)
